**Estimation of Foot Plantar Center of Pressure Trajectories with Low-Cost Instrumented Insoles Using an Individual-Specific Nonlinear Model**

### **Xinyao Hu 1, Jun Zhao 1, Dongsheng Peng 1, Zhenglong Sun <sup>2</sup> and Xingda Qu 1,\***


Received: 29 November 2017; Accepted: 30 January 2018; Published: 1 February 2018

**Abstract:** Postural control is a complex skill based on the interaction of dynamic sensorimotor processes, and can be challenging for people with deficits in sensory functions. The foot plantar center of pressure (COP) has often been used for quantitative assessment of postural control. Previously, the foot plantar COP was mainly measured by force plates or complicated and expensive insole-based measurement systems. Although some low-cost instrumented insoles have been developed, their ability to accurately estimate the foot plantar COP trajectory was not robust. In this study, a novel individual-specific nonlinear model was proposed to estimate the foot plantar COP trajectories with an instrumented insole based on low-cost force sensitive resistors (FSRs). The model coefficients were determined by a least square error approximation algorithm. Model validation was carried out by comparing the estimated COP data with the reference data in a variety of postural control assessment tasks. We also compared our data with the COP trajectories estimated by the previously well accepted weighted mean approach. Comparing with the reference measurements, the average root mean square errors of the COP trajectories of both feet were 2.23 mm (±0.64) (left foot) and 2.72 mm (±0.83) (right foot) along the medial–lateral direction, and 9.17 mm (±1.98) (left foot) and 11.19 mm (±2.98) (right foot) along the anterior–posterior direction. The results are superior to those reported in previous relevant studies, and demonstrate that our proposed approach can be used for accurate foot plantar COP trajectory estimation. This study could provide an inexpensive solution to fall risk assessment in home settings or community healthcare center for the elderly. It has the potential to help prevent future falls in the elderly.

**Keywords:** postural control; falls in the elderly; fall risk assessment; low-cost instrumented insoles; foot plantar center of pressure

#### **1. Introduction**

Postural control refers to the control and maintenance of body's center of mass (COM) within the base of support during static or dynamic activities [1]. It has drawn much attention from research community in recent decades. The main functional goals of postural control include postural orientation and postural equilibrium, both of which require the integration of sensory information from visual, vestibular and somatosensory systems to stabilize the body and coordinate the movement strategies [2]. Therefore, postural control is a complex skill based on the interaction of dynamic sensorimotor processes [3], and can be challenging for people with deficits in sensory functions. For example, elderly who suffered age-related degeneration in sensory systems [4,5] and people with pathological conditions (such as cerebral palsy [6], stroke [7] and Parkinson's disease [8,9]) were found to be associated with impaired postural control.

The foot plantar center of pressure (COP) has often been used for quantitative assessment of postural control. For example, Rocchi et al. examined 14 COP measures by using the principle component analysis and found that five of these measures, including root mean square distance, mean velocity, principal sway direction, centroidal frequency of the power spectrum, and frequency dispersion, can effectively reflect the postural control mechanism among patients with Parkinson's disease [9]. Liu et al. used the velocity of COP trajectory during quiet upright standing to quantify the intensity of postural sway among young adults, healthy old adults and fall prone old adults [10]. Lafond et al. suggested that the velocity of COP trajectory was the most reliable measure for assessing postural stiffness among the elderly [11]. Biswas et al. have shown that the characteristics of COP can enhance the predictive power to a constructed index for dynamic stability [12]. In other studies, COP measures were also used to assess postural stability among stroke patients [13], patients with post-stroke hemiparesis [14] and patients with rheumatic disease [15]. In a recent study, Johansson et al. investigated how the COP sway can be used to predict future falls. They found that fall risk increased by 75% for participants with the COP sway lengths ≥ 400 mm during quiet standing with eyes open. They also suggested that fall risk could almost be doubled if the sway lengths ≥ 920 mm during quiet standing with eyes closed [16].

Conventionally, the COP trajectory is measured by force plates or force mapping systems [17]. However, such systems are restricted in laboratory settings. As such, they cannot be used to assess postural control in daily activities. To address this, a variety of insole-based plantar pressure measurement systems have been developed. Some of them are commercially available, such as the F-scan measurement system (Tekscan, Inc., Boston, MA, USA) and Novel Pedar system (Novel Inc., Kirkland, WA, USA). These systems allow the COP trajectory to be captured in a more extended space compared to force plates and force mapping systems. However, they have to rely on cables for data acquisition and power supply, which makes them obstructive and compromise their wearability. More importantly, these systems are too expensive for personal or daily use in home settings.

Providing inexpensive and wearable solutions for postural control assessment will provide vital information of fall risks, and thus could be useful for preventing future falls especially among elderly [18]. Recent advancement in microelectronics technology makes such wearable solutions feasible. For instance, Balaga et al. proposed a method that used a single body-fixed inertial sensor to quantitatively describe the postural control strategy during the lying-to-sit-to-stand-to-walk transfer tasks [19]. Similarly, wearable sensors (including those integrated in smart phones) were reported to be useful for fall risk assessment [20–23]. Many low-cost instrumented insole systems have also been developed [24–33]. Among the existing low-cost insole systems, different sensor technologies were implemented, such as force sensitive resistors (FSRs) [25,33], fabric or textile pressure sensing arrays [28], or piezoelectric sensors [32], etc. The numbers of pressure sensors or sensing arrays also varied among different studies, ranging from four FSRs [31] to 48 pressure sensor arrays [32].

Regardless of the sensor types and the number of sensors, the "weighted-mean approach" was most commonly used to estimate COP trajectories. In the weighted-mean approach, the force/pressure detected by each sensor was weighted by its corresponding coordinate (i.e., sensor location) and then summed. The COP trajectory was calculated by dividing the sum of weighted force/pressure by the overall force/pressure [26–30,32,33]. The accuracy of the weighted-mean COP trajectory estimation approach is not robust. In particular, when implementing the weighted-mean approach, the accuracy of COP trajectory estimation is dependent on the areas covered by the pressure sensors, which will be a potential source for errors. In other words, the weighted-mean approach depends on the locations and numbers of the pressure sensors. Theoretically, more sensors or sensor arrays implemented in the insole will result in higher accuracy. However, more sensors or sensor arrays always means high cost and system complexity, which makes it practically infeasible today.

In the present study, we aimed to develop and evaluate low-cost instrumented insoles for the estimation of foot plantar COP trajectories. In order to address the limitations of the weighted-mean approach, we proposed a novel individual-specific nonlinear model. This model was expected to accurately estimate foot plantar COP trajectories using the data from a small number of low-cost FSR sensors. An experiment involving a variety of postural control assessment tasks was carried out to provide data for least square error approximation and model coefficients specification. The accuracy of the foot plantar COP trajectories estimation by this model was examined by comparing the estimated COP data with the reference data from a commercial insole-based plantar pressure measurement system. Accuracy comparisons were also carried out between the proposed nonlinear model and the weighted-mean approach.

#### **2. Materials and Methods**

#### *2.1. Hardware Design*

The block diagram of the proposed instrumented insole is shown in Figure 1. It mainly consists of three parts: a 12 integrated FSR (FSR402, Interlink Electronics, Los Angeles, CA, USA) based insole, a lower-shank mounter block, and a PC end graphic user interface (GUI). Each FSR402 sensor has a 12.7 mm diameter sensing area made of fiberglass resin that is attached to a base 18.1 mm in diameter. The lower-shank mounter block consisted of a Micro-Computer Unit (MCU, STM32 32-bit ARM Cortex, ARM, Ltd., Cambridge, UK), a Bluetooth module (HC-06, Wavesen Co. Ltd. Guangzhou, China), a customized A2D module and a battery. Twelve FSRs were strategically adhered onto a silicone made insole, the size of which was US 9. The layout of the FSRs is depicted in Figure 2. This layout is similar to what has been suggested by Howell et al. [34], where the 12 sensors can cover the important foot plantar pressure distribution areas such as great toe, metatarsophalangeal joint, arch of the foot, and heel. An insole coordinate system was used to help identify the actual location of each sensor. As illustrated in Figure 2, the *x*-axis is the tangent line to the bottom edge of the insole, which defines the medial–lateral direction (ML); and the *y*-axis is the tangent line to the left edge of the insole, which defines the anterior–posterior (AP) direction. The origin is the intersection between the *x*-axis and the *y*-axis.

Prior to usage, each FSR was calibrated and conditioned following the techniques suggested by Hall et al. [35]. This was to eliminate the creep effect (i.e., the change of FSR resistance over prolonged time) and minimize the hysteresis effect. After that, each FSR sensor was connected to a Voltage-to-Current (V2C) converting circuit, as recommended by the manufacturer [36]. The thickness of the connecting cable is only 0.4 mm so that it will not lead to uneven surface of the insole. The V2C converting circuit converts the FSR resistance value to an inverse voltage output, which were subsequently converted into readable voltage output through a 10-bit analog-2-digital (A2D) module. The force measured by each sensor was obtained based on its voltage output following a fourth order polynomial equation, as suggested by Hall et al. [35]. The data were transmitted wirelessly by a Bluetooth module to a PC end, where the GUI was designed by a customized MATLAB script to visualize the pressure output in real time.

**Figure 1.** The block diagram of the instrumented insole.

**Figure 2.** The layout of the 12 Force Sensitive Resistor (FSR) sensors and the corresponding coordinates.

#### *2.2. The Experiment*

The experiment was carried out for model development and validation. A convenience sample of 10 young (age = 22.6 ± 1.5 years, height = 173.8 ± 4.4 cm, weight = 64.9 ± 4.4 kg) and 10 older adults (age = 65.7 ± 3.4 years, height = 167.5 ± 3.9 cm, weight = 65.2 ± 2.9 kg) were recruited from the local community. These participants all had the shoe size of US 9. The young participants were between 21 and 24 years old. The elderly participants were all above 60 years old and living independently, and could perform the activities of daily living (ADLs) without external assistance. Both young and elderly participants reported no injuries, illness, or medical surgery history. All participants signed the informed consent form approved by the Shenzhen University ethics committee.

Previous research has demonstrated that the F-scan system can measure the COP trajectory accurately [37]. Therefore, the foot COP trajectories obtained from the F-scan system were considered as the reference measurement (i.e., the response data). Prior to the experiment, the F-scan sensor sheets were tailored to make sure it had the same shape and size as the instrumented insoles. They were then pasted underneath the instrumented insoles (Figure 3a). Cautious actions were taken before data collection to make sure that the instrumented insoles and F-scan sensor sheet were adhered evenly and firmly. The insoles were inserted into a pair of sports shoes (size US 9), as depicted in Figure 3b.

**Figure 3.** (**a**) the F-scan sensor sheets (green) were tailored and adhered underneath the instrumented insole (blue); (**b**) the insoles were inserted into a pair of sports shoes.

Participants were asked to wear the shoes with the instrumented insoles. Both the lower-shank mounted block of the instrumented insoles and the signal box of the F-scan system were attached to the lower shank (Figure 4). Then, participants were instructed to perform a variety of postural control assessment tasks including (1) quiet standing with open eyes, (2) quiet standing with closed eyes, (3) standing up from a chair with armrests (wooden chair, seat height 435 mm, armrests height 250 mm), (4) sitting down to a chair with armrests, (5) standing up from a chair without armrests (wooden chair, seat height 435 mm), and (6) sitting down to a chair without armrests. During quiet standing trials, participants were asked to stand quietly with a self-chosen comfortable posture for 10 s. During standing up and sitting down trials, participants were asked to keep the feet on the ground when they started standing up or sitting down. When performing standing up and sitting down tasks with the chair with armrests, participants were instructed to use the armrests to help the body ascending and descending. These tasks were performed in a random order. Three trials were collected for each task. Each of these trials was separated to form three blocks of trials, which would be used for cross validation purposes. Prior to data collection, participants were given approximately 5 min practice to get familiar with these tasks. The activation of the F-Scan system and our proposed instrumented insoles is initiated with the application of external pressure. Thus, these two insole systems were synchronized by manually applying external pressure onto them at the same time. The sampling frequency was set at 50 Hz.

**Figure 4.** Both the lower-shank mounted block of the instrumented insole and the signal box of the F-scan system were attached to the lower-shank.

We noted that the instrumented insole was placed in between the foot and F-scan sensor sheet. The overlapping between the instrumented insole and F-scan sensor sheet might modify the foot plantar pressure distribution and induce errors in F-scan measurement. Such errors might compromise the accuracy of the reference data. A test was carried out to examine the possible errors caused by the overlap. One male volunteer (age: 31, height: 180 cm, weight: 65 kg, shoe size: US 9) participated. Firstly, the participant was asked to wear the experimental shoes that had both the proposed instrumented insole and F-scan sensor sheet (overlapping condition), and perform the above-mentioned postural control assessment tasks. Each task was performed 20 times in a random order. Then, the participant wore the same shoes with F-scan sensor sheet only (no overlapping condition) and performed each postural control assessment task 20 times in a random order. Two COP parameters including COP range and mean COP were compared between the 'overlapping' and 'no overlapping' conditions by using *t*-tests. As shown in Table 1, when the level of significance was set at 0.05, no significant differences existed in the COP parameters between the 'overlapping' and 'no overlapping' conditions. This suggests that the possible errors due to the overlapping did not significantly affect the COP estimation results. Thus, we used the F-scan measurement data in the overlapping condition as the reference data in the following analysis.


**Table 1.** Comparisons of COP parameters between the 'overlapping' and 'no overlapping' conditions.

#### *2.3. The Individual-Specific Nonlinear Model for COP Estimation*

The foot plantar COP is practically defined as the centroid of all the external forces acting on the plantar surface of the foot [38]. Therefore, in the previous weighted-mean approach, the COP trajectories were estimated by a spatial average of all the forces measured by pressure sensors. However, with a reduced number of pressure sensors, the external forces acting on the foot plantar might not be registered completely. Thus, errors might be induced.

To address the limitation of the weighted-mean approach, a multivariate nonlinear regression model was proposed where the FSR sensor outputs were considered as the predictor data. Similar to the weighted-mean approach, in the proposed model, the output of FSR sensor was weighted and summed to determine COP locations (Equations (1) and (2)). However, instead of using the sensor locations, the model coefficients (i.e., weighting parameters) were determined by a least square error approximation process. The sensor locations (i.e., the coordination of each sensor defined in Figure 2) were only used to set the initial values of the model coefficients. The least square error approximation process would tune the model coefficients to achieve improved COP trajectory estimation capability. This model is as follows:

$$\mathbf{X\_{COP}} = \frac{\sum\_{i}^{12} \mathbf{C}\_{i}^{\mathbf{x}} \mathbf{F}\_{i}}{\mathbf{C}\_{\text{tot}}^{\mathbf{x}} \mathbf{F}\_{\text{tot}}},\tag{1}$$

$$\mathcal{Y}\_{COP} = \frac{\sum\_{i}^{12} \mathcal{C}\_{i}^{y} F\_{i}}{\mathcal{C}\_{tot}^{y} F\_{tot}},$$

where *XCOP* and *YCOP* are the COP coordinates defined in the coordinate system (Section 2.1). *XCOP* represents the COP location along the medial–lateral direction, *YCOP* represents the COP location along the anterior–posterior direction. *Cx <sup>i</sup>* , and *<sup>C</sup><sup>y</sup> <sup>i</sup>* (*i* = 1, 2, 3 ... 12) are the model coefficients that weight each sensor output (*Fi*). *Cx tot* and *<sup>C</sup><sup>y</sup> tot* are the model coefficients that weights the sum of all forces registered by FSRs. *<sup>F</sup>tot* is the sum of all forces registered by FSRs (*Ftot* <sup>=</sup> <sup>12</sup> ∑ *i Fi*). *C<sup>x</sup> <sup>i</sup>* , *<sup>C</sup><sup>y</sup> <sup>i</sup>* , *Cx tot* and *Cy tot* are determined through the least square error approximation process.

An iterative least square error approach was carried out to determine the model coefficients. The model coefficients were written in the vector form as:

$$
\widehat{\mathbf{C}}^{\widehat{\mathbf{x}}} = (\mathbf{C}\_i^{\mathbf{x}}; \mathbf{C}\_{tot}^{\mathbf{x}}),
\tag{3}
$$

$$
\widehat{\mathbf{C}^y} = (\mathbf{C}\_i^y; \mathbf{C}\_{\text{tot}}^y)\_\prime \tag{4}
$$

The reference measurement (i.e., the foot COP trajectories obtained from the F-scan system) in the vector form was as follows:

$$X\_{\rm COP} = (X\_{\rm COP,1}; X\_{\rm COP,2}; \dots X\_{\rm COP,m}),\tag{5}$$

$$\mathbf{Y}\_{\text{COP}} = (\mathbf{Y}\_{\text{COP},1\prime}\mathbf{Y}\_{\text{COP},2\prime}\dots\mathbf{Y}\_{\text{COP},m})\_{\prime} \tag{6}$$

where *m* is the number of observations in the training data (which related to the time taken in each postural control assessment task). Similarly, the predictor data (*F*) can be written as a *m* × *i* matrix (*i* is the number of FSR sensors) as follows:

$$F = \begin{pmatrix} F\_{11} & F\_{12} & \cdots & F\_{1i} & \sum\_{i}^{12} F\_{1i} \\ F\_{21} & F\_{22} & \cdots & F\_{2i} & \sum\_{i}^{12} F\_{2i} \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ F\_{m1} & F\_{m2} & \cdots & F\_{mi} & \sum\_{i}^{12} F\_{mi} \end{pmatrix} = \begin{pmatrix} \mathbf{F}\_{1}^{T} & \sum\_{i}^{12} \mathbf{F}\_{1} \\ \mathbf{F}\_{2}^{T} & \sum\_{i}^{12} \mathbf{F}\_{2} \\ \vdots & \vdots \\ \mathbf{F}\_{m}^{T} & \sum\_{i}^{12} \mathbf{F}\_{m} \end{pmatrix}. \tag{7}$$

The goal of the approximation process is to find *C <sup>x</sup>* and *<sup>C</sup> <sup>y</sup>* that can minimize the square error between the reference measurements and the model predicted values. Let *E<sup>C</sup> <sup>x</sup>* <sup>=</sup> *XCOP* − *FC x* T *XCOP* − *FC x* and *E<sup>C</sup> <sup>y</sup>* <sup>=</sup> *XCOP* − *FC y* T *XCOP* − *FC y* , so

$$
\widehat{\mathbf{C}}^{\mathbf{x}^\*} = \underset{\widehat{\mathbf{C}}^{\mathbf{x}^\*}}{\text{arg min}} \, E\_{\widehat{\mathbf{C}}^{\mathbf{x}^\*}} \tag{8}
$$

$$
\widehat{\mathcal{C}}^{\mathfrak{g}^\*} = \underset{\widehat{\mathcal{C}}^{\mathfrak{g}}}{\text{arg min}} \underset{\widehat{\mathcal{C}}^{\mathfrak{g}}}{\text{min}} \,\tag{9}
$$

In multivariable calculus, to find the minimum value of *E<sup>C</sup> <sup>x</sup>* and *<sup>E</sup><sup>C</sup> <sup>y</sup>* , it requires solving the following partial derivative functions [39]. This process went iteratively until the closed form optimal solution for *C <sup>x</sup>* and *<sup>C</sup> <sup>y</sup>* were found:

$$\frac{\partial E}{\partial \widehat{\mathbf{C}^{\widehat{\mathbf{x}}}}} = 2\mathbf{F}^{T} \left( \mathbf{F}\widehat{\mathbf{C}^{\widehat{\mathbf{x}}}} - \mathbf{X}\_{\mathrm{COP}} \right) = \mathbf{0},\tag{10}$$

$$\frac{\partial E\_{\widehat{C}}}{\partial \widehat{C}^{\theta}} = 2F^{T} \left( F \widehat{C}^{\theta} - \mathbf{Y}\_{\text{COP}} \right) = 0. \tag{11}$$

Three-fold cross validation was implemented for model evaluation. The data from the postural control assessment tasks were equally assigned into three groups, and each group contained data from tasks (1)–(6) (as described in Section 2.2). For each evaluation procedure, two groups were selected as the training set while the other group as the evaluation set. This repeated three times until all the group combinations were tested. Then, the mean value of root mean square errors (RMSE), the correlation coefficients (CC), the maximum error (MaxE) and minimum error (MinE) between the estimated COP trajectories and the reference measurements were calculated. In addition, in order to determine the statistical significance of the CC, the *p*-values were examined for CC. In addition, the COP trajectory was also estimated by the weighted mean approach. Similarly, three-fold cross validation was implemented to evaluate the weighted mean approach. Comparisons were carried out between the nonlinear model and weighted mean approach in RMSE, CCs, MaxE and MinE.

#### *2.4. Graphic User Interface (GUI)*

The GUI is shown in Figure 5. The left-hand side panel shows the pair of instrumented insole and the locations of FSR sensors. Each red dot presents the instant sensor pressure, the size of which corresponds to the pressure magnitude. The green dot indicates the estimated location of the COP. The right-hand side panel shows the COP trajectories along the time. It also shows the pressure output of each FSR sensor and instant *Ftot* normalized by body weight (BW).

**Figure 5.** The GUI developed in the MATLAB to show the output of each sensor, *Ftot* normalized by body weight (BW), and the estimated foot plantar COP trajectory.

#### **3. Results**

Figure 6 shows an example of COP trajectory estimation by the proposed nonlinear model. The COP trajectories along the ML direction and AP direction were plotted against the reference measurements. This example shows that the estimated COP trajectories were trending closely with the reference, suggesting an accurate estimation.

Table 2 shows the RMSE, CC, MaxE and MinE results of the COP trajectories calculated by the proposed nonlinear model compared with the reference measurements. The overall mean and standard deviation were also summarized. The RMSE reflects estimation accuracy. For the ML COP, the mean RMSEs were 2.23 (±0.64) and 2.72 (±0.83) for the left and right foot, respectively. For the AP COP, the mean RMSEs were 9.17 (±1.98) and 11.19 (±2.98), respectively. The RMSEs along the AP were larger than that along the ML direction because the COP trajectories typically have a larger moving range along the AP direction. The small RMSEs suggested that the proposed COP trajectory estimation model had high COP trajectory estimation accuracy.

The CC accounts for the similarity of the estimated COP trajectory time series, compared to the reference measurements. The mean value of CCs was 0.91 (±0.05) to 0.93 (±0.02) along the ML and AP direction, respectively. In addition, the *p*-values of each CC were all smaller than 0.0001, indicating that the correlation is highly significant.

The MaxE and MinE indicate the least and best estimation each model can achieve. The MaxE for ML COP was 26.119 mm and 25.254 mm for the left and right foot, respectively; for AP COP, it was 124.86 mm and 116.18 mm for the left and right foot, respectively. As for the minimum error, both the weighted mean approach and nonlinear model can achieve a high accuracy with the minimum error less than 1 mm.

Figure 7 shows the comparison of the RMSE between the nonlinear model and weighted mean approach. For all participants, the RMSE between the estimated COP by the nonlinear model and the reference measurements are smaller compared to their counterparts estimated by the weighted mean approach. Figure 7 also shows the CCs. Overall, the weighted mean approach yielded a fairly good CC, which was approximately 0.77–0.85 along the ML and AP direction, respectively. However, the nonlinear model yielded a higher CC. These results indicate that the COP trajectories estimated by the nonlinear model have a more similar trending with the reference COP.

The least square error approximations were completed by a customized MATLAB script running on a laptop computer (Intel® Core (TM) i7-7500U 2.7GHz, 8.00GB RAM, 256GB HD, Windows x64 operation system, HP Inc., Palo Alto, CA, USA). The number of iterations in least square error approximation was between 5–8 (Mean ± SD = 6.5 ± 1.0 iterations) across different trials. It took 0.053–0.124 s (Mean ± SD = 0.112 ± 0.005 s) to complete the approximation procedure.

**Figure 6.** Representative plot of the comparison of COP trajectory estimation by the nonlinear model and the reference data. The data were obtained from the left foot and for the different tasks: (1) quiet standing with open eyes, (2) quiet standing with closed eyes, (3) standing up from a chair with armrests, (4) sitting down to a chair with armrests (task 3 and 4 have multi-contacts phases where participant's hands are in contact with armrests or the seat) (5) standing up from a chair without armrests, and (6) sitting down to a chair without armrests.



**Figure 7.** Comparisons of the RMSE and CC between the nonlinear model and the weighted mean approach.

#### **4. Discussion**

This study presents an individual-specific nonlinear model that can help estimate the foot plantar COP trajectories with an instrumented insole. Among the 20 participants involved in this study, the average RMSE between the estimated COP trajectories by the proposed model and the reference measurements was less than 3 mm along the medial–lateral direction of the foot, and less than 12 mm along the anterior–posterior direction. In addition, the estimated COP trajectories by this nonlinear model established high correlation coefficients with the reference measurements (0.91–0.93 with *p*-values all less than 0.0001). To our knowledge, the results are the smallest errors and highest correlation coefficients reported in relevant studies where the foot plantar COP trajectories were estimated by a small number of low-cost pressure sensors [28,29]. For instance, Shu [28] reported a mean relative difference of 7.6 and 9.9 mm between the estimated COP trajectories and the reference measurements during normal standing and standing on one leg. Dyer et al. [29] reported the RMSE ranging from 7 mm to 24 mm during locomotion. Overall, the results suggested that the proposed nonlinear model had excellent COP trajectory estimation capability.

Previously, the weighted mean approach was the most commonly used method to estimate the foot plantar COP trajectory by instrumented insoles [40]. However, as mentioned earlier, this approach may not lead to desirable estimation, especially when the number of low-cost pressure sensors is small. To address the limitation of the weighted mean approach, we proposed the nonlinear model here and used this model to estimate COP trajectories during a variety of postural control assessment tasks. Meanwhile, to identify the improvement of the proposed nonlinear model, the weighted mean approach was also used for COP trajectory estimation by using the same set of data. The results show that the mean RMSE between the estimated COP trajectories by the weighted mean approach and the reference measurements were ~4–6 mm along the ML direction, and ~18–20 mm along the AP direction. These nearly doubled the RMSEs obtained by the nonlinear model. The correlation coefficients were ranging from 0.80 to 0.86, which were also lower than those calculated by the nonlinear model approach. These results suggested that the proposed nonlinear model can lead to improved COP trajectory estimation compared to the weighted mean approach.

There might be two reasons that can help explain why the nonlinear model works better than the weighted mean approach with a small number of sensors. First, in the weighted mean approach, the accuracy depends largely on the numbers of sensors and their predefined locations. As we discussed earlier, the small number of sensors might induce errors. Different sensor placement strategy will also influence the accuracy of COP trajectory estimation. However, in the nonlinear model, least square error approximation will help rectify the errors due to the small number of sensors and their placement strategy. Another reason is that the weighted mean approach cannot address individual differences, which may induce additional intra-subject errors. In addition, the least square error approximation process in the nonlinear model is dependent on individual data from the sensors. In other words, the model coefficients were determined by each participant's own experimental data. Thus, the proposed nonlinear model is individual-specific and is able to address individual differences.

This study demonstrated that the foot plantar COP can be accurately estimated by a small number of low-cost pressure sensors. Many instrumented insole systems have been developed and are even commercially available. Clinically, more accurate COP estimation help better postural control assessment. However, there is always a trade-off between the number of sensors and estimation accuracy. The increased number of sensors will complicate the instrumented insole system and possibly compromise the reliability of the whole system. In addition, a larger number of sensors or sensor array will increase the cost substantially and make it unaffordable for the home-dwelling elderly. Therefore, an instrumented insole with a small number of low-cost sensors like what we proposed in the present study has its own merits.

From a practical point of view, this study can benefit the home-based postural control assessment for elderly and patients with pathological conditions. The impaired postural control is considered as an important risk factor of falls [41,42]. Though consensus is still lacking, postural control parameters have been widely suggested as indicators of fall risks [43–46]. Based on these, the authors believe that this can benefit the fall prevention research.

There are still some limitations in this study. First, this study mainly focused on postural control assessment during static stance and sit-to-stand transitions. Although previous research suggested that increased postural sway during stance can be a risk factor for prospective falls in community-dwelling elderly individuals [10], future work needs to be carried out to test the validity of this nonlinear approach during other activities of daily living, such as walking and stair negotiation. Second, we did not examine insole size other than US 9 just for convenience. However, as the proposed COP estimation model in the present study was individual-specific, we believe that this model would be applicable for other insole sizes as well.

#### **5. Conclusions**

Falls are still a major safety and health problem among aged population. Fall risk assessment is an effective approach to reduce fall accidents among the elderly. A substantial number of falls in the elderly result from loss of balance. Thus, the plantar COP, as an indicator of postural control performance, is an important fall risk assessment parameter. This study presented a low-cost instrumented insole system that uses a nonlinear model for COP trajectory estimation. Results show that this system is able to provide accurate COP trajectory data. Compared to traditional COP trajectory estimation approaches (i.e., weighted mean approach), the proposed nonlinear model performed better in terms of estimation accuracy. Based on this, we suggest that the proposed instrumented insole system could serve as an inexpensive solution to fall risk assessment in home settings or community healthcare centers for the elderly. It has the potential to help prevent future falls in the elderly.

**Acknowledgments:** This work was supported in part by the Natural Science Foundation of China (11702175; 31570944), the Natural Science Foundation of Guangdong Province (2015A030313553; 2016A030310068), and the Science, Technology and Innovation Committee of Shenzhen City (JCYJ20160422145322758; JCYJ20150525092940994). The authors would like to thank Wenzhen Chen and Jialun Cai for their help during data collection.

**Author Contributions:** X.H. and X.Q. conceived and designed the experiments; X.H., D.P. and J.Z. performed the experiments; D.P., Z.S. and X.H. analyzed the data; and X.H., Z.S. and X.Q. wrote the paper.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Review* **Shoe-Insole Technology for Injury Prevention in Walking**

#### **Hanatsu Nagano and Rezaul K. Begg \***

Institute for Health and Sport (IHES), Victoria University, Melbourne, VIC 3032, Australia; hanatsu.nagano@vu.edu.au

**\*** Correspondence: rezaul.begg@vu.edu.au; Tel.: +61-3-9919-1116; Fax: +61-3-9919-1242

Received: 29 March 2018; Accepted: 29 April 2018; Published: 8 May 2018

**Abstract:** Impaired walking increases injury risk during locomotion, including falls-related acute injuries and overuse damage to lower limb joints. Gait impairments seriously restrict voluntary, habitual engagement in injury prevention activities, such as recreational walking and exercise. There is, therefore, an urgent need for technology-based interventions for gait disorders that are cost effective, willingly taken-up, and provide immediate positive effects on walking. Gait control using shoe-insoles has potential as an effective population-based intervention, and new sensor technologies will enhance the effectiveness of these devices. Shoe-insole modifications include: (i) ankle joint support for falls prevention; (ii) shock absorption by utilising lower-resilience materials at the heel; (iii) improving reaction speed by stimulating cutaneous receptors; and (iv) preserving dynamic balance via foot centre of pressure control. Using sensor technology, such as in-shoe pressure measurement and motion capture systems, gait can be precisely monitored, allowing us to visualise how shoe-insoles change walking patterns. In addition, in-shoe systems, such as pressure monitoring and inertial sensors, can be incorporated into the insole to monitor gait in real-time. Inertial sensors coupled with in-shoe foot pressure sensors and global positioning systems (GPS) could be used to monitor spatiotemporal parameters in real-time. Real-time, online data management will enable 'big-data' applications to everyday gait control characteristics.

**Keywords:** gait; insole; injury prevention

#### **1. Introduction**

Walking is a fundamental locomotor task essential to healthy, active living, but it is accompanied by injury risk, particularly among the senior population. Walking is a continuum of gait cycles repeated thousands of times daily, and suboptimal features of the gait cycle can increase the probability of injury. Older adults (particularly females), for example, are prone to falls-related acute injuries due to gait impairments [1,2], and impaired foot pressure control during the loading response can cause foot problems [3]. The knee joint can be also affected over time due to functionally undesirable weight bearing, possibly resulting in osteoarthritis (OA) [4]. While walking is critically important both from a functional perspective and to ensure adequate exercise, injury risks must be minimised by optimising some essential biomechanical features of the gait pattern. Biomechanical interventions for injury prevention should, however, fulfil certain practical requirements, including low cost, easy engagement, immediate effects, and little physical effort; otherwise, interventions are unlikely to be adopted voluntarily and maintained in the longer term [5]. Footwear interventions have the potential to satisfy these requirements.

Typically, shoes are constructed with a number of components, all of which can influence gait mechanics [6]. An elevated heel is a factor in lateral instability and may result in caution-related adaptations reflected in spatio-temporal parameters [6]. Compared to standard soles, hard soles can more effectively provide tactile sensation for quicker reactions to maintain balance. Footwear-collars improve balance due to increased tactile sensation around the ankle while reducing swing foot clearance [6]. The outsole provides the interface with the walking surface and affects the frictional demands of walking and associated risk of slipping [7]. In contrast, the insole has direct contact with the sole of the foot and directly controls foot pressure and ankle joint motion that, in turn, influences the individual's gait pattern [8]. While some features of 'safe shoes' tend to be avoided, such as firm shoe-lace fixation, particularly in individuals with impaired activities of daily living (ADL) [9], insole interventions have the potential to be more readily accepted due to their practicality when applied to various footwear types.

As summarised in Table 1, three main types of insole modification can be identified as having the potential to support safer walking. Until recently, most high-grade insoles were produced using custom-moulding, which was designed to accommodate the individual's foot shape and influence foot pressure distribution. While this approach has provided a springboard, sensor technologies are now available that can provide a highly detailed biomechanical analysis of foot pressure and gait patterns to considerably advance shoe-insole development. Such progress could revolutionise injury prevention. For example, three-dimensional (3D) motion capture systems (e.g., Optotrack, Vicon, Optitrack) can accurately model gait motions, which is useful for identifying suboptimal gait features. By utilising this sensor technology, insole development can be undertaken to optimise gait control.

Foot pressure mapping, another very useful in-shoe sensor technology (e.g., F-scan, Pedar), is often synchronised with 3D motion capture to reveal the foot's pressure distribution and centre of pressure (CoP) in real-time [10]. The advantages of insole sensor systems are portability and wireless communication. Foot pressure measurement can not only be utilised in developing new insoles but also has the potential to acquire and store data to record gait patterns. Body-mounted inertial sensors are a related technology with similar potential to sample gait parameters in the natural environment. The portability of sensor-based systems will be their essential advantage in future gait assessments and will gradually overcome the limitations of laboratory-based 3D motion capture systems.



CoP = centre of pressure.

In the current review, typical locomotor injuries are first explained and then insole developments and their significance are thoroughly discussed. Finally, the concept of a wireless gait measurement insole will be introduced and future directions in gait-related sensor technology outlined.

#### **2. Biomechanics of Locomotive Injuries Based on Gait Analysis**

In daily locomotion, both acute and overuse injuries should be considered. The primary cause of acute injury during locomotion is falls, particularly in the older population [1]. In contrast, certain types of ankle and knee pathology can be classified as overuse injuries due to the accumulation of negative gait features over time rather than a single traumatic event. This section introduces the biomechanics of falls and lower limb injuries primarily at the ankle and knee joints.

#### *2.1. Falls*

Approximately 33% of senior adults fall annually and up to 20% of cases lead to serious injuries [1,2,19]. Falls in this context can be defined as "an unintentional coming to the ground or the lower level due to balance loss" [20]. Biomechanically, falls occur when balance is disturbed and not able to be restored, with the consequence that the individual makes forceful contact with either the walking surface or surrounding objects. This twofold process comprises, therefore, (1) an event that disturbs balance and (2) failure in balance recovery.

Of the many balance-disturbing events, tripping has been identified as the leading cause of falls, accounting for up to 53% of all falls [21]. Tripping is due to physical contact of the swing foot with the walking surface or an object on it, which creates the momentum to significantly destabilise balance. To prevent tripping, therefore, swing foot clearance should provide a sufficient vertical margin, particularly at the mid-swing event—Minimum Foot Clearance (MFC)—illustrated in Figure 1. At MFC, the vertical margin of the swing foot from the walking surface is low (i.e., 1–2 cm) while moving at maximum speed and, as a consequence, causes high impact in the case of tripping [22]. Previous research by Moosabhoy and Gard suggests that, to prevent tripping, ankle dorsiflexion should be the most effective lower limb joint control strategy, whereby one degree of ankle dorsiflexion at MFC can be predicted to elevate the toe by .3 cm [23], a response that can significantly reduce the probability of tripping [24].

**Figure 1.** Minimum Foot Clearance (MFC).

Biomechanically, dynamic balance is defined by the relationship between body centre of mass (CoM) and base of support (BoS) [25]. In gait analysis, sensors, such as infrared light-emitting diodes (IREDs), light-emitting diodes (LEDs), and passive reflective markers, are usually attached to anatomical landmarks to model the subject's motion relative to a pre-registered laboratory coordinate system (*x*, *y*, *z*). In static conditions, when the CoM is preserved within BoS, balance is considered to be secure, whereas in dynamic conditions, including walking, extrapolated CoM (XCoM) is used rather than CoM, as in the equation below [26].

$$\text{\textbullet CoM} = \text{CoM position} + \frac{\text{CoM velocity}}{\sqrt{\frac{\text{gravity}}{l}}}$$

In the above equation, *l* indicates the distance between the ankle and the end of invertedpendulum movement, CoM [27]. XCoM has been considered to more accurately represent the positional threshold within BoS because CoM position by itself does not differentiate static CoM from fast moving CoM. The distance between XCoM and either the lateral or anterior-posterior boundary of the BoS, depending on CoM movement direction, is defined as the margin of stability (MoS) (Figure 2). A greater MoS indicates that balance is secure while a negative (<0) MoS indicates balance loss and that the body is falling. From an injury prevention perspective, the most important feature of these relationships is that CoM motion is highly dependent on foot CoP control [28], which, as described above, can be influenced by footwear manipulations.

**Figure 2.** Description of balance in the transverse plane. MoS = margin of stability; BoS = base of support; CoM = centre of mass; XCoM = extrapolated centre of pass; Relevant boundary depending on the direction of CoM movement.

#### *2.2. Foot-Related Conditions*

Older adults' footwear often has inappropriate features, perhaps due to prioritising comfort over safety [9]. This is reflected in the tendency to avoid shoes with fixation or heel counter, and sandals are commonly worn (22%) when they experience falls [9]. Improper footwear structures that cause undesirable foot pressure distribution and ineffective impact distribution can lead to chronic pain over time, which discourages active walking [29]. Footwear, including the insole, should be designed to promote walking mechanics that do not impose excessive stress on the foot and lower limb joints [3]. Foot problems often arise from inadequate weight bearing, which is reflected in undesirable ankle motion and foot pressure distribution [30,31]. Pressure distribution is the important element of gait control in preventing conditions, such as ulcer development, around the plantar areas, particularly among the diabetic population [32]. Fifteen million people in the U.S. have diabetes and 15–20% of those experience hospitalisation [3]. Another consequence of inadequate foot pressure distribution is Hallux valgus, which is especially common among high-heel wearers [14,33]. Hallux deformities are one of the major causes of foot pain [3]. Control of foot pressure is reflected in CoP movement, which is not only responsible for foot disorders but also controls dynamic balance. Inversion ankle sprain, for example, is a common acute injury caused by excessively lateral CoP [34], while Hallux valgus is a result of overpronation during mid to late stance [35]. As reported in previous studies [13,17] and as shown in Table 1, insole modifications can alter CoP control and foot pressure distribution,

possibly reducing the incidence of debilitating foot conditions caused by inadequate CoP and foot pressure control.

Foot segment orientation is controlled by ankle motion, including dorsiflexion/plantarflexion and eversion/inversion. Pronation is the combination of dorsiflexion and eversion while inversion and plantarflexion form supination [36]. For adequate loading, the foot should be pronated in the early stance phase to absorb impact while quickly accommodating the walking surface environment [36,37]. Foot contact sometimes occurs with a supinated ankle but immediately starts to pronate, the essential countermovement to maximise the range of pronation, with accompanying tibial internal rotation [36,38]. Towards toe-off via foot-flat, the foot supinates, controlled first by eccentric work until foot-flat followed by the concentric work of supinators approaching toe-off [36,39]. This series of functional ankle motions most efficiently oscillates foot contact impact through the stance phase and into forward progression. These ankle motions are responsible for efficient weight transfer but also affect CoP movement, which can be measured using either force plates or in-shoe foot pressure monitoring systems.

Ankle kinematics can be monitored using 3D motion capture systems. For modelling the foot, a typical marker setup includes the toe (the distal and superior surface of the foot), second metatarsal head, fifth metatarsal head, lateral and medial malleolus, and heel. To visualise ankle motion, the tibia should also be modelled by incorporating lateral and medial epicondyles in addition to the lateral and medial malleolus at the ankle (Figure 3).

**Figure 3.** Illustration of marker setup for gait analysis; knee adduction moment due to shank external rotation.

#### *2.3. Knee Osteoarthritis*

Knee osteoarthritis (OA) is a painful condition that discourages people from walking and increases the risk of falling [39]. OA is frequently associated with pain caused by micro-fractures due to bone-to-bone collision with reduced articular cartilage. Knee adduction moment is the kinetic marker most reliably related to the progression of OA, especially at the medial compartment [40,41]. A comprehensive review suggested that approximately 10–15% of the senior population show clinical evidence of OA [42], but it is possible that gait modification interventions using insoles slow the progression of OA if knee adduction moment can be reduced [43].

There are two peaks in knee adduction moment during a gait cycle. The first peak is associated with OA and timing approximates the first Ground Reaction Force (GRF) peak and maximum knee flexion [44]. Reducing knee adduction moment is, therefore, the challenge in OA prevention. Biomechanically, this can be achieved by (1) reducing GRF and (2) creating a shorter moment arm by reducing external shank rotation in the frontal plane after foot contact [45,46]. Shock absorption at the heel could reduce peak GRF, while tibial realignment in early stance is associated with reduced moment arm and knee adduction moment. Biomechanical gait data obtained using a combination of 3D motion capture systems and force plates (i.e., kinematic and kinetic analysis) can compute knee joint kinetics via an analysis technique known as inverse dynamics. As introduced later, a lateral wedge insole may be able to control CoP and realign the tibia to reduce knee adduction moment.

In summary, safe walking is essential to maintaining mobility without injury and, as shown in Table 1, shoe insoles can control gait functions to optimise walking mechanics [8,47]. In the current review, the primary focus is insole modifications to improve gait function.

#### **3. Shoe Insole Technology for Gait Control and Injury Prevention**

Shoe-insoles provide the interface between the foot and the footwear. Potential modifications include (1) modifying insole geometry (i.e., changing ankle angle) [18]; (2) increasing contact area (i.e., custom-moulding) [11]; (3) adjusting resilience (i.e., lower for shock absorption and greater for reusing mechanical energy) [12]; (4) treating insole surface (e.g., texture installation) [48]; (5) providing assistive support (e.g., heel counter) [49]; and (6) incorporating portable sensor systems, such as pressure sensors or inertial sensors [50]. Fundamental functions relating to reducing the above injury risks are as follows.

#### *3.1. Shock Absorption*

Impact energy at heel contact can be dissipated to minimise soft tissue damage, and softer materials are likely to reduce shock more effectively [51]. Ethylene vinyl acetate (EVA) foam is one of the more common shoe-insole materials, is usually used with a density range between 300 and 400 m/s3, and is ideal for semi-customised shoe-insole moulding in commercial production [52]. The choice of insole material characteristics, such as density, can be used to control elasticity and resilience.

Technology for shock absorption at the heel has been marketed commercially using materials such as gels. One biomechanical explanation for the effectiveness of shock absorption products is that they prolong the time from initial foot–ground contact until complete compression of the footwear. Although GRF and foot pressure distribution can be modified by construction materials [53], some studies have reported that shock absorbers may not reduce injury risks [54]. In addition to shock absorbing materials, insole structure and hardness of the mid-sole are important for reducing load [51,55].

Shock absorption utilising either a material's elastic properties or a spring mechanism may be effective in storing mechanical energy at impact and then recovering the energy later in the stance phase toward toe-off as demonstrated by Zhang et al. using spring-loaded Axillary clutches [56]. Energy recovery can therefore engender more efficient walking by minimising the energy required at push-off. As detailed further in the following Section 3.2. Ankle, foot–ground impact can be viewed as an unexploited source of external energy because in human walking only 60–70% of impact energy can be oscillated through the loading response while the remainder is lost as vibration, sound, or heat [57,58].

#### *3.2. Ankle*

Ankle motions responsible for shock absorption during early loading are a combination of dorsiflexion and eversion (i.e., pronation) [36,59–61]. Flat-foot contact is not considered desirable for shock absorption because the interval from initial foot contact to foot flat determines impact distribution over time due to the eccentric work of dorsiflexors [37,38,61]. Ageing appears to be a factor in reduced dorsiflexion at foot contact [62].

A pronated ankle during early loading rotates the shank internally and triggers knee flexion [63–65], but both these kinematic adaptations have the potential to reduce knee adduction moment [51]. An over-pronation problem arises following mid-stance when the ankle should begin to supinate [36]. Given the anatomical constraints on foot kinematics during the stance phase, during early-loading foot orientation for shock absorption is important and after mid-loading over-pronation should be avoided.

Stored mechanical energy should be transferred throughout the loading response from the pronated ankle up to supinated orientation at mid-stance towards toe-off [36]. Quantification of mechanical energy transfer is possible by calculating the recovery rate (see below), the percentage of mechanical energy transferred to oscillate the loading response from heel contact to toe-off [64–66].

$$\text{Recovery Rate} \left( \% \right) = 100 \times \left[ \Delta \text{KE} + \Delta \text{PE} - \Delta (\text{KE} + \text{PE}) \right] / \left( \Delta \text{KE} + \Delta \text{PE} \right)$$

where ΔKE = increase in kinetic energy; ΔPE = increase in potential energy; Δ(KE + PE) = increase in the sum of kinetic energy and potential energy, all measured in the double support phase of the gait cycle based on CoM kinematics.

Step-to-step CoM transition during double support requires mechanical energy to produce GRF to continue walking [67]. More efficient loading is possible by utilising the impact at heel contact to initiate toe-off without dispersing the mechanical energy transferred to other lower limb joints (i.e., the knee). A higher recovery rate is, therefore, considered advantageous in reducing the heel contact forces transferred to other lower limb joints. For efficient loading, insole geometry can be re-designed to support pronation of the subtalar joint at heel contact while ensuring no disturbance to later ankle supination [36,59–61,68]. In the presence of foot deformity or other pathological conditions, however, careful consideration is required prior to any insole modification. For example, a custom-moulding insole modification may be necessary for seriously deformed feet, such as deformation due to neuropathy in diabetic patients [69]. Levinger et al. [45,46] suggested that a lateral wedge insole may be effective in reducing knee adduction moment by supporting pronation during early stance. They have, however, also stated that a lateral wedge insole may not be suitable for an already pronated foot.

#### *3.3. Foot Pressure Control*

The most common method for foot pressure distribution by shoe-insoles is to increase the contact area between the foot and the insole surface [70]. This can be achieved by custom-moulding the shoe insole to accommodate differences in foot shape [11]. Semi-customisation is also possible using an insole surface material that gradually adapts to the foot's shape. Foot CoP trajectory in a normal gait starts at the heel and ends at the toe, with a lateral curvature [71]. This lateral excursion seems to be a reflection of ankle supination during mid-stance.

As indicated above, balance control is determined by the relationship between CoM and BoS. A specially designed insole can control the CoP which changes CoM motion and it is particularly important to regulate excessive lateral CoP excursion to stabilise sideways balance [59]. A potential solution is an insole incorporating enhanced texture to consistently guide the optimal CoP path (Figure 4).

**Figure 4.** Example of texture installation to guide the CoP (WO2016015091A1—Injury Reduction Insole).

If the CoP is variable over multiple gait cycles, associated CoM motions could also become variable and dynamic balance control could become unstable. Ensuring adequate cutaneous stimulation to assist the optimal CoP path is, therefore, a possible approach to improving balance. As described below, texture installation may also assist balance control by improving the reaction speed to balance perturbations.

#### *3.4. Reaction Speed*

Stimulation of cutaneous receptors increases afferent feedback and may, therefore, decrease reaction time. Ideally, the CoP should travel through (i.e., stimulate) high sensitivity cutaneous receptors without excessive pressure [72] and if balance is lost, recovery can be faster. Stimulation of the plantar surface is possible either by incorporating enhanced texture or using vibration devices [15,73]. The findings from these previous studies of this manipulation, however, leave some doubt as to how effectively such augmented stimulation can enhance afferent feedback and proprioception [15,73,74].

Priplata et al. [15] tested the effect of vibrating insoles on balance control and found that older adults had improved dynamic balance due to plantar stimulation, but the study methodology was questioned by Lafond et al. [74]. One of their primary concerns was the methods employed for estimating balance, which is a more general problem arising from the lack of standardised biomechanical analysis methods to evaluate footwear effects on gait and balance. Similarly, some studies have reported enhanced proprioceptive reaction effects of textured insoles [73,75,76], while others have not [77], and these discrepancies may, again, have been due to methodology but also to the fact that the textured or vibration insoles tested were not the same [78]. If cutaneous receptors are stimulated more systematically rather than stimulated by randomly installing textures over the entire insole surface, gait and balance could be controlled more effectively [78]. A further methodological limitation of previous insole research has been sample size; however, an ongoing study by Hatton et al. [79] with a large sample is expected to reveal the effects of texture stimulation on gait patterns with greater confidence.

There are a few reports of texture installation that is not applied over the entire surface. For example, SoleSensor is a commercially available shoe-insole that takes advantage of tactile stimulation to enhance reaction speed [16] using only a narrow tube peripheral to the insole surface. When foot pressure shifts peripherally and balance is disturbed, stimulation is provided by the tube that increases reaction speed. Similarly, Ritchile et al. applied textures only to the medial portion of the insole and reported that it supported functionally important mid-foot supination [68]. Due to the variety of texture installations with respect to area, size, shape, and hardness, further research is necessary to confirm the most effective stimulation properties. Research evidence suggests, however, that stimulation of cutaneous receptors has considerable promise for enhancing gait and balance [78].

#### *3.5. Lateral Wedge Insole*

The lateral wedge insole is designed to stabilize the ankle in a more everted position, helping to more internally align the tibia [46]. This adaptation reduces the tibial moment arm from the GRF vector and consequently decreases knee adduction moment [40,41]. For the Varus deformity, an everted ankle reduces knee adduction moment, softening the compression of the medial knee structure between the femur and tibia. The degrees of eversion support are usually between 5◦ and 15◦ [80].

Despite these biomechanical theories, the lateral wedge insole effects on knee OA are controversial. Weinhandl et al. [41], for example, reported no apparent effects in a young group who wore a lateral wedge insole for one week. As with the stimulation insoles, the conflicting results may be attributable to differences between the lateral wedge insoles in each study. Paradoxically, if the optimum structure can be identified, lateral wedges appear to have potential as a knee OA treatment [81]. Sawada et al. [82] reported that an individual's foot alignment also determines the effectiveness of a lateral wedge insole in reducing peak knee adduction moment. Walking speed was also suggested to affect experimental results in terms of knee adduction moment, but this hypothesis has been rejected by one study [83]. An additional modification that may enhance the lateral wedge insole is arch-support [84]. Based on data from 90 participants [81], soft, rather than hard, lateral wedge insoles have also been found to be more effective.

An additional advantage of lateral wedge insoles is CoP control [45], although this feature has not been examined thoroughly in previous reports. As discussed, CoP excursion through the stance phase demonstrates a characteristic lateral curvature. This response is functional in leading the pronated ankle into a supinated position to efficiently oscillate mechanical energy from heel contact to toe-off. This deviation may, however, be associated with lateral balance disturbance and ankle inversion sprain [85]. Thus, eversion support has potential for injury prevention by regulating excessive lateral CoP excursion. Careful consideration is required when incorporating this feature into shoe-insoles to correct excessive Valgus knee so as not to disturb functional supination.

#### **4. Concerns for Custom-Moulding**

In populations with significant foot deformities, it may be difficult to wear conventional shoes and customised insoles may be required. Custom-made insoles can accommodate individual-specific foot shapes and maximise the contact area between the foot and insole [11,30]. This promotes foot pressure distribution, and for those with significantly deformed feet, customisation appears to be essential [86,87].

Caution is, however, required for custom-moulding for two reasons. First, every part of the foot does not contribute equally to weight bearing; for example, the mid-foot has been reported to have little responsibility [14]. It is, therefore, possible that some parts of the foot may be more vulnerable to foot pressure than others. The second caution is that customisation could reflect and accentuate negative foot control. If a misaligned foot is scanned and moulded without correction, progression of the deformity may advance due to inadequate foot posture and additional foot problems may arise as a consequence [29]. Lateral CoP excursion is associated with the risk of inversion sprain [34]. If a foot is susceptible to inversion sprain, for example, custom-moulding without CoP modification may not help or could even further increase the risk of injury. A lateral wedge assists foot pronation and reduces knee adduction moment, but the lateral wedge may not adequately assist the foot with excessive Valgus [45]. It can, therefore, be implied that a foot with excessive Valgus may benefit from a medial wedge rather than a lateral wedge. Furthermore, custom-moulding was found not to treat Hallux valgus by regulating hyperpronation of the subtalar joint during the later stance phase when weight is concentrated on the metatarsal area [35]. The results suggest, therefore, that custom-moulding alone may not correct inadequate ankle joint kinematics.

Although custom-moulding is important, the insole's fundamental geometry should support optimum foot pressure control and energy efficient loading by providing adequate pronation–supination coupling. Ideally, the thin insole surface layer should be individual-specific, while the layers below can be rigidly constructed to assist optimum foot control. Telfer et al. [31] described the potential of 3D foot scanning and 3D printing for custom-moulding. This is a more finely-tuned approach to insole design, and future customisation of insoles and footwear is likely to take advantage of this technology. Other recent studies have also reported that 3D foot scanning could be an effective method for custom-made footwear despite further research being required to more comprehensively test the precision of this technique, including an increased number of tested subjects and wider measurement parameters to define the foot segment [88,89].

#### **5. Gait Analysis for Insole Development**

While shoe-insoles assist walking, detailed gait analysis has not been widely utilized to test their effects on gait patterns. While pressure data are instructive, as mentioned above, gait analysis can record human walking precisely to identify various problems, including the risk of falling and balance control and lower limb joint (ankle and knee) biomechanics problems. The utility of 3D gait analysis in identifying the causes of foot and knee pain has been discussed by Rao et al. [29], and Menant et al. [6] used these techniques to determine how footwear features influenced gait biomechanics. Heel collars, for example, also improve balance by providing increased tactile sensation around the ankle via the extended contact area provided by the collar, while a high collar was found to increase the risk of tripping by reducing swing foot height at MFC. In addition to identifying condition-specific biomechanical parameters, such as those associated with knee pain described above, spatio-temporal gait parameters also reflect walking fundamentals (Figure 5). For example, gait impairments due to injury, ageing, psychological conditions, or medications show similar gait patterns, including slower gait velocity (due to shorter step length), increased step width, and prolonged double support time [90]. Utilising 3D motion capture systems, the fundamental spatio-temporal gait parameters can be easily measured by markers attached only to the heel and toe. Alternatively, other gait assessment tools, such as the GaitRite mat, can also record stride cycle parameters. Motion capture systems have also been used for testing commercial products, such as footwear and anti-slip strips [91,92]. Commercially available 3D motion capture systems are now relatively affordable (e.g., Optitrack, NaturalPoint), and the complex programming for 3D gait analysis can be overcome using low-cost commercial software that automatically extracts a range of gait parameters given a pre-specified standardized marker setup at data collection.

**Figure 5.** Spatio-temporal gait parameters, step length, and step width.

#### **6. Potential of a Wearable, Sensor-Integrated Insole for Real-Time Gait Monitoring**

Digital gait analysis began using multiple standard video cameras to monitor human movement from various angles to estimate 3D motion. Later, 3D motion capture systems utilising either infrared light emitting diodes (active systems) or reflective markers (passive systems) provided highly accurate position-time data and are now widely used in experimental gait research. While these systems are used as the gold standard for gait assessment, limitations in this technology include a complicated setup, high time demands, a lack of portability, and a requirement for specialized skills for system operation and data analysis. In Section 5, potential directions to overcome these problems have been proposed, and one approach is wearable sensor technology, which, despite its limitations, receives considerable ongoing research attention due to its immensely practical application in gait measurement.

Small, inexpensive body-mounted inertial measurement units (IMU) can measure angular velocity and linear acceleration [93]. IMU data can also be transmitted to Android devices using Bluetooth [94]. These innovations create the potential for online management, such as mass data storage and automatic recording of personal gait data. Difficulties in utilising IMUs for gait monitoring remain in deriving position-time data due to complex noise filtering [93]. To date, a number of techniques have been proposed to obtain greater accuracy in estimating positional data but further efforts may be required until sufficiently reliable kinematic data estimation is achieved.

Incorporation of IMUs into an insole is a promising approach to gait management because IMU attachment to the foot has the advantage of concurrently recognising gait cycle events and, using them, estimating walking speed [95]. It is, however, still difficult to precisely estimate walking speed and other associated parameters, such as stride length, using IMUs alone. It is, accordingly, fruitful to consider adding other technology, such as foot pressure sensors (Figure 6) and global positioning systems (GPS) within the insole. In-shoe foot pressure sensors can more precisely detect foot contact than IMU data, while the GPS can track the walking path and, therefore, total distance travelled. In-shoe pressure monitoring insoles are available [10,96], and real-time monitoring can alert the user to inadequate pressure distribution, such as excessive plantar pressure, to prevent foot ulcers or identify a lateral CoP excursion warning of potential balance loss.

**Figure 6.** Foot-pressure monitoring system, Pedar (Novel, Munich, Germany, www.novel.de).

The VitaliSHOE project utilises wearable IMUs and pressure sensors integrated into the shoe-insole to detect the risk of falling in the senior population [97]. In an early paper, they showed the successful detection of transition between stance and swing based on both IMU and insole pressure data [97]. They identified, however, limitations due to gait measurement at a high walking speed and fragility of the IMU associated with mechanical stress [97]. In a later project [98], temporal variables were measured with high reliability but they also acknowledged the difficulties in acquiring reliable spatial data (e.g., step length) from IMUs [98]. It could also be interesting to utilise GPS technology to obtain spatial data in studies conducted in real-world settings, and smart shoes incorporating a mobile GPS have already been piloted as a fruitful direction for gait analysis in everyday settings [99].

#### **7. Long-Term Effects**

When developing shoe-insoles, the long-term effects of wearing them should be carefully evaluated to prevent overuse injuries, restriction of natural lower limb motion, and potential anatomical deformity [100]. Long-term use of a lateral wedge insole was found to reduce adductor moment semi-permanently, a positive adaptation; but if structured inadequately, adverse effects are possible [101]. As discussed earlier, foot control modifications can cause biomechanical changes to potentially all lower limb joint actions. Even very small negative features of the gait cycle can be accumulated by the thousands of steps taken every day and can eventually cause injury. Laboratory-based gait testing is useful in inspecting whether there are hazards in gait control. If gait kinematics and kinetics are maintained, it can be speculated that no negative long-term effects may arise. It is, however, still important to conduct human trials for a prolonged period to ensure that insoles are unlikely to cause orthopaedic problems.

#### **8. Conclusions**

Shoe-insoles have potential as an effective intervention to encourage safe walking. Biomechanical gait analysis is available for monitoring insole effects on walking performance. Insole modifications could support more adaptive ankle angles, improve foot pressure distribution, absorb shock, and reduce proprioceptive reaction time. Long-term effects of shoe-insoles should be tested to support and confirm experimental biomechanical evidence for their safety and effectiveness. Integration of wearable sensors into shoe-insoles will be a very important future direction for real-time gait measurement. Taking advantage of online data management, it will be possible to achieve the goal of detailed gait analysis available to everyone in performing the gait activities of everyday life.

**Author Contributions:** Conceptualization, H.N. and R.B.; Investigation, H.N.; Writing-Original Draft Preparation, H.N.; Writing-Review & Editing, R.B.; Supervision, R.B.

**Acknowledgments:** The authors thank W.A. Sparrow for proof-reading the manuscript, Caitac for providing the insole image and Shinichi Tajima for illustrations.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### **Self-Tuning Threshold Method for Real-Time Gait Phase Detection Based on Ground Contact Forces Using FSRs**

### **Jing Tang 1,2, Jianbin Zheng 1,2, Yang Wang 1,2,\*, Lie Yu 3, Enqi Zhan 1,2 and Qiuzhi Song <sup>4</sup>**


Received: 8 December 2017; Accepted: 31 January 2018; Published: 6 February 2018

**Abstract:** This paper presents a novel methodology for detecting the gait phase of human walking on level ground. The previous threshold method (TM) sets a threshold to divide the ground contact forces (GCFs) into on-ground and off-ground states. However, the previous methods for gait phase detection demonstrate no adaptability to different people and different walking speeds. Therefore, this paper presents a self-tuning triple threshold algorithm (STTTA) that calculates adjustable thresholds to adapt to human walking. Two force sensitive resistors (FSRs) were placed on the ball and heel to measure GCFs. Three thresholds (i.e., high-threshold, middle-threshold andlow-threshold) were used to search out the maximum and minimum GCFs for the self-adjustments of thresholds. The high-threshold was the main threshold used to divide the GCFs into on-ground and off-ground statuses. Then, the gait phases were obtained through the gait phase detection algorithm (GPDA), which provides the rules that determine calculations for STTTA. Finally, the STTTA reliability is determined by comparing the results between STTTA and Mariani method referenced as the timing analysis module (TAM) and Lopez–Meyer methods. Experimental results show that the proposed method can be used to detect gait phases in real time and obtain high reliability when compared with the previous methods in the literature. In addition, the proposed method exhibits strong adaptability to different wearers walking at different walking speeds.

**Keywords:** adaptability; force sensitive resistors; self-tuning triple threshold algorithm

#### **1. Introduction**

From a medical and industrial perspective, wearable devices have evolved and continue to develop in terms of providing assistance to humans [1]. Gait analysis through wearable devices is an extensive area in the field of biomechanics that provides knowledge in terms of identifying pathologies, evaluating athletes' performance, design of sports products and rehabilitation engineering [2]. Generally, the wearable devices used for gait analysis are designed by equipping various sensors on it.

Thus far, different kinds of sensor types, including force sensitive resistors (FSRs) [3–5], air pressure sensors [6], inertial sensors [7–14], inclinometers [15], foot switches [16], and electromyography (EMG) sensors [17], are readily available in the industry and can be applied to gait analysis. Yan et al. [18] estimated the gait events through a sensory apparatus, which combine three subsystems such as a primary phase estimator, a desired gait event detector and a phase error compensator. This method

is sensitive to the tuning of the adaptive oscillators parameters. However, different parameters may change the results. In the literature, several works [11,12] have detected the gait phase using the inertial sensors that could measure the body segment orientations and joint angles. Bejarano et al. [13] proposed an adaptive algorithm based on the inertial and magnetic sensors to detect the gait events. However, the inertial sensor is sensitive to temperature, shock and magnetic disturbances, which would result in the misdetections of gait phases. In order to overcome these limitations, Muller et al. [14] proposed a novel online gait phase detection algorithm, which can be used indoor and in the presence of magnetic disturbances. However, the delays of the heel off and initial contact are considerably high. Nevertheless, an alternative solution to inertial sensors is FSRs [16]. As described by Catalfamo et al. [19], force platforms, such as FSRs, represent the gold standard method for gait analysis. Yu et al. [20] used proportion an adaptive method, which calculates the sums and proportions of ground contact forces (GCFs). However, one parameter is affected by the attachment of the shoe to the foot, which would affect the reliability of the whole system. Additionally, force platforms measure the GCFs to detect gait phases, and each gait phase has a unique GCF pattern [7].

As reported by Smith et al. [21], 80% of errors in gait phase detection using FSRs was due to the setting of threshold value. In order to set appropriate thresholds for gait phase detection, a number of researchers [4,5,19] had presented their methods. Mariani et al. [4] defined 5% body weight as a threshold. However, different subjects had different body weights such that different thresholds should be set for different subjects. In addition, no matter how slowly or fast the subject walked, only one threshold was set for the same subject in all experiments. As a result, this method of using body weight percentage as a threshold was not adaptable to different subjects and different walking speeds (i.e., slow and fast). Lopez–Meyer et al. [5] and Catalfamo et al. [19] used the maximum and minimum GCFs of gait cycles to compute the threshold. However, the use of the maximum and minimum GCFs for the threshold computation could not detect gait phases in real time because the maximum and minimum GCFs were obtained in data post-processing.

The main purpose of this study is to develop a gait phase detection method, which can detect gait phases in real time during different walking trials. Therefore, we propose a self-tuning triple-threshold algorithm (STTTA), which obtains the maximum and minimum GCFs of the present gait cycle to calculate adjustable thresholds. In addition, the acquirement of the maximum and minimum GCFs, and the calculations of adjustable thresholds are completed in the walking process. This paper also proposes a gait phase detection algorithm (GPDA), which provides the rules that determine calculations for STTTA. In order to evaluate the reliability of STTTA, three previous methods in the literature are introduced to obtain comparative results.

We hypothesize that the proposed STTTA can gain high and stable reliabilities for different subjects walking at different speeds. To better test the adaptability, the same initial threshold values are set for all subjects in all experiments.

#### **2. Methods**

#### *2.1. Subjects*

This study included fourteen males and ten females (age = 24.5 ± 2.0 years, weight = 67.3 ± 8.8 kg; mean ± SD) with no history of foot diseases. All of them volunteered to participate in our experiments. The participants walked on the treadmill for 30 s duration at designated constant speeds of 2 km/h, 3 km/h, 4 km/h, 5 km/h and 6 km/h, respectively.

#### *2.2. Instrumentation and Data Processing*

As shown in Figure 1, the sensor units comprised two FSRs (LOSON LSH-10, LOSON Instrumetation, Nankin, China), which were located in the sole of the ball and the sole of the heel. The FSRs collected GCF data at a frequency of 2000 Hz with a high resolution of 16 bits AD converter using an ARM11 computer (S3C6410). The measuring range of each FSR was 0–200 kg. The accuracy

(including linearity and repeatability) of each FSR was ±0.5% full scale (FS). Standard load cells (5 kg, 10 kg, 20 kg, 25 kg, 50 kg, 100 kg and 200 kg) were used to calibrate the FSRs. As the FSRs output a weak micro-voltage signal, an amplification circuit was equipped. Because of this circuit, the output signal of each FSR was amplified to 0–5 V, which correlated with the measured mass of 0–200 kg.

**Figure 1.** FSRs placed inside one shoe with one in the ball and the other in the heel. A lid is made to enlarge the press area.

After data acquisition, the GCF data from FSRs were filtered by second order Butterworth low pass filter with a cut-off frequency of 10 Hz. The results of the gait phase detection were processed in Matlab (version2012, MathWorks, MA, USA) using the proposed method.

#### *2.3. Previous Methods*

In previous, the GCFs measured by an FSR can be divided into on-ground and off-ground statuses through setting a threshold *T*:

$$G = \begin{cases} \text{on} - \text{ground}, \text{ F} \ge T \\ \text{off} - \text{ground}, \text{ F} < T \end{cases} \tag{1}$$

where *F* is the GCF from the ball or heel, and Tis the threshold. *G* is the result that "on-ground" status means "FSR pressed" and "off-ground" status means "FSR not pressed".

In our experiments, three previous ways were applied to the threshold computations. To be specific, one way was to define body weight percentage as a threshold, which had been affirmed to detect gait phase in real time. The other two ways were to use the maximum and minimum GCFs of gait cycles for threshold calculation. Meanwhile, the two ways carried out data post-processing to calculate an appropriate threshold, and were finally chosen as reference methods.

#### 2.3.1. Mariani Method

Mariani et al. [4] defined 5% of body weight to compute the threshold *T*:

$$T = 0.05 \cdot \text{mg}\_{\text{\textdegree}} \tag{2}$$

where m is the subject's mass and g is the acceleration of gravity.

#### 2.3.2. Timing analysis module Method

Catalfamo et al. [19] employed a Timing analysis module (TAM) software (version5.24, Tekscan, South Boston, MA, USA) as their reference method. After data acquisition of each experiment, the maximum and minimum GCFs (i.e., *T*max and *T*min, respectively) were searched out for threshold calculation:

$$T = T\_{\rm min} + (T\_{\rm max} - T\_{\rm min}) \times \frac{10}{100}.\tag{3}$$

Additionally, *T* needed to be calculated severally for each set of FSR in each experiment.

#### 2.3.3. Lopez–Meyer Method

Lopez–Meyer et al. [5] used the average value of the maximum GCFs and the average value of the minimum GCFs to compute the threshold *T*. For each experiment, the threshold computation needed all the local maximum GCFs and local minimum GCFs (i.e., *T*max(*i*) and *T*min(*j*), respectively) of gait cycles. In a complete gait cycle, there were only one *T*max and only one *T*min for each set of FSR. In one experiment, there were many gait cycles. For one set of FSR, there were *k T*max and *l T*min in one experiment. However, *k* did not equal *l* as there might exist an incomplete gait cycle in the experiment:

$$T\_{MAX} = \frac{1}{k} \sum\_{i=1}^{k} T\_{\max}(i)\_{\prime} \tag{4}$$

$$T\_{MIN} = \frac{1}{l} \sum\_{j=1}^{l} T\_{\min}(j)\_{\prime} \tag{5}$$

$$T = T\_{MIN} + \mathfrak{a} (T\_{MAX} - T\_{MIN})\_\prime \tag{6}$$

where *α* was a proportional factor for the threshold adjustment to compensate for interindividual variability in pressure levels. In our experiments, the selection of proportional factor was made that α was set to 0.084.

#### *2.4. Self-Tuning Triple-Threshold Algorithm*

With respect to the Mariani method, different thresholds should be set for different subjects, and the same threshold is set for one subject in five experiments with different walking speed each time. The Mariani method is not adaptable to different subjects and different walking speeds. Meanwhile, the TAM and Lopez–Meyer methods both carry out data post-processing to calculate the thresholds and are incapable of real-time application. In order to seek a method to detect gait phase in real time and be adaptable to different walking conditions, we propose a self-tuning triple-threshold algorithm (STTTA) that uses three thresholds to obtain the maximum and minimum GCFs for threshold computations.

#### 2.4.1. Setting of Three Thresholds

In this section, the magnitudes of three thresholds are amplified. Figure 2 demonstrates the GCFs processing through three thresholds, including high-threshold, middle-threshold and low-threshold (i.e., *TH*, *TM* and *TL*, respectively). Specifically, *TH* is utilized to search the maximum GCF (i.e., *T*max(*j*)) of the present gait cycle, and chosen as the main threshold that divides the GCFs into on-ground and off-ground statuses. *TM* is used to search time points (points b, e and h shown in Figure 2) for the adjustments of *TH* and *TL*. Meanwhile, *TL* is employed to search the minimum GCF (i.e., *T*min(*j*)) of the present gait cycle, and the time point (point f) for the adjustments of *TM*.

At the present gait cycle, the three thresholds are newly calculated for the next gait cycle. The three thresholds (i.e., *TH*(*i* − 1), *TM*(*i* − 1) and *TL*(*i* − 1)), by which the GCFs are processed at the present gait cycle, have been calculated at the last gait cycle.

**Figure 2.** Three thresholds used for the GCFs processing.

As Figure 2 shows, the present gait cycle starts at point a and ends at point g. Point a is the end point of the last gait cycle, and also the start point of the present gait cycle. Point b is used to calculate the *TL*(*i* − 1). Points c and d make a region [c, d]. When the GCF (*F*(*k*)) is larger than *TH*(*i* − 1), it goes to region [c, d]. Then, the *T*max(*j*) can be searched out as follows:

$$T\_{\text{max}}(j) = MAX = \begin{cases} F(k), \ F(k) \ge T\_H(i-1) \& F(k-1) < T\_H(i-1) \\\ F(k), \ F(k) > MAX \& F(k) > T\_H(i-1) \& F(k-1) \ge T\_H(i-1) \\\ MAX, \ F(k) < MAX \& F(k) > T\_H(i-1) \& F(k-1) > T\_H(i-1) \end{cases},\tag{7}$$

where *k* is the total number of the collected GCFs from one FSR in each experiment. The numerical range of *k* is 1~60,000 because each experiment lasts 30 s with a sampling frequency of 2000 Hz.

At point e, *TH*(*i*) is calculated as a new threshold for the next gait cycle:

$$T\_H(i) = \beta \cdot (T\_{\text{max}}(j) - T\_L(i-1)) + T\_L(i-1),\tag{8}$$

where *β* is a proportion factor and chosen to be a constant. *i* means that the present gait cycle is the *i*-th step of one subject walking in one experiment, while *j* means that the *T*max(*j*) is the *j*-th maximum GCF of the total gait cycles. However, *i* was not necessarily equal to *j* as incomplete gait cycles may have occurred.

At point f, *TM*(*i*) is newly calculated for the next gait cycle:

$$T\_M(i) = \gamma \cdot (T\_{\text{max}}(j) - T\_L(i-1)) + T\_L(i-1),\tag{9}$$

where *γ* is a proportion factor and chosen to be a constant.

In common with points c and d, the points f and g also make a region [f, g]. The calculation of *T*min(*j*) is made in the regions [f, g], which means that the *F*(*k*) is smaller than *TL*(*i* − 1):

$$T\_{\rm min}(j) = MIN = \begin{cases} F(k), \ F(k) \le T\_{\rm L}(i-1) \& F(k-1) > T\_{\rm L}(i-1) \\\ F(k), \ F(k) < MIN \& F(k) < T\_{\rm L}(i-1) \& F(k-1) \le T\_{\rm L}(i-1) \\\ MIN, \ F(k) > MIN \& F(k) < T\_{\rm L}(i-1) \& F(k-1) < T\_{\rm L}(i-1) \end{cases} \tag{10}$$

Additionally, point g is the end point of the present gait cycle, and also the start point of the next gait cycle.

However, *TL*(*i*) is newly set at point h, as the same to *TL*(*i* − 1) calculated at point b:

$$T\_L(j) = \lambda \cdot T\_{\min}(j) + (1 - \lambda) \cdot T\_M(i),\tag{11}$$

where *λ* is a proportion factor and chosen to be a constant.

#### 2.4.2. Setting of Initial Threshold Values

Before the experiments, it is necessary to set initial values for the three thresholds. In this paper, *TH*(1), *TM*(1) and *TL*(1) are the initial values of *TH*, *TM* and *TL*. However, the setting should follow that *TL*(1) < *TM*(1) < *TH*(1).

#### *2.5. Gait Phase Detection Algorithm*

The gait cycle can be divided into stance-phase and swing-phase. The transitions between gait phases are gait events. As proposed by Pappas et al. [3,22], swing-phase, stance-phase, heel-strike and heel-off are the most common gait phases and gait events for a single foot. For the proposed STTTA, *TH* is the main threshold, which divided the GCFs into on-ground and off-ground statuses:

$$\mathcal{G} = \begin{cases} \text{on}-\text{ground}, \ F \ge T\_H \\ \text{off}-\text{ground}, \ F < T\_H \end{cases} \text{.} \tag{12}$$

As is the same for Equation (1), *F* is the GCF from the ball or heel, and *TH* is the threshold. *G* is the result that "on-ground" status means "FSR pressed" and "off-ground" status means "FSR not pressed".

When the GCFs from the ball and heel are processed from Equations (1) and (12), the gait phases and gait events can be distinguished by following the rules in Table 1.

**Table 1.** Rules of gait phase detection algorithm.


#### *2.6. Reference Methods*

In order to obtain the reliability of the proposed STTTA, reference methods should be determined. The TAM and Lopez–Meyer methods carried out data post-processing to calculate an appropriate threshold for each FSR at each walking speed. In addition, the TAM method had been chosen as a reference method to test the reliability of gait events detection. On the other hand, the Lopez–Meyer method had been tested on a shoe-based wearable sensor system, which had been compared with the "GAITRite system". The Lopez–Meyer method acquired 95% confidence to compare detection results with the GAITRite system. Therefore, in this paper, the TAM and

Lopez–Meyer methods were both selected as reference methods. Finally, both the STTTA and Mariani methods were compared with the reference methods to obtain their reliabilities.

#### **3. Experimental Results**

#### *3.1. Selection of Coefficients*

Before the experiments, the values of *TH*(1), *TM*(1) and *TL*(1) were determined. The value selections of *TH*(1), *TM*(1) and *TL*(1) followed the rules that *T*min < *TL*(1) < *TM*(1) < *TH*(1) < *T*max. Based on data analysis, the value range of *T*min was 0–5 N and the value range of *T*max was 300–1300 N. Therefore, the values of *TH*(1), *TM*(1) and *TL*(1) could be selected within a wide range. To evaluate this in our

experiments, we randomly chose a set of initial threshold values where *TL*(1) = 15 N, *TM*(1) = 20 N and *TH*(1) = 25 N.

After the selection of *TH*(1), *TM*(1) and *TL*(1), the proportion factor β, γ, and λ in Equations (8), (9) and (11) were optimized to determine the highest reliability of STTTA. From Equations (9) and (11), the values of *TM* and *TL* were affected by each other. As a result, one of γ and λ should be firstly determined such that the value of λ was chosen to be 0.5. Then, using the data from nine subjects as training data, the reliabilities of the proposed STTTA were calculated by using various values of β and γ, which is shown in Figure 3. The selections for β and γ to acquire the highest reliability was made such that β = 0.071 and γ = 0.042.

Finally, the same *TH*(1), *TM*(1) and *TL*(1) were used for all 24 subjects in all experiments. In addition, data from the remaining subjects were processed using the selected β, γ, and λ.

**Figure 3.** The reliabilities of the proposed STTTA in various values of *β* and *γ*.

#### *3.2. The Results of Gait Phase Detection and Experimental Comparisons*

Figure 4 showed the detection results of the proposed STTTA. In Figure 4a, the GCFs from the ball and heel were displayed. The self-adjustments of three thresholds for the GCFs from the ball and heel is demonstrated in Figure 4b,c. It was qualitatively illustrated in Figure 4d where the gait phases and gait events were detected through the proposed GPDA. In order to clarify the self-adjustment effect, the main threshold TH of the proposed STTTA was compared with the thresholds calculated by the previous methods. As pictured in Figure 4e,f, the amplitudes of the threshold TH were adjustable and adaptable to walking conditions, while constant thresholds were figured out for the previous methods.

**Figure 4.** *Cont*.

**Figure 4.** *Cont*.

**Figure 4.** (**a**) the GCFs of heel and ball measured by FSRs; (**b**) three self-tuning thresholds for the processing of GCF from the ball; (**c**) three self-tuning threshold for the processing of GCF from the heel; (**d**) the result of gait phase detection through GPDA; (**e**) thresholds calculated for the ball by four kinds of methods; and (**f**) thresholds calculated for the heel by four kinds of methods.

#### *3.3. Real-Time Application for Gait Phase Detection*

After data acquisitions, the data processing included the initial threshold setting, acquirement of the maximum and minimum GCFs, and threshold computations, was made prior to the GPDA. When the threshold computations were achieved, the gait phase detection could be implemented through the GPDA.

In this gait phase detection system, the sampling frequency was set to 2000 Hz. Data acquisitions, data processing and gait phase detection could be all accomplished within one sampling period. The time delay between data acquisition and gait phase detection was less than 0.5 ms. Therefore, the proposed STTTA could be used to detect gait phases in real time.

#### *3.4. Adaptability to Different Walking Conditions*

Needless to obtain the body weights of all subjects, the proposed STTTA acquired higher average reliabilities than the Mariani method as shown in Table 2. In addition, irrespective of the reference method applied (TAM method or the Lopez–Meyer method), higher reliability was gained for the STTTA compared to the Mariani method. The comparative results affirmed that the proposed STTTA was more reliable than the Mariani method in real-time gait phases detection.


**Table 2.** Comparative reliability results for the proposed method.

To demonstrate the adaptability of the proposed STTTA, the experimental results of one male subject and one female subject were taken as examples. As shown in Figure 5a–d, the Mariani method could achieve high reliability at one walking speed, and then reduce significant reliability at other walking speeds. However, the proposed STTTA gained stable reliabilities at five walking speeds. The stability to speeds proved that the proposed STTTA was adaptable to different walking speeds.

**Figure 5.** *Cont*.

**Figure 5.** Reliabilities of the Mariani and STTTA method for one male subject at five walking speeds: (**a**) TAM method as reference method; (**b**) Lopez–Meyer method as reference method. Reliabilities of the Mariani and STTTA method for one female subject at five walking speeds; (**c**) TAM method as reference method; (**d**) Lopez–Meyer method as reference method.

#### **4. Discussion**

#### *4.1. The Reason of Using Three Thresholds for GCFs Processing*

At the beginning, one threshold was expected to obtain the maximum and minimum GCFs of the present gait cycle, and the time point of a new threshold calculation for the next gait cycle. As shown in Figure 6a, the region (c,d) was used to search the maximum GCF of the present gait cycle, and the point d was used as the time point for new threshold calculation. However, the new threshold would be calculated twice (i.e., it should be only once) occasionally at one gait cycle as shown in Figure 6b. The threshold *T*(*i*), which was the correct threshold calculated for the next gait cycle at time point d. Nonetheless, it possibly existed that the GCFs after time point d were larger than the newly computed threshold *T*(*i*) in a new region (c',d'). In this situation, the algorithm would search the maximum GCFs in region (c',d'), and figured out an unexpected new threshold *T*'(*i*), which replaced the *T*(*i*) for threshold processing of the next gait cycle, at time point d'. Additionally, the secondly calculated threshold *T*'(*i*) was a false threshold for the next gait cycle. Similarly, using two thresholds also resulted in the same problem.

However, using three thresholds avoided the above-mentioned problem. As Figure 2 showed, the maximum and minimum GCFs were severally obtained in region (c, d) and (f, g); meanwhile, the new thresholds (i.e., *TH* and *TL*) were severally computed at time point e and h. The middle threshold TM was to search the time point e and h, which avoided the new threshold computation at time point d or g. To sum up, the use of three thresholds avoided the problem of twice threshold computations.

#### *4.2. Adaptability to Variable Speed Walking*

Generally, when the subjects walk faster, their feet hit the ground harder, i.e., the magnitude of GCFs gets bigger. As a result, the magnitude of GCFs changes with the variation of walking speeds; meanwhile, the maximum GCF increases (or decreases) as the walking speed increases (or decreases). According to the formula in Equation (8), when the subjects walk slowly, the maximum GCF possesses small magnitude such that a small threshold is calculated. Similarly, when the subjects walk fast, the formula calculated a big threshold. As to variable speed walking, one computed threshold should be set for each walking speed to conduct correct gait phase detection. If the subject suddenly changes walking speed, the ground contact force changes accordingly. The subject will take one step, and the thresholds will be adjusted to the new speed because three thresholds of the next gait cycle are calculated by the GCFs of the present gait cycle. The proposed STTTA uses the maximum and minimum GCFs from the last gait cycle to calculate a threshold (i.e., *TH*) for the present gait cycle. The maximum and minimum GCFs of the last gait cycle are the closest to those of the present cycle. As a result, appropriate thresholds are calculated for each gait cycle, which correlates with speed variation. To sum up, the proposed STTTA is adaptable to variable speed walking.

#### *4.3. Limitation of the Research*

In our study, only healthy subjects have been studied on level ground, not taking the pathological subjects into account. The experiments were done on a treadmill because the method is not suitable for irregular terrain and stairs walking. Uneven road conditions and obstacles on the ground may lead to misdetections.

In the experiments, some rules of GPDA were not observed. However, the specific walking habits of individuals may employ the unused rules. The GPDA rules could be used to detect abnormal gaits for podiatric diagnoses.

**Figure 6.** (**a**) the description of using one threshold for threshold adjustment; (**b**) demonstration of threshold adjustment using one threshold and analysis of twice threshold computations.

#### **5. Conclusions**

This paper proposes a self-tuning triple-threshold algorithm that calculates adjustable thresholds to adapt to human walking. In addition, the adjustable thresholds are calculated in the walking process, and the same initial threshold values are set for all subjects in all experiments. For all subjects, high average reliabilities are gained when the proposed method is severally compared with two reference methods. For each subject, the proposed method acquires stable reliabilities in five experiments with different walking speeds each time. It comes to a conclusion that the proposed STTTA can be used to detect gait phases in real time, and shows adaptability to different walking conditions on level ground. In the future work, the method of gait phase detection would be studied on irregular terrains or when climbing stairs. Furthermore, it needs more types of sensors to detect the sub-phases of swing.

**Acknowledgments:** We would like to thank all participants for helping us with data acquisitions. This work was supported by the National Key R&D Program of China "The study on Load-bearing and Moving Support Exoskeleton Robot Key Technology and Typical Application" (2017YFB1300500).

**Author Contributions:** Jing Tang, Jianbin Zheng and Yang Wang proposed the conception and design of this research, acquired the data and drafted this article. Lie Yu and Enqi Zhan helped with the analysis and interpretation of data. Qiuzhi Song revised for important intellectual content. All authors revised and approved the final paper version to be published.

**Conflicts of Interest:** None of these authors has conflict of interest in this research.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **A Real-Time Wireless Sweat Rate Measurement System for Physical Activity Monitoring**

#### **Andrew Brueck 1, Tashfin Iftekhar 1, Alicja B. Stannard 2, Kumar Yelamarthi <sup>1</sup> and Tolga Kaya 3,\***


Received: 20 October 2017; Accepted: 8 February 2018; Published: 10 February 2018

**Abstract:** There has been significant research on the physiology of sweat in the past decade, with one of the main interests being the development of a real-time hydration monitor that utilizes sweat. The contents of sweat have been known for decades; sweat provides significant information on the physiological condition of the human body. However, it is important to know the sweat rate as well, as sweat rate alters the concentration of the sweat constituents, and ultimately affects the accuracy of hydration detection. Towards this goal, a calorimetric based flow-rate detection system was built and tested to determine sweat rate in real time. The proposed sweat rate monitoring system has been validated through both controlled lab experiments (syringe pump) and human trials. An Internet of Things (IoT) platform was embedded, with the sensor using a Simblee board and Raspberry Pi. The overall prototype is capable of sending sweat rate information in real time to either a smartphone or directly to the cloud. Based on a proven theoretical concept, our overall system implementation features a pioneer device that can truly measure the rate of sweat in real time, which was tested and validated on human subjects. Our realization of the real-time sweat rate watch is capable of detecting sweat rates as low as 0.15 μL/min/cm2, with an average error in accuracy of 18% compared to manual sweat rate readings.

**Keywords:** sweat sensor; sweat rate; dehydration; IoT; PDMS

#### **1. Introduction**

With the popularity of smartwatches, a new lifestyle is growing, particularly in urban areas, where people are more conscious of how they treat their own bodies [1]. Numerous devices are available for professionals and exercise/technology enthusiasts that monitor body vitals such as heart rate monitors [2], temperature sensors [3], and even optical glucose measurement devices [4]. One of the growing exercise activities to stay fit and healthy is endurance sports such as marathons, biking, and triathlons [5]. Not only is there a significant increase in the number of professional athletes competing in these events, but also there has been growing interest among non-professional athletes every year in the last decade [6]. There are several smartwatches on the market that can help monitor exercise intensity and performance. Even though heart rate monitoring is considered the most effective way of monitoring the effort level, athletes are well aware that fluid intake before, during, and after a hard training is quite important [7,8]. Fluid intake of less than what is required by the body would lead to dehydration (or hypohydration), and cause both physical and mental performance degradation [8–10]. On the other hand, too much fluid intake, more than required (hyperhydration), can be fatal [7].

It is imperative that most of the fluid losses during a race or training occur through sweating in order to regulate the core body temperature [11]. Physical exercise leads to an increase in core body temperature, and sweating through skin cools down the body via thermoregulation. Sweat glands that are located under the skin are responsible for sweating action. Eccrine sweat glands are the main sweat glands located in human skin, and are responsible for regulating core body temperature [12]. On the other hand, apocrine sweat glands are located at armpits and groin areas that are activated during emotional and sexual arousement [13]. Although the density of eccrine sweat glands varies significantly throughout the body, some averages can be given, such as 104 glands/cm2 for forearms, 155 glands/cm2 for the forehead, etc. [14].

Research on the physiology of sweat is more than a century old, when contents of sweat were first analyzed [13]. What is in sweat is well understood; it is mostly water, with electrolytes such as sodium, potassium, urea, lactate, and other trace minerals [12,14]. Several sweat-based hydration detection monitoring systems were proposed that included optical [15,16], electrochemical [17,18], or amperometric [17,19] techniques. It is known that the amounts of individual constituents of sweat not only vary from person to person, but also show significant variations in different regions of the same individual [14,20,21]. Without establishing a base level of sweat on a particular individual's particular body region, comparisons cannot be made. Furthermore, the concentration of sweat also changes with the rate of sweat, which is directly related to the intensity of the exercise and environmental conditions [11,22,23]. Therefore, sweat rate plays a significant role in physical conditioning, and is a useful variable for determining the performance of an athlete, particularly in endurance sports. A real-time measurement of sweat rate can ultimately provide more deterministic results on how to interpret sweat concentration data that could be used to gather physiological information about an athlete.

Sweat-rate sensing was first performed by utilizing traditional sweat collection methods such as whole-body wash [24,25] or absorbent pads [26]. Although accurate, these approaches could only provide a limited number of data points, and cannot be considered in real time. Toward sweat rate sensing, flow rate sensors were considered as an alternative. It was widely accepted that microflow sensing (flow rates around μL/min) is implemented with thermal flow rate sensors, particularly calorimetric or thermal pulse flow sensors [27,28]. Matzeu et al. demonstrated the detection of sweat rate using image analysis where they took images of the sweat filling a commercial sweat collector (Megaduct, a larger volume than a Macroduct) [29]. This approach was arguably the first real-time sweat rate detection system. However, it requires cameras and post-processing, altering the use of the system as a wearable device. Coyle et al. also proposed a preliminary sweat rate sensing textile patch employed by discrete humidity sensors [30]. Although promising, this device was only capable of measuring the evaporated sweat rather than the total secreted sweat from a region, which is called transepidermal water loss, and is not the same as the sweat rate. Wei et al. used a cellulose-based sweat absorber to measure sweat levels [31]. Real-time sweat rates were obtained, but these devices were not validated through manual measurements. Furthermore, their system relied on the weight of the absorbed sweat, which would cause hidromeiosis, i.e., the local dysfunction of sweat glands due to the blockage of sweat ducts.

Building on our years-long sweat concentration research [16,32–34], we have recently proposed a concept that could be used for sweat rate sensing applications, and stated that it could potentially be utilized for sweat-rate sensing applications [35]. Building upon our expertise on the Internet of Things (IoT) [36–39], we are bringing our concept into life by proposing a wearable, real-time sweat rate device. Although we have shown the feasibility of the concept in our previous work [30], this paper provides an innovative implementation of the concept, and expands the scope with human subject validation (compared with manual data collection). A unique modification in building a microfluidic calorimetric sweat rate sensor allowed us to feed the sensor with sweat via a commercial sweat collector (Macroduct). Sweat rate data was obtained, processed, analyzed, and sent to the cloud for instantaneous and longitudinal monitoring and analysis. A smartphone app was developed that displayed the sweat rate information obtained wirelessly from a LilyPad Simblee board [40]. A custom

printed circuit board (PCB) was designed with surface mount components. Human trials were also conducted to validate the proposed IoT-based sweat rate monitoring system.

#### **2. Materials and Methods**

The flow rate sensor was built using a polydimethylsiloxane (PDMS) silicone elastomer kit (Dow Corning) by using two pieces of PDMS. A silicon wafer with a long rectangle channel mold was used. PDMS was poured into the silicon mold to create the inverse channel, which was 26 mm × 3 mm × 1 mm. Inlet and outlet holes were punched through the PDMS channel. The upper PDMS block had a 10-ohm resistor embedded in it that acted as a heater, and holes were also punched through to allow two copper–constantan T-type thermocouples (Omega, with 30 gauge wiring) to insert. Upper and lower PDMS pieces were attached together by using the plasma cleaner (Harrick PDC-32G) to form the closed channel. Thermocouples were then slid into place through the already punched holes. Loctite was used to ensure a tight seal around the thermocouples and the heating resistor.

A PCB was designed and fabricated to minimize the footprint of the electronics. A power transistor (NPN MJD200T4G) was used to supply enough current to the heating resistor. An instrumentation amplifier (AD8237ARMZ) was utilized to convert the differential temperature readings into voltage. An entire PCB was powered via the Simblee BLE's (Bluetooth Low Energy) regulated 3.3-V output. The voltage output of the amplifier was connected to the Simblee board for wireless communication.

The characterization of two copper–constantan T-type thermocouples was done by using a Revlon turbo/lightweight/1875 W hairdryer. Thermocouples were physically kept farther apart, and one of them was heated. The process was reversed for thermocouples in order to obtain a symmetric temperature difference. The temperature reading of the thermocouples was obtained from two Omega HHM9007R multimeters (OMEGA, Norwalk, CT, USA). The voltage output of the differential temperature readings of the thermocouples was obtained from the instrumentation amplifier. A non-zero reference voltage was used for the amplifier. The temperature value and corresponding output voltage were then recorded.

The characterization of the sweat rate sensor device was performed by pumping deionized (DI) water through the channel with a syringe pump (KDS Scientific 100) at flow rates from 3 μL/min to 100 μL/min, which were equally spaced logarithmically. The temperatures and the output voltage of the amplifier were recorded once the flow rate reached a steady state value (about 10 min).

The Macroduct sweat collector (ELITechGroup Biomedical Systems, South Logan, UT, USA) was used to guide the sweat to the PDMS channel. A 3.7-V 500 mAh rechargeable LiPo PKCell battery was used to power the Simblee board, which then powered the PCB. The entire device was housed in a box that was custom designed using TinkerCAD (Version 3.9, Autodesk, CA, USA), and three-dimensional (3D) printed by Makerbot Replicator + (New York City, NY, USA).

Human trials were conducted using an exercise bike. In order to record the sweat rate sensor manually, a Macroduct was placed on the left upper forearm of the sweat rate sensor prototype, and the volume of sweat in the Macroduct tubing was recorded over time. The manual rate of sweat was then calculated by taking into account the dimensions of the Macroduct tubing. The sweat rate sensing device was put on the right upper forearm. Output voltages of the device were wirelessly transmitted to a smartphone (HTC One M7 WLS and iPhone 6s) via the LilyPad Simblee BLE board, and a mobile app was utilized to read the data every 15 s or 30 s. A cloud interface was built around ThingSpeak. A Raspberry Pi 3 was used as the receiver and connection to the cloud. An online stopwatch (Free-stopwatch.com) was used to keep time accurately for the sweat rate testing. Five subjects were tested after they gave informed consent. Institutional Review Board (IRB) approval was obtained at Sacred Heart University (IRB #170922A). Subjects were asked to bike on the exercise bike at 80% workload of their maximum heart rate, which was calculated as 208 − (0.7 × age). Room temperature was 24 ◦C, and the relative humidity was recorded as 40%. Subjects were allowed to drink water before, during, and after ad libitum.

#### **3. Device Architecture**

The device was built on the commercial Macroduct sweat collector with some modifications. The coiled tubing of the Macroduct was shortened to 1 cm, and the tubing was connected to the inlet of the calorimetric sweat rate sensor to guide the sweat into the sweat rate sensing channel. Temperature information readings, as well as heating element control circuitry, were designed and custom built on a PCB. Information gathered from the sensor was processed at the PCB electronics, and the voltage output of the sweat rate information was sent to the Simblee board for wireless transmission. The circuitry was powered by the Simblee board, which had a LiPo battery as its own power supply. The individual components of the entire system and the assembled device photos are shown in Figure 1a,b, respectively.

**Figure 1.** Components of the sweat rate sensing device: (**a**) Individual parts that were used to build the prototype; (**b**) overall device view after the assembly and connections were made.

Two system architectures have been developed to test the sweat rate sensor's IoT functionality, as illustrated in Figure 2. Both architectures are compatible with the current system, and no hardware modifications needed to be made on the device. The first architecture purely focuses on monitoring the instantaneous sweat rate of just one user, and provides a warning to the user so that appropriate actions can be taken when the level goes beyond a threshold value. This monitoring is accomplished through a cell phone application with which the LilyPad Simblee can directly communicate. The second architecture is designed such that the sweat rate from multiple users could be obtained simultaneously, longitudinal data analytics could be performed, and data could be stored for further analysis by a clinical expert. This architecture requires an internet gateway, which was implemented by a Raspberry Pi 3 on the receiver side. A ThingSpeak IoT platform was utilized for data analytics, storage, and visualization.

**Figure 2.** Architectural components of an Internet of Things (IoT) framework for instantaneous (architecture 1) and longitudinal (architecture 2) sweat rate analysis.

#### **4. Results**

Figure 3a shows the initial sweat rate sensing concept where the initial theoretical operation was introduced by the authors [30]. A heating element is placed in a flow channel. Two thermocouples (upstream and downstream) are placed on each side of the heater (2-mm spacing) to sense the temperature. Once there is a flow, temperature around the upstream sensor (*Tu*) cools off because of the fluid flow. Temperature around the downstream sensor (*Td*), on the other hand, increases with the flow as the temperature profile is pushed further with the flow. Once a critical flow rate value is reached (detailed theoretical analysis can be found in [31]), *Td* starts to cool off as well (due to high flow rates). The temperature difference, Δ*T* = *Td* − *Tu*, reaches a maximum at this critical flow rate, as illustrated in Figure 1b. In practice, this threshold occurs at high flow rates that are well beyond the human sweat rate. Therefore, a more realistic region of Δ*T* was represented with a solid line, and the theoretical (but not practical) extrapolation of Δ*T* is shown as the dashed line in Figure 2b.

**Figure 3.** Theoretical principles of the calorimetric flow rate device. (**a**) Two thermocouples (upstream and downstream) were placed along the heating element. Sweat entered the channel from the bottom, and exited from the top as an outlet; (**b**) The temperature difference between thermocouple readings increased with the flow and reached a maximum. Realistic flow rates were well below this maximum. The temperature difference is a monotonously increasing function of the flow rate.

The overall circuit diagram of the sensor and its electronics was provided in Figure 4. The base voltage of a transistor (NPN MJD200T4G) is biased using a simple voltage divider. The current that the transistor draws flows through a 10-ohm resistor connected between the emitter and the ground. This resistor acts as a heater, and is embedded in the PDMS device. The temperature upstream and downstream of the heater are both sensed by copper–constantan T-type thermocouples (Omega). The thermocouples produce a voltage depending on the temperature with a slope of 40 μV/◦C. Both of the thermocouples are inputs to a linear instrumentation amplifier (AD8237ARMZ), which takes the difference between the two and amplifies the result to obtain voltage values that can be detected by the microcontroller. External resistors set the gain of the amplifier, and the reference voltage is set by a 10-kΩ potentiometer. The output voltage of the amplifier is then sent to the Simblee board for data conversion and wireless transfer. The Simblee board has a regulated 3.3-V power supply that powers the transistor network, as well as the amplifier. An external LiPo 3.7 V 500 mAh PKCell battery powers the Simblee board.

**Figure 4.** Schematic diagram of the IoT platform. A power transistor was used to set the current to the heating resistor. The temperature difference was converted into voltage via the instrumentation amplifier, and fed to the Simblee board for wireless communication.

The circuitry was first tested by changing the temperature around the thermocouples using a hairdryer to characterize the temperature to voltage conversion. It can be seen from the Figure 5 inset that a linear voltage increase occurs with temperature changes. In this experiment, Δ*T* was changed on a range of approximately 4 ◦C to show the trend. By setting the instrumentation amplifier gain to 1000, and knowing that the thermocouple voltage conversion is 40 μV/◦C, an ideal function can be derived as *Vout* = 40 μV/◦C × 1000 × Δ*T* + *Vref*, where *Vref* represents the reference voltage for the amplifier. The reference voltage in Figure 5 inset was subtracted for visualization purposes.

**Figure 5.** Characterization of the IoT device. The inset shows the linear relationship between the output voltage and the temperature difference. The main figure shows the output voltages that were obtained from different flow rates (generated with a syringe pump). A quadratic polynomial was fit to the data so that the output voltages can be converted into flow rate in real human trials.

Once the circuitry operation has been validated, the sweat rate device was characterized by using a syringe pump for various volumetric flow rates (in μL/min). Figure 5 shows the output voltage values with respect to varying flow rates, which were aligned with the theoretical illustration in Figure 2b. It can be seen that *Vout* starts to increase back from the minimum point at around 80 μL/min, which would make detection difficult (two flow rate data points would correspond to the same voltage value). On the other hand, flow rates below 1 μL/min do not result in a significant change in output voltage, reducing the resolution of the measurement. Therefore, the volumetric flow rate detection limit is determined to be between 1–80 μL/min. It is important here to note that the sweat rate is given in flow rates per unit area (cm2). Although these experiments were conducted by pushing the liquid to the PDMS channel directly from a syringe pump, corresponding sweat rate values could be calculated by using the area of the Macroduct, which would feed the sweat to the sensing area. As the diameter of the Macroduct is 2.7 cm, the minimum sweat rate that can be detected with the current system would be 0.15 μL/min/cm2. Although the upper limit of detection for the device is calculated to be around 13 μL/min/cm2, this limit would never be reached, because it is known that the sweat rate for humans does not exceed 5 μL/min/cm2. Hence, the range of the device fits well to the application. A second order polynomial was fit to the experimental data, which can be used to obtain the flow rate (FR) from the output voltage (*Vout*).

In order to validate the sweat rate device, human subject trials were conducted. A commercial Macroduct was used as the control. The Macroduct collects the sweat in the spiral tubing as the collector makes contact with the skin, and the small pinhole at the bottom allows the sweat to go through the tube. Spiral Macroduct tubing was marked every 45◦ turn, which corresponds to a 5-mm distance between each mark. The exact locations of the marks were determined by uncoiling the spiral tube, as shown in Figure 6a. The volume of the sweat between each mark was calculated using the inner diameter of the tubing (0.64 mm). The time it took the sweat to go from one mark to another was noted, and the volumetric flow rate was calculated accordingly. This method allowed us to record sweat rate values approximately every 2 μL in the tube. The Macroduct tube has a dye in it that made it easier to visualize the location of the sweat. An actual Macroduct usage on a subject is shown in Figure 6b.

**Figure 6.** Manual determination of the sweat rate was achieved with a Macroduct sweat collector. (**a**) Uncoiled Macroduct tube to measure the distance between each mark; (**b**) Photo of the Macroduct being filled with the sweat during an actual subject trial.

Two male and three female participants completed the exercise protocol (Table 1). Exercise duration ranged from 18 min to 30 min, while the average sweat rate was 0.76 ± 0.41 <sup>μ</sup>L/min/cm2. Pre and post-weights were measured with clothes. Therefore, weight loss during the exercise only provides a qualitative measure, and can be assumed that the overall weight loss was greater. No fluids

were ingested during the trial. In order to provide more quality information regarding the validation of the device, error values in accuracy were devised from the data by calculating the percent deviation of the sweat rate (*SR*) from the device data (*SR*device) to the manual Macroduct collection (*SR*control), and are presented in the Table 1 as the percentage error (calculated as 100*x*(*SRdevice* − *SRcontrol*)/*SRcontrol*). The percent error in accuracy for all of the subjects ranged from 4–30% (Table 1).

**Table 1.** Results from the validation of the sweat rate sensor device.


The sweat rate device was worn on the opposite arm to the manual data collection, and data was collected simultaneously. Sweat rate readings were gathered every 30 s from the phone app. Figure 7 shows four of the five subject tests. It is apparent that the manual readings from the Macroduct collection started 7–10 min after the exercise started, as it takes time for the Macroduct reservoir to begin filling. Similarly, device data for subjects 1 and 2 were also delayed. Subject 4 and 5 data from the device started from the very beginning, due to the residual DI water left from rinsing the device. This residual water was left in the tube intentionally to lower the latency of the device. As expected, subjects' sweat rate patterns varied during the exercise protocol, with most of the patterns being relatively constant after 12 min of exercise.

**Figure 7.** Human trials were conducted to validate the sweat rate sensing device. Manual data collection from the Macroduct was recorded by visually observing the sweat going through the spiral tube. Sweat rate from the device was recorded from the Smartphone.

The IoT framework was tested with the actual sweat rate sensor platform. Figure 8a shows the real-time data of the sweat rate sensor, where different flow rates were introduced to the device. The flow rate was calculated from the obtained voltage values using the second-degree polynomial obtained earlier. ThingSpeak offers a great way of post-processing the data using embedded MATLAB functions (as seen in green buttons). The mobile app, as shown in Figure 8b,c, was designed to give instantaneous feedback to the user if the sweat rate values reached excessively high levels. Currently,

this threshold was set to 1.5 μL/min/cm2, which can easily be adjusted depending on the situation. If the threshold is exceeded, the app gives a warning (shown in Figure 8c) on the screen.

**Figure 8.** The IoT framework with a sweat rate sensor prototype. (**a**) The ThingSpeak web interface was used for real-time monitoring of the data; (**b**,**c**) The smartphone app was developed based on the Simblee's interface to give instantaneous feedback to the user.

#### **5. Discussion**

The current device has a size of 6 cm × 6 cm × 5 cm, which equates to a form factor of 180 cm3. Considering that the form factor of a sports watch is around 25 cm3, the size of the current sweat rate sensing device is big. From the end product point of view, the PDMS device can be made much thinner, and a smaller battery could be used. Lastly, a PCB can be embedded into the PDMS. These efforts could potentially reduce the form factor to below 100 cm3, which would be acceptable for commercialization.

The overall current consumption of the device is around 80 mA. Most of the current is consumed at the heating element. The Simblee board is capable of providing enough current to the device. The current consumption of the instrumentation amplifier is 0.1 mA, which can be considered negligible. During the data transmission, the Simblee board can consume as high as 10 mA. The relatively large current consumption of the entire device results in the need for a high-capacity battery. If the current is lowered on the heating element, the temperature differences, and hence the output voltage swings, get significantly smaller, limiting the sensitivity of the conversion. One way to reduce the current consumption is to implement the heater using thin films, i.e., microfabrication techniques. This approach would significantly lower the form factor and allow the device to become more sensitive at below 0.15 μL/min/cm2. The authors had theoretically proved that the effective range of the device can be shifted to lower flow rates by downscaling the device dimensions, particularly the distance between the heater and the respective thermocouples [35].

In order for the Macroduct sweat collector to start filling up the coiled tube, the interface between the skin and the Macroduct needs to fill first. However, because the Macroduct conforms to the skin, we assumed this volume is less than 1 μL. There is about 1 cm of coiled tube before the sweat enters the sweat-sensing channel, resulting about 4 μL of sweat that is not analyzed. Furthermore, both of the thermocouples that are in the sweat channel need to be immersed in sweat. Considering this distance is around 1 cm (from the location where the sweat enters the channel to the downstream thermocouple), and the cross-section of the channel is 3 mm × 1 mm, potentially 30 μL of sweat would be required. Overall, almost 35 μL of sweat would be needed to start receiving sweat rate data. An average sweat volume rate of 3 μL/min (approximately 0.5 μL/min/cm2) would result in a 10-min latency, which is consistent with our observations for subjects 1 and 2 (Figure 7). However, the sweat channel and Macroduct tubing can be filled with DI water from PDMS outlet in order to reduce the latency. It can be seen from subjects 4 and 5 (Figure 7) that the latency was completely removed.

It is apparent that the local sweat rate cannot be a good representative of the whole body sweat rate. However, a rough estimate can be performed. Taylor et al. provided a detailed study on regional sweat rate variations [14]. In their seminal work, morphologically normalized and referenced adult values (weight: 70 kg, height, 1.7 m, body surface area, BSA: 1.8 m2) were defined, and the active sweat gland density for the forearm was given as 104 glands/cm2 [14]. The total number of active eccrine sweat glands on the entire body was estimated to be around two million [14]. An average of 1 μL/min/cm2 of local sweat on the forearm then would result in a 0.4 L (~0.4 kg) of sweat loss in 20 min. Our preliminary weight loss measurements revealed that subjects lost at least 0.2–0.3 kg during the exercise. Although these numbers are approximations, it is evident that the values are within the same scale.

The average error in the accuracy of the sweat rate device was found to be 18%. However, it must be noted here that manual data collection with the Macroduct also presents inherent challenges in terms of its own accuracy. The time it takes the sweat to travel from one marked location on the tube to the next was visually observed, which could easily yield data-recording errors. Also, each Macroduct was attached to forearms using stretchable bands. The amount of pressure that is applied to the skin can also affect the functionality of both data collection methods.

All of the subsystems that were used in the IoT framework were off-the-shelf components with open-source software libraries available, and minimal effort would be required if the system were to be modified for a different analysis.

#### **6. Conclusions**

We have presented a wireless real-time sweat rate sensing platform for athlete hydration detection applications. A calorimetric flow rate sensor was utilized in combination with a Macroduct sweat collector to detect the rate of sweat. Data collection was carried out by establishing an IoT framework through either: (1) displaying the sweat rate information on a smartphone app, or (2) displaying on the cloud via ThingSpeak. Controlled flow rate characterization was performed using a syringe pump to calibrate the device. Human subject trials were conducted to validate the device.

The lower limit of sweat rate detection for the current device is around 0.15 μL/min/cm2. Considering that the device was developed for athletes who would sweat more due to heavy exercise, this limit could potentially not be a practical limit in field tests. For example, one of our subjects (subject 4) was a very light sweater, but we were still able to gather her sweat rate information via our device.

The theoretical concept of the calorimetric flow rate system was investigated by our group [36]. Our current work significantly expanded on this knowledge, and we built a standalone real-time sweat rate device. Our implementation of combining the Macroduct sweat collector with the calorimetric sweat rate sensor (inlet from the bottom, and outlet from the top by using a double PDMS mold) is novel. Furthermore, the device was validated with five subjects, resulting in an average error in accuracy of 18%, which is considered to be very promising. Our device is unique in approach and validation, which separates us from camera-recorded sweat rate measurements [29] or the absorbent patch approach [30,31] that was discussed earlier.

We believe the current developments on IoT interfaces and physiological sensor devices will lead to cloud-based health awareness and tracking tools for unique exercise routines. Our current approach will help pave the way for wireless, real-time sweat-based hydration systems.

**Acknowledgments:** This work was funded in part by the National Science Foundation under Grant #1542368. Zach Nelson helped with 3d designs and prints.

**Author Contributions:** Andrew Brueck performed the experiments and contributed to the writing. Tashfin Iftekhar helped with PDMS fabrication. Alicja B. Stannard helped with subject tests and contributed to writing. Kumar Yelamarthi contributed to writing and building and testing the IoT interface. Tolga Kaya wrote the article, performed the data analysis, and supervised the project.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Estimation of Handgrip Force from SEMG Based on Wavelet Scale Selection**

### **Kai Wang 1, Xianmin Zhang 1, Jun Ota <sup>2</sup> and Yanjiang Huang 1,2,\***


Received: 15 January 2018; Accepted: 22 February 2018; Published: 24 February 2018

**Abstract:** This paper proposes a nonlinear correlation-based wavelet scale selection technology to select the effective wavelet scales for the estimation of handgrip force from surface electromyograms (SEMG). The SEMG signal corresponding to gripping force was collected from extensor and flexor forearm muscles during the force-varying analysis task. We performed a computational sensitivity analysis on the initial nonlinear SEMG-handgrip force model. To explore the nonlinear correlation between ten wavelet scales and handgrip force, a large-scale iteration based on the Monte Carlo simulation was conducted. To choose a suitable combination of scales, we proposed a rule to combine wavelet scales based on the sensitivity of each scale and selected the appropriate combination of wavelet scales based on sequence combination analysis (SCA). The results of SCA indicated that the scale combination VI is suitable for estimating force from the extensors and the combination V is suitable for the flexors. The proposed method was compared to two former methods through prolonged static and force-varying contraction tasks. The experiment results showed that the root mean square errors derived by the proposed method for both static and force-varying contraction tasks were less than 20%. The accuracy and robustness of the handgrip force derived by the proposed method is better than that obtained by the former methods.

**Keywords:** surface electromyography; handgrip force; force-varying muscle contraction; nonlinear analysis; wavelet scale selection

#### **1. Introduction**

Various dynamic handgrip tasks are required in a number of applications, including manual fabrication and machining, handling tools, monitoring for surgery, human–robot interactions and so on [1–3]. Usually, these tasks require long duration muscle contractions. With respect to the characteristics of usage, a high level of repetition and force is required in the handgrip tasks, which can be considered a risk factor for work-related upper extremity muscular disorder (WUED) [4]. Therefore, accurately monitoring the handling force in real time is important for preventing WUED [5,6]. To obtain real-time handgrip force measurements, one indirect measurement method is based on surface electromyography (SEMG), which is a common, non-invasive technique for analyzing muscle contractions for real-world applications [4,7,8]. However, SEMG signals are affected by physiological factors and non-physiological factors during prolonged muscle contractions, which affects force estimation from SEMG [1,9,10]. These influence factors are complex and varied under a prolonged muscle contraction environment. Therefore, it is necessary to obtain valid signals from the SEMG to realize accurate prediction of handgrip force in a long duration gripping motion.

In previous studies, a number of correlations between SEMG of a target muscle and force output of a target task have been discussed. In muscle contraction, the activation of motor units within a target muscle follows a mechanism (motor unit rotation and substitution) to maintain a given force. However, muscle fatigue resulting from prolonged muscle contractions can reduce the force generation performance of the target muscles. To overcome the decline in performance, more fresh motor units are recruited to replace the burden of fatigued motor units and compensate for a lack of performance [7,9,11,12]. Many researchers have reported that the increase in amplitude of SEMGs collected from long duration muscle contractions is caused by complex electrophysiological events (i.e., motor unit rotation or substitution), motion artifact, and so on, which, as noises, affect the SEMG signals at different frequencies and reduce the force estimation performance [8,13–17]. According to this phenomenon, many time-frequency and time-scale techniques have been used to eliminate the influence of these noises on force estimation from SEMG.

Many researchers used Fast Fourier Transform (FFT) to observe that the power spectrum of SEMG shifts to a lower frequency range when complex noises (electrophysiological effect of fatigue, random noises) develops [18–21]. The frequency banding technique based on FFT showed that the low frequency band amplitude is increased linearly during endurance isometric contraction tasks [22]. Kumar et al. (2003) found that the amplitude of the SEMG signal in the high frequency band does not change significantly with the development of complex physiological noise, through wavelet analysis [23]. To analyze the changes in SEMG in varying muscle contractions, time-frequency techniques have typically been adopted. A comparison of different time-frequency techniques indicated that continuous wavelet transformation (CWT) produces precise results with good representation of time and frequency localization [24]. The frequency-band analysis technique based on wavelet transformation has been generally used to explore the valid signals from SEMG for accurately estimating force. Sparto et al. (1999) used different wavelet scales of the Daubechies (db) wavelet function of order 6 to report that a decrease in high frequency wavelet coefficients and an increase in low frequency wavelet coefficients may be correlated with fatiguing muscle contractions [25]. Soo et al. (2010) performed a comparison of 55 types of wavelet scale combinations (WSC) and reported that a correlation between two scales (scale 2 and scale 3 of db2) and handgrip force provides a stable estimate of handgrip force from SEMG during endurance isometric contractions [6].

Many CWT-based methods, such as frequency band analysis, have been utilized to obtain a suitable wavelet scale for estimating accurate force from SEMG. These methods have used a single scale or particular WSC to analyze and filter the original signal [14,22]. An original signal can be decomposed into signals under different wavelet scales (i.e., different frequency band ranges) and retains the time domain features through CWT. In practical grasping tasks, each wavelet scale signal contains a variety of complex physiological and non-physiological factors. It is difficult to identify which factors are present at different scales. Currently, the nonlinear differences in quantification between each wavelet scale and the measured force are not clear. Thus, it is necessary to analyze the nonlinear correlations (i.e., nonlinear degree) between each wavelet scale and force in the estimation of handgrip force from SEMG.

The purpose of this paper is to find an appropriate WSC for accurate evaluation of handgrip force from SEMG during force-varying muscle contractions of long duration. The suitable WSC was obtained by using the sequence combination analysis (SCA) method based on Monte Carlo sensitivity analysis. As a simple and reliable sensitivity analysis method, Monte Carlo simulation can calculate the degree of nonlinearity between variables and the actual target output, regardless of how many factors the variable contains [26,27]. We first obtained the nonlinear correlations between wavelet scales and handgrip force during prolonged force-varying muscle contraction in the Monte Carlo sensitivity analysis, and then selected the appropriate WSC by analyzing the nonlinearity degree of each correlation. Finally, we evaluated the proposed method by comparing it to former methods. The root mean square error (RMSE) was used to evaluate the difference between the estimated force and the measured force during the WSC selection procedure, as well as comparisons between the

proposed method and former methods. The contribution of this study is that we have investigated the nonlinear correlations between the wavelet scale and handgrip force. These correlations can enhance the selection of an appropriate WSC to accurately estimate handgrip force during force-varying muscle contractions of long duration.

#### **2. Experimentation and Methodology**

#### *2.1. Experimental Setup*

Twenty healthy participants (16 men and 4 women), whose ages ranged from 22 to 31 years, participated in the experiment. At this stage, we did not consider the differences between males and females. Ten subjects (group A: 8 men and 2 women) were used for determining the suitable WSC for estimating handgrip force in the force-varying contractions as a training group. The other ten subjects (group B: 8 men and 2 women) were used as a validation group, to evaluate the performance of the suitable WSC. Before the experiment, all participants were informed of the experiment process and the risk of muscle fatigue, and agreed to perform the experiment.

In the gripping tasks, SEMG data was recorded from the flexor digitorum superficialis (FDS) muscle and extensor carpi radialis (ECR) muscle of the forearm, as shown in Figure 1. These are the dominant forearm muscles that contribute to handgrip force [28,29]. The SEMG was measured using reusable integral bipolar dry surface electrodes (Green Sensor Electrodes SX230, Biometrics Ltd., Newport, UK) with a skin contact area of 10 mm<sup>2</sup> for each electrode and a center-to-center recording distance of 20 mm. Electrode locations were determined in accordance with a previous study [6,29]. EMG signals were amplified 20 times and bandpass prefiltered (15–450 Hz), together with the force signals that were obtained at a sampling rate of 1000 Hz (14 bits), using a portable electrophysiological amplifier system (Biometrics Ltd., DataLOG-MWX8-Bluetooth, UK, input resistance > 1015 ohms, CMRR-60 Hz > 96 dB) and stored on a computer. Handgrip forces (kg) were measured using an adjustable handgrip dynamometer. The sampling rate was set to 1000 Hz, and the data was digitized at a 14-bit resolution. To reduce the influence of grip width, a grip width of 65 mm was selected [5]. The force signal was visualized online, on a computer screen, to provide feedback about the exerted force by subjects during the data collection procedure, as shown in Figure 1.

**Figure 1.** Experimental setup. SEMG: surface electromyogram.

#### *2.2. Experimental Procedure*

At the beginning of the experiment, each participant was asked to perform a maximal voluntary contraction task (MVC) three times, and from this, the average value of maximum grip strength (kg) and maximum amplitude (mV) of the SEMG were calculated. In the experiment, the participants in Group A first performed the force-varying analysis task to obtain the suitable WSC using the proposed method. Then, all participants in Groups A and B performed the static and force-varying validation

tasks to evaluate the performance of the proposed method. The force-varying analysis task, static validation task, and force-varying validation task are defined as follows.

2.2.1. Force-Varying Analysis Task (Force-Varying Contraction)

Participants in Group A were required to perform a force-varying fluctuating contraction. In the force-varying analysis task, a series of varying handgrip forces, occurring within ten minutes, was increased or decreased respectively by squeezing and releasing the dynamometer. Participants were asked to increase their handgrip force gradually to provide an adequate correlation between handgrip force and SEMG for each minute, as shown in Figure 2.

**Figure 2.** The process of the force-varying analysis task.

2.2.2. Static Validation Task (Isometric Muscle Contraction)

Static tasks were performed at four force levels, which included 20%, 30%, 50%, and 70% of the MVC in a random order during the isometric contraction [6,29]. The participants sustained these contractions until the generated force dropped below 5% from the target force level for 5 s. The experiment with each force level was executed only once. To avoid the effect of muscle fatigue, participants were allowed to rest for 15 min before performing the next force level experiment [6].

#### 2.2.3. Force-Varying Validation Task (Voluntary Contraction)

In the force-varying validation task, a force level that varied between 0% and 80% of the MVC was performed over a 10-min period. Before the force-varying validation task, the participant rested for 60 min. A sample of the recorded handgrip force and SEMG for both muscles is illustrated in Figure 3.

**Figure 3.** The process of the force-varying validation task.

#### *2.3. Proposed Method*

To accurately estimate the handgrip force from the SEMG signals during the long duration force-varying grip motion, this paper utilized the Monte Carlo sensitivity analysis to determine the nonlinear degree of the relationship between each wavelet scale and handgrip force from the initial nonlinear SEMG-handgrip force relationship, and used a nonlinear correlation-based scale-selection method (SCA) to obtain the suitable WSC of the scale.

#### 2.3.1. Initial Nonlinear SEMG-Handgrip Force Relationship

The initial nonlinear SEMG-handgrip force relationship was built using the ten wavelet (db2) scales (scale number: 1 2, 3, 4, 5, 6, 7, 8, 9, 10) from the SEMG signals and handgrip force collected from the force-varying analysis tasks of group A. These wavelet scales are most relevant to handgrip force [23]. This model is based on the second-order polynomial [29,30]. Given the predictors *e*(1), ... , *e*(*P*), the response of handgrip force *F*(*estimation*) is predicted by:

$$F(estimation) = \sum\_{p=1}^{P} \left(\theta\_{1p}\varepsilon(p) + \theta\_{2p}\varepsilon^2(p)\right) \tag{1}$$

where 1 ≤ *p* ≤ *P* = 10, *P* is the total number of wavelet scales, and the vectors of coefficients *θ*<sup>1</sup> = - *θ*11,..., *θ*1*p*,..., *θ*1*<sup>P</sup>* and *θ*<sup>2</sup> = - *θ*21,..., *θ*2*p*,..., *θ*2*<sup>P</sup>* are produced by ordinary least squares. In this paper, the predictors, *e*(*p*), are the wavelet coefficients' intensities of the wavelet scale of *p*, which are collected from the original SEMG *xn,* with data size, *N*, through CWT and root mean square (RMS)

$$c\_n(p) = \sum\_{n=0}^{N-1} x\_n \psi^\* \left[ \frac{(n'-n)\delta t}{p} \right] \tag{2}$$

where 0 ≤ *n* ≤ *N* − 1, *cn*(*p*) is the wavelet coefficient, *δt* is the sampling time, *ψ* is the mother wavelet, and (\* ) indicates the complex conjugate [31]. Then, the valid intensity of wavelet coefficients can be calculated after pre-processing of the RMS with the window frame size of *M*

$$c\_n(p) = \sqrt{\frac{1}{M} \sum\_{n=kM}^{(k+1)M-1} c\_n(p)^2} \tag{3}$$

where 0 <sup>≤</sup> *<sup>k</sup>* <sup>≤</sup> *<sup>N</sup> <sup>M</sup>* − 1. A moving average filter (200 ms) [32] was used to smooth the *en*(*p*).

#### 2.3.2. Sensitivity Analysis

For a nonlinear index, the sensitivity analysis is important [26]. In contrast to other sensitivity analysis methods, Monte Carlo simulation, as a variance-based global sensitivity analysis method, can obtain the sensitivity of interactions between the parameters [27]. This paper used the multivariate Monte Carlo simulation to obtain the sensitivity value of each wavelet scale related to force, predicted from the original nonlinear model (Equation (1)). The Monte Carlo simulation is often used to calculate the expected estimation of parameters in known mathematical models with random variables [26,27]. The main steps of Monte Carlo simulation can be categorized as follows:


In this paper, the random variable was distributed evenly in a specific range (0–100) as a basic step to realize the Monte Carlo simulation. In each iteration of the simulation, each model parameter used a given number of random variables to obtain the expected sensitivity value. The given number of random variables was obtained by a finite difference-based convergence experiment for all participants in Group A [27]. This experiment showed a relationship between the number of random variables and the expected sensitivity estimator, as shown in Figure 4. This experiment was performed at a specific variable number range, *r*, that was between 100 to 140,000, with an interval of 100. By using this range, *r*, the sensitivity of each wavelet scale was obtained from the Monte Carlo simulation at each given variable quantity and these sensitivities were calculated by first order differential calculation.

$$d\_i(p) = |y\_{i+1}(p) - y\_i(p)|\tag{4}$$

$$i = \frac{r}{100} \tag{5}$$

where *p* = 1, 2, ... , 10, is the ordinal number of the wavelet scale, *r* = 100, 200..., 140,000, and *yi*(*p*) is the sensitivity value of the wavelet scale of *p* with a specific random variable number (*r* = *i* × 100). Then, the sum, *di*(*integration*), of the sensitivity differences under a variable number for Group A can be calculated as follows:

$$d\_i(\text{integration}) = \sum\_{p=1}^{10} d\_i(p) \tag{6}$$

**Figure 4.** The results of the convergence experiment for determining the number of random variables.

Figure 4 depicts the influence of the number of random variables on the sensitivity distribution. The mean difference (red line) decreased as the number of random variables increased, and tended to stabilize from 60,000. The upper difference (blue line) and the lower bound difference (yellow line) became stable when the number of random parameters exceeded 100,000. Therefore, this paper used 100,000 random parameters, which were uniformly distributed between 0–100, to perform the Monte Carlo simulation for analyzing the sensitivity of each wavelet scale to handgrip force.

#### 2.3.3. Sequence Combination Analysis (SCA)

Based on the Monte Carlo simulation, we obtained the nonlinear degree of each scale-handgrip force correlation through the force-varying analysis tasks. To choose the suitable WSC, we designed a set of scale permutations and combination rules based on the mean value of sensitivity of each scale. The mean sensitivity value was calculated from all participants in Group A. The distribution and mean sensitivity value of each scale is shown in Figure 5. Through frequency band analysis, Soo et al. (2010) observed that wavelet scales 2 and 3 have large effects on fatigue resistance during muscle isometric contractions [6]. By using the Monte Carlo sensitivity analysis, it was shown that these two scales have lower sensitivity to handgrip force than other scales. Therefore, this paper determined a specified sequence and combination rule of the wavelet scale. The rule is that the scale with low sensitivity is preferred for retention and the scale with high sensitivity is eliminated, in turn. By considering the wavelet scales described in Section 2.3.1, we can obtain ten WSCs (combination serial numbers I, II, III, IV, V, VI, VII, VIII, IX, X) based on the proposed sequence combination analysis. The details of the WSCs are illustrated in Figure 6. The coefficient tern structure and parameters of the original nonlinear SEMG-handgrip force relationship were adjusted to match the different WSCs. This paper uses RMSE as the evaluation index to determine which WSC has the appropriate estimation accuracy for handgrip force.

#### *2.4. Performance Comparison*

#### 2.4.1. Former Method

In this paper, two former methods were utilized to be compared with the proposed method. Formerly, force estimation was realized by calculating the amplitude of the SEMG signal, for example, using RMS and the pre-defined SEMG-force relationship. Former method A builds a linear SEMG-force relationship by using the normalized SEMG signal, which is bandpass-filtered between 15 and 450 Hz [5]. Former method B builds a SEMG-force model based on the frequency-band technique, which determines a suitable WSC (wavelet scales 2 and 3) to estimate the handgrip force from SEMG signals during an isometric muscle contraction task [6].

#### 2.4.2. Performance Index

To obtain the estimated error, a comparison between the estimated and actual handgrip force was conducted using the RMSE:

$$\text{RMSE} = \sqrt{\frac{1}{T} \sum\_{t=1}^{T} \left( F(actual)\_t - F(estimated)\_t \right)^2} \tag{7}$$

where *F(actual)* is the sample data of the observed griping force from the grip dynamometer and *F(estimation)* is the predicted force from SEMG. The sampling number of handgrip force is 1 ≤ *t* ≤ *T*. A value of RMSE closer to zero indicates that the model has a smaller random error component.

#### **3. Results and Discussion**

#### *3.1. Nonlinear Correlation between Wavelet Scale and Handgrip Force*

Figure 5 describes the distribution between the wavelet scale and its sensitivity to handgrip force for Group A. These sensitivity values can be used to evaluate the degree of nonlinear correlation between the wavelet and handgrip force. According to the observation of the sensitivity distribution of the two muscles, scales 1, 2, 3, 4 have lower degree nonlinear correlations with handgrip force, the degree of nonlinear correlation between scales 6, 7, 8, 9, 10 and handgrip force are generally higher, and scale 5 is in the middle sensitivity position. According to Soo et al. (2010), the four wavelet scales with a low degree of nonlinear correlation to handgrip force are in the 181.82–727.27 Hz central frequency range, the five scales with a higher degree of nonlinear correlation to handgrip force are between 72.72 Hz and 121.21 Hz of central frequency and the center frequency of scale 5 is in the middle at 145.45 Hz [6]. Wakeling et al. (2002) reported that the frequency ranges for recruitment of slow and fast twitch muscle fibers are approximately 143 and 397 Hz of SEMG, respectively [33]. Thus, it can be approximated that slow fiber signal actions at scale 5, and fast fibers, are located at low sensitivity scales. Begg et al. (2007) reported that the SEMG of low frequency signals contain motion artifacts [34], which may result in high sensitivity. Many researchers reported that the power spectrum of SEMG shifts to a lower frequency range when muscle fatigue develops [11,18,19]. According to the center frequency of the wavelet scale, the wavelet scales 6–10 belong to the low frequency range of the SEMG signal. This is the reason why the wavelet scales of low frequency signals (i.e., scale 6–10 in Figure 5) have a high degree of sensitivity.

From the results shown in Figure 5, we can observe that each scale has a different sensitivity distribution. To simplify the distribution, we calculated the mean sensitivity of each scale to express the relationship between the ten wavelet scales and their sensitivity to handgrip force. Then, wavelet scales were arranged, in ascending order, in regard to mean sensitivity value, as shown in Figure 6.

**Figure 5.** Sensitivity distribution of the ten wavelet scales to handgrip force.

#### *3.2. Selection of Wavelet Scale Combination*

Figure 6 describes a layout of the SCA. The layout shows the sequence and combination of the ten wavelet scales. For instance, the WSC II covered all scales, except scale 8, and the WSC X only had scale 2. The performances of force estimation under the ten WSCs were evaluated by RMSE during the force-varying validation tasks. The RMSE was calculated based on the difference between the handgrip force estimated from SEMG and the handgrip force measured from the grip dynamometer.

**Figure 6.** Sequence combination of wavelet scales based on the sensitivity value.

The performance evaluation consisted of a varying RMSE process of the ten WSCs during the force-varying validation task. Figure 7 illustrates that the performance evaluation of the ten WSCs is a time-varying RMSE process. Each WSC has a different error curve during the force-varying validation task. Figure 7 shows that WSC I, II, III, IV, and V obtained from the extensor muscles have unsatisfactory estimated accuracies, and WSC VI to X have similar estimation accuracies in the amplificatory graph. In comparison to the five WSCs (I, II, III, IV, and V), the removal of highly sensitive wavelet scales can help to improve the estimation accuracy of handgrip force over a long duration griping motion. WSC I, II, III and IV obtained from the flexor muscle have unsatisfactory accuracies of estimated force, and WSC V to X have similar estimated accuracies. To choose the suitable WSC, a statistical method was used to evaluate the differences between estimated accuracy for ten WSCs, as shown in Figure 8.

**Figure 7.** Varying RMSE process of each wavelet scale combination (WSC) in the force-varying validation task.

**Figure 8.** Comparison of the ten WSCs.

Figure 8 shows the handgrip force estimation error that was derived from Group A under different WSCs. The WSCs of I, II, III, IV, and V in the extensor muscle have larger predicted errors of handgrip force than the others. However, after amplification of local errors, it was found that the minimum force estimation error occurred in WSC VI. The suitable WSC selection for the flexor muscle is similar to that for the extensor muscle. The WSC V of the flexor muscle is considered to have the minimum RMSE range and mean in estimation of the long duration handgrip force from SEMG.

#### *3.3. Performance Comparison*

Figures 9 and 10 show the comparisons between RMSE of handgrip force, estimated by the proposed method and former methods, in both force-varying and static validation tasks. Figure 9A,C present the RMSE obtained by the participants in Groups A and B within the first minute of the force-varying validation task. In the first minute of the task, the SEMG is affected by a low degree of physiological effect of fatigue and motion artifact. Figure 9B,D shows the performance comparisons in the last minute of the validation task. In the last minute of task, the complex noise (electrophysiological effect of fatigue, motion artifact, and so on) seriously affects SEMG amplitude and frequency, and reduces the correlation between SEMG and handgrip force. This is the reason why the RMSE in the first minute is smaller than that in the last minute. Figure 10 shows the RMSE derived by the proposed method and former methods during the static task. Since the static task is more likely to produce fatigue than the force-varying task, we compared the estimated accuracy of these three methods in the first 10 s and the last 10 s of the isometric contraction task, at four force levels. The results show that the RMSE in the first 10 s is smaller than that in the last 10 s.

**Figure 9.** Results of the force-varying validation tasks.

**Figure 10.** Results of the static validation tasks.

In the force varying validation task, the experimental results (Figure 9) show that the proposed method has a better performance than the others in terms of force estimation, especially in the force range of 10–25%. This is because the proposed method can obtain a better pre-defined SEMG-handgrip force relationship than the former methods. In contrast with the high degree of force (greater than 41% MVC), the performances of these three methods are slightly different. Because high force muscle contraction requires a large number of motor units to maintain the particular level of force, the high force muscle contraction is different from the low force contraction. Based on the principle of motor unit recruitment [1,9], the three methods can obtain a similar pre-defined SEMG-handgrip force relationship at a high force level. Since former method A contains all wavelet scales, except for scale 1, its pre-defined SEMG-handgrip force relationship is based on the scales with considerable noise (motion artifacts and random noise); thus, the performance of this method is weak compared to the other methods. The similar results can be obtained in the static validation task, as shown in Figure 10.

The proposed method has a significant performance improvement for fatigued muscle contractions with long durations (see Figures 9B,D and 10B,D). The proposed method reduces the force-varying validation RMSEs by 27%, 53%, 54%, and 61%, compared to former method A, in the handgrip force ranges of 56–70%, 41–55%, 26–40% and 10–25% of MVC, respectively. The RMSE of the proposed method is reduced by 4%, 2%, 11% and 21%, respectively, in these four ranges of force, compared to former method B. In the static validation tasks, the proposed method can significantly improve the performance of force estimation at the low force level. A large amount of noise caused by the electrophysiological effect of fatigue and motion artifacts affects the amplitude and frequency of SEMG, which causes the pre-defined SEMG-handgrip force relationship built by former methods to no longer be accurate.

Compared to the first minute, the estimated accuracy of former method B in the last minute decreased by an average of 11% in the force range of 10–40% of MVC. From the perspective of sensitivity of each scale to handgrip force, this is because the wavelet scales 2 and 3 have lower sensitivities to gripping force prediction than the other scales and thus, cannot characterize flexibles change during force-varying contractions for low forces. The signal noise induced by the intramuscular electrophysiological effect and motion artifacts has consequences for the correlation between the two scales and handgrip force and further reduces the estimated performance of former method B. Previous studies have shown that low strength muscle contractions are mainly controlled by slow-twitch muscle fibers (approximately 143 Hz of SEMG), and high strength muscle contractions are mainly controlled by fast-twitch muscle fibers (approximately 397 Hz) [33,35]. In former method B, the SEMG signal of 143 Hz is missing due to the limitation of scale pre-layout in frequency band analysis technology [6]. This is the reason why the performance of former method B is worse than our proposed method.

In contrast to the former methods, the proposed method can effectively inhibit the influence of complex electrophysiological noises on the predicted accuracy of the handgrip force from SEMG during long duration fatiguing muscle contractions. However, the prediction errors for low-level forces are greater than those for high level forces. The reason for this is that the intramuscular electrophysiological effect and motion artifacts can reduce the degree of correlation between different levels of force and SEMG to varying degrees. During the long duration muscle contractions, all participants overcame fatigue to produce a specified handgrip force with extra shaking of the forearm. Motion artifact noise may be caused by the shaking of these forearm muscles. When the fatigue reduces the force generated from the muscle, the rotation and substitution between fatigued motor units and fresh motor units is performed to generate and maintain the requisite force level. In this situation, the complex electrophysiology within muscles affects the changes in the SEMG signal. Thus, the intramuscular electrophysiological effect and motion artifacts ultimately change the amplitude and frequency of the SEMG signal.

#### **4. Conclusions**

In this paper, we proposed a nonlinear correlation-based method to select appropriate wavelet scales to accurately estimate handgrip force from SEMG during a force-varying muscle contraction of long duration. The sensitivity analysis of the Monte Carlo simulation was used to explore the nonlinear correlation between ten wavelet scales and handgrip force. A nonlinear, correlation-based SCA method was proposed to determine the suitable WSC. The results derived via SCA indicated that the suitable scales for the extensor muscles is WSC VI, consisting of scales 1 to 5, and WSC V including scales 1 to 6 is suitable for flexor muscles. The appropriate WSC, obtained by the proposed method, was compared to two former methods through a set of static and force-varying contraction tasks. The experimental results showed that the proposed method has a better estimation performance than the former methods, especially for low level contractions (0–30% maximal voluntary contraction).

The proposed method enhances the performance of grip strength estimation from SEMG. However, the actual working muscles not only provide force output but also drive limb posture changes. Further research using the proposed nonlinear correlation-based scale selection method is needed to explore the correlation between wavelet scales of SEMG and muscle loads or limb posture in various work tasks, such as human–robot interactions, and gait analysis based on SEMG of lower limb muscles. An accurate force estimation based only on electromyography data can be used for gait analysis instead of using inverse dynamics or similar methods [36,37]. In addition, the participants in this study were limited in age, health, and sex ratio (i.e., 16 healthy male students and four healthy female students); thus, further research should take into account the gender differences in the results of the proposed method.

**Acknowledgments:** This work was supported by the [Natural Science Foundation of China] under Grant [51505151,91748111]; [Natural Science Foundation of Guangdong Province] under Grant [2015A030310239].

**Author Contributions:** Kai Wang, Xianmin Zhang, Jun Ota and Yanjiang Huang conceived and designed the experiments; Kai Wang and Yanjiang Huang performed the experiments and analyzed the data; Kai Wang and Yanjiang Huang wrote the paper; Xianmin Zhang, Jun Ota and Yanjiang Huang revised the paper. All of the authors have read and approved the final manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **An Automatic Gait Feature Extraction Method for Identifying Gait Asymmetry Using Wearable Sensors**

#### **Arif Reza Anwary 1, Hongnian Yu 1,\* and Michael Vassallo <sup>2</sup>**


Received: 6 January 2018; Accepted: 20 February 2018; Published: 24 February 2018

**Abstract:** This paper aims to assess the use of Inertial Measurement Unit (IMU) sensors to identify gait asymmetry by extracting automatic gait features. We design and develop an android app to collect real time synchronous IMU data from legs. The results from our method are validated using a Qualisys Motion Capture System. The data are collected from 10 young and 10 older subjects. Each performed a trial in a straight corridor comprising 15 strides of normal walking, a turn around and another 15 strides. We analyse the data for total distance, total time, total velocity, stride, step, cadence, step ratio, stance, and swing. The accuracy of detecting the stride number using the proposed method is 100% for young and 92.67% for older subjects. The accuracy of estimating travelled distance using the proposed method for young subjects is 97.73% and 98.82% for right and left legs; and for the older, is 88.71% and 89.88% for right and left legs. The average travelled distance is 37.77 (95% CI ± 3.57) meters for young subjects and is 22.50 (95% CI ± 2.34) meters for older subjects. The average travelled time for young subjects is 51.85 (95% CI ± 3.08) seconds and for older subjects is 84.02 (95% CI ± 9.98) seconds. The results show that wearable sensors can be used for identifying gait asymmetry without the requirement and expense of an elaborate laboratory setup. This can serve as a tool in diagnosing gait abnormalities in individuals and opens the possibilities for home based self-gait asymmetry assessment.

**Keywords:** inertial measurement unit; accelerometer; gyroscope; asymmetry; feature extraction; wearable sensors; gait analysis

#### **1. Introduction**

Gait asymmetry (GA) is an indicator of different diseases and disease progression. It results in reduced gait efficiency and activity levels. Objective assessment of GA is important in the treatment and rehabilitation of patients with various conditions such as falls or after orthopaedic surgery. Gait is the result of a series of rhythmic alternating movement of arms, legs, and trunk which create forward movement of the body [1]. It relies on complex mechanisms depending upon the closely integrated actions of musculoskeletal, nervous system (central and peripheral), visual, vestibular, auditory systems; joint mobility and the smooth propulsive movement of the center of gravity. Every individual's gait pattern should be symmetrical with right and left sides performing identical movements. This is not the case since every individual has a unique gait pattern and the limb movement of one side is not exactly repeated on the other side, which results in GA. GA analysis provides bilateral locomotive information of gait parameters (e.g., length and period of stride, step, stance and swing), kinematic and kinetic measurements (e.g., angular joint trajectories, angular joint velocities, joint forces, and reaction forces), muscular measurements (e.g., muscle contraction, muscle force) and energy expenditure (e.g., oxygen consumption, heart rate) [2]. This is employed in different

domains such health, sports and rehabilitation. GA can be a determinant of recovery in patients with different diseases such as Parkinson's disease [3] and stroke [4]. It can be used to monitor and improve an athlete's performance [5] as well as a patient's progress in orthopedics and rehabilitation [6]. In biometrics and biomedical engineering areas, gait analysis is used as an assistive tool to characterize human locomotion [7]. GA is important in elderly patient fall risk assessment [8], and is a predictor of functional and cognitive decline [9].

The tools and methodologies used to assess GA are often arbitrary [10] and often studied in non-natural controlled conditions. Currently available methods include reflective skin markers [11], ear-worn sensors [12] and camera based motion analysis systems [13]. Other methods for measuring GA include ground reaction forces [14], dynamic electromyography [15], and instrumented walkways [16]. These methods are laboratory based, complicated, costly, and often carried out by technical or clinical staff, which may be difficult for patients to use at their homes.

Clinical scales used to analyse gait parameters are subjective or semi-subjective and a poor replacement to laboratory based methods for identifying changes in GA. Different assessment tools such as the Gait Abnormality Rating Scale [17], Figure of 8 Walk Test [18], Four Square Step Test [19], The Functional Gait Assessment [20], Groningen Meander Walking Test [21] and Berg Balance Scale [22], are used to observe a patient's gait and balance. Their use has a high chance of intra and inters observer variation and human error. This may affect the accuracy of diagnosis, follow-up and treatment [2]. Therefore, a more objective way of assessing GA is required.

A variety of wearable sensors including accelerometer, gyroscope, magnetometer, foot pressure sensor, inclinometer, and goniometer [23,24] are generally used to measure various characteristics of human gait. Inertial Measurement Unit (IMU) sensors are used in different situations such as monitoring of post-operative gait abnormalities [25], stride variability [24], measurement of GA [26], fall-related gait characteristics measured on a treadmill in daily life [27], nature of the Parkinsonian gait [28] and human waking foot trajectory [29]. Accelerometer based gait parameters such as times of stance, swing, single support and double [30]; stride and stance phase [31]; gait velocity, cadence and step length [32]; step number, moving distances, every step instant speed and average speed [33]; step counting [34,35]; times of heel strike, toe strike, heel-off, and toe-off [36]; stride length and duration [37]; walking distance, time and speed [38] were investigated.

Although, IMU based gait analysis methods are available, specific GA was actually found in few studies [30]. Some studied a single gait parameter [24], single sensor [12] or applied simple statistical methods for comparisons [38]. Force-sensitive resistors placed in insoles to detect ground contact were used to estimate the stance time for GA [39]. Microsoft Kinect-based GA [13], IMU and pressure sensitive shoe insole was used to detect gait onset and toe-off detection [40] and IMU-based knee flexion/extension angle measurements [41] and GA using gyroscopes [42] were also used. Systems developed for the detection of GA require a fully automatic system for data collection, feature extraction and quantitative measurements where both limbs are evaluated.

In order to broaden the use of accurate quantitative GA monitoring in clinical screening and research, an affordable GA tool is required which can be used in clinic or home. This study aims to design and implement an automatic lower limb gait features extraction method based on accelerometer and gyroscope data to increase the reliability and validity of monitoring GA. We set out to develop an affordable multi-sensor based synchronous data collection system for a comprehensive physical gait analysis extracting 24 commonly reported GA features. We develop a novel android app for collecting synchronous accelerometer and gyroscope data from both legs. Features include total distance, total time, total velocity, stride, step, cadence, step ratio, stance, and swing. We also estimate the mean, standard deviation, variance, minimum and maximum values. The paper has the following contributions: (1) designing and developing a novel android app to collect real time synchronous IMU data from legs; (2) proposing a gait asymmetry feature extraction method (a novel stride detection technique, a stance and swing detection technique, and a method for estimating travelled distance); and (3) creating a data set using our developed app and designed MetaWear casing, Velcro elastic belt, and buckles for validating the designed app and the proposed method. The paper is organized in the following sections: Section 2 presents the design and the methods of the proposed system. Section 3 delivers the experimental results and discussion. The conclusion is given in Section 4.

#### **2. Design and Methods**

#### *2.1. Subjects Selection*

We recruited a convenience sample of 20 subjects with 10 healthy young subjects (nine male, mean age 25.3 years, standard deviation 4.64, range 19–35 years), and 10 older subjects (nine male, mean age 69.4 years, standard deviation 7.28, range 62–86 years). Older subjects 1, 3, 5, 6, 9 and 10 did not have any health problems. Subject 2 has a right foot drop and drags the foot and toes. Subject 4 has pain in the right leg lower muscle and walks without any support. Subject 7 has pain in the lower part of his left leg and uses a crutch during walking. Subject 8 has pain in both ankles and walks with support of a walker. According to World Health Organization (WHO), the life expectancy at birth is 71 years in Bangladesh [43]. Therefore, 65+ is considered old age in Bangladesh although would be viewed as young old in the Western countries. The subjects are purposefully chosen for this study to provide a variety of gaits for evaluation.

#### *2.2. Experimental Protocol and Calibration*

The experiment is performed in two different locations for young and older. The older subjects are residents in a care home. All subjects perform a walk in a straight corridor comprising of 15 strides of normal forward walking, a turn-around and another 15 strides. Accelerometer and gyroscope data from sensors attached on two foot locations are recorded in a database synchronously using our Android app. The distance carried out by walking on the corridor is measured by a tape. The several older subjects perform less than 15 strides. Calibration was performed individually where the distance travelled is measured manually and the result compared to the output from the sensor. The Qualisys motion capture system (Qualisys AB, Kvarnbergsgatan 2, Gothenburg, Sweden) is used for calibration in the validation stage as well.

#### *2.3. Sensor Placing Location*

From our investigation [44], it is found that placing a sensor in different foot locations gives quite different signal patterns. It is also observed that the orientation of the sensor has a significant effect on output data. In order to increase the sensor accuracy and reliability, and reduce the variability, all sensors are fitted tightly to the barefoot. We choose the barefoot rather than sensors attached to a shoe because wear and tear in the shoe can affect the position of the sensor and accuracy of the data output. The plantar aspect is not covered as this is the part of the foot in contact with the floor and is not practical for the subject to walk on the sensor. In this study, the sensors are placed at metatarsal foot locations of both legs (Figure 1) for collecting data since the best performance [44] can be achieved.

**Figure 1.** IMU sensors placement in right and left metatarsal foot locations of the barefoot.

#### *2.4. Sensor Selection*

There are wearable wireless IMU sensors commercially used for health rehabilitation and activity monitoring [45]. An affordable wearable Bluetooth, long autonomy, minimum consumption, multiple synchronized data transmission supported IMU sensor is important for our investigation. With all these considerations the recently introduced sensor MetaWear CPro (MBIENTLAB Inc., San Francisco, CA, USA) [46] (price: \$30.00) was chosen. The power consumption of our sensor is low during sleep mode. The sensor (Figure 2) is in an active state when connected by Bluetooth to our android device and only goes to sleep mode once it is disconnected. It is sensitive to acceleration and rotatory movements that occur during normal human locomotion.

The IMU sensors are sampled at a frequency range of 20 Hz to 200 Hz [47]. In practice, a low sampling rate for the accelerometer possibly produces excellent recognition and accuracy in posture and activity classification [48]. For this study, the accelerometer range is ±8 m/s<sup>2</sup> and gyroscope range is ±500 degrees/s with the sampling rate of 50 Hz. A casing for the sensor is designed using SolidWorks [49] and printed using a 3D printer. A Velcro elastic belt and buckles are used to adjust and attach the sensor (Figure 2).

**Figure 2.** Proposed MetaWear casing, Velcro elastic belt, buckles and IMU sensor: (1) Buckle and Elastic Belt: the buckle is sewn onto an elastic belt for fastening to Velcro; (2) Bottom case which keeps the sensor safe from pressure, temperature and water; (3) Lock Open Edge which helps to open the cover from bottom case; (4) Sensor Lock Mechanism: The four locks keep the sensor sideways movement and orientation; (5) Cover Lock Mechanism which tightly locks with the case; (6) Velcro-Elastic Joint: The elastic belt is sewed with Velcro; (7) Velcro which adjusts and tighten when the sensor is attached; and (8) IMU sensor and battery.

#### *2.5. Android App Design, Development and Data Collection*

The sensor provides an Android API library for interacting with the MetaWear board on an Android mobile phone. A minimum of Android 4.3 (SDK 18) is required to use this library, however some features will not properly function due to the underlying Bluetooth LE (BLE) implementation. For getting the best results, it is recommended to use an Android 4.4 (SDK 19) or higher and BLE 4.0 or higher. Based on these criteria, an Android app is developed using Android Studio 2.2 to collect real time accelerometer and gyroscope data, and store data on an external SD card in a csv file. The data soring format in a csv file is date (dd/mm/yyyy), time (HH:MM:SS.ss), system clock (milliseconds), accelerometer (X, Y, Z) and gyroscope (X, Y, Z), shown in Figure 3.

The app is designed through the following steps: (1) requirement analysis through the literature review and interview with the expert and users; (2) reviewing the market available software development platform; (3) initial design of the mobile app using Android Studio 2.2; (4) testing and debugging with the user feedback for improvement; (5) documentation for development. We develop the proposed app shown in Figure 3. The details of the design are beyond the scope of this paper and will be reported in the future.

Pressing the SCAN button the availability of all the devices for data collection is checked. The order of data collection is selected by Slot number. Each sensor then automatically connects

with the corresponding mac address by showing CONNECTED. Pressing the DATA RECORDING switch makes a dialog box to get the file name for storing data. Pressing the OK button starts data collection. Pressing the STOP button stores the collected data.


**Figure 3.** Proposed android app to collect data from MetaWear CPro.

A stride is a whole gait cycle with stance and swing events. This horizontal movement produces high acceleration during walking and this movement is the subject of investigation in this study for GA monitoring. The data with the horizontal movement information from the feet are analysed using our method to find GA information. Figure 4 shows the raw data.

**Figure 4.** Raw accelerometer and gyroscope data from right and left feet of older subject 1.

The red, green and blue lines in Figure 4 stand for accelerometer readings on *x*, *y* and *z* axis, respectively, with *g* units (9.81 m/s2) in the sensor frame. We can observe from the raw data that the accelerometer reading on *x* is the highest and *z* is the lowest before the commencement of walking for right leg indicative of the initial gravitational force. Similarly, the accelerometer reading on *y* is the highest and *z* is the lowest before the commencement of walking for left leg. The initial data is not aligned to zero means that the sensors are not placed perfectly upright position with the earth frame in the foot locations due to the initial gravity part of *y* and *z*. For this study, the sensors do not need to be perfectly upright which in any case is not user-friendly and impossible. The discrepancy between the sensor frame, the foot frame and the earth frame are compensated for in this study.

#### *2.6. Gait Asymmetry Feature Extraction Method*

#### 2.6.1. Raw Data Processing

It is noted that accelerometers are sensitive to altitude and impact forces, while gyroscopes are sensitive to temperature changes and suffer from a low-changing bias. Consequently accelerometers have poor dynamic features and gyroscopes have poor static features [50]. To provide a robust absolute orientation vector in the form of quaternion or Euler angles, the sensor combines the measurements from 3-axis accelerometer and 3-axis gyroscope sensors. The algorithm in [46] fuses the raw data in an intelligent way to improve each sensors output. This includes offset calibration of each sensor, monitoring of the calibration status and Kalman filter fusion to provide distortion-free and refined orientation vectors [46]. The IMU sensor provides accelerometer *A*(*ax*, *ay*, *az*) and gyroscope *G*(*gx*, *gy*, *gz*) with respect to time *t*. As the accelerometer is sensitive to acceleration due to movement and the local gravitational force, the input data consists of the user acceleration and gravitational acceleration.

#### 2.6.2. Coordinate Systems

In this study, there are three coordinate systems, the foot frame describing the foot rotation, the sensor frame describing the motion of the sensor and the global or Earth frame. Since the sensor is attached to the foot tightly using an elastic Velcro belt, we assume that the sensor does not slip or move during walking time. Therefore we consider that the foot frame and sensor frame are the same. Our approach is to transfer the sensor frame to the Earth frame and then to remove the gravitational component. The high gravitational force of the Earth frame is downward towards Earth. The *Ax* axis is aligned along the foot axis of the IMU sensor, *Az* points downwards so that it is aligned with gravity so that the three axes from a right handed coordinate system shown in Figure 5a.

#### 2.6.3. Quaternion

Quaternion is a concept related to the foundations of algebra and number theory. While the accelerometer and gyroscope sensors enable the tracking of translational and rotational movements, the accurate measurement of the sensor orientation is important to interpret sensor information. Quaternions are a mathematical construct that consist of four individual numeric complex number components that can be used to represent the orientation of a ridged body or coordinate frame in a three dimensional space.

**Figure 5.** (**a**) Sensor frame and earth frame of accelerometer and gyroscope axes; (**b**) The orientation of frame *E* is achieved by a rotation, from alignment with frame *S*, of angle of *φ*, *θ*, and *ψ* around the axis *Sxyz*.

Many quaternions are available to estimate the orientation from accelerometer, gyroscope and magnetometer data. We use the Madgwick technique [51] which fuses accelerometer, gyroscope and magnetometer for estimating quaternion. An arbitrary orientation of frame *S* relative to frame *E* can be achieved through a rotation of angle θ around an axis of *Sxyz* defined in frame *E* shown in Figure 5b where the mutually orthogonal unit vectors *Sx, Sy, Sz* and *Ex, Ey, Ez* define the principle axis of coordinate frames *S* and *E*, respectively.

*Sx*, *Sy* and *Sz* define the components of the unit vector *Sxyz* in the *x*, *y* and *z* axes of the frame *S*. To denote the relative frames of orientations and vectors, *<sup>S</sup> <sup>E</sup>q*ˆ in Equation (1) represents the orientation of frame *E* relative to *S* and *Sxyz* is a vector described in *S* [51].

$$\begin{aligned} \, \_E^S \hat{q} = \begin{bmatrix} q\_1 & q\_2 & q\_3 & q\_4 \end{bmatrix} = \begin{bmatrix} \cos \frac{\theta}{2} & -S\_3 \sin \frac{\theta}{2} & -S\_3 \sin \frac{\theta}{2} & -S\_5 \sin \frac{\theta}{2} \end{bmatrix} \end{aligned} \tag{1}$$

The compensated gyroscope measurement *<sup>S</sup>ω<sup>c</sup>* then is used in place of the gyroscope measurements *<sup>S</sup>ω*, where the magnitude of the angular error in each axis *<sup>S</sup>ω*<sup>∈</sup> is equal to a quaternion derivative of unit length and then the integral gain *ς* directly defines the rate of convergence of the estimated gyroscope bias *<sup>S</sup>ω<sup>b</sup>* expressed as the magnitude of a quaternion derivative [51]. The complete orientation of *<sup>S</sup> <sup>E</sup>q*ˆ*est*,*<sup>t</sup>* is achieved and Figure 6 shows a block diagram representation of the complete orientation filter implemented for an IMU. The details derivation is described in the Supporting Information.

**Figure 6.** The process diagram of the complete orientation filter for an IMU.

We apply the technique shown in Figure 6 to our collected data for body acceleration to the Earth frame with a sampling frequency of 50 Hz, *β* gain of 0.1 and *ς* gain of 0.5. The gravity components are removed and the conversion of the accelerometer from gravitational force *g* to user acceleration of movement (*AMxyz*) m/s<sup>2</sup> is achieved by multiplying 9.81. The three axis data are transformed due to the fact that looking at specific axes is sensitive to the sensor orientation [52]. Figure 7 shows the acceleration due to user movement *AMxyz* = [*amx*, *amy*, *amz*] for both feet of older subject 1.

Figure 8 shows the acceleration of total *ATxyz* and gyroscope *GTxyz* towards *x*, *y* and *z* directions estimated using Equation (2):

$$\left| AT\_{\rm xyz\_i} \right| = \sqrt{am\_{\rm xl}^2 + am\_{\rm yl}^2 + am\_{\rm zl}^2} \text{ and } \left| GT\_{\rm xyz\_i} \right| = \sqrt{g\_{\rm xl}^2 + g\_{\rm yl}^2 + g\_{\rm zl}^2} \tag{2}$$

**Figure 7.** Acceleration due to user movement *AMxyz* after removing gravity component.

**Figure 8.** The total acceleration *ATxyz* and gyroscope *GTxyz*.

2.6.4. Stride, Stance, Swing and Step Events Detection

Human walking can be described and characterized in the context of a gait cycle. A stride is the distance between a point on one foot at the first foot contact and the same point on that foot at the next foot contact. It is the equivalent of a gait cycle made up of two steps. Each stride contains stance and swing relevant phases. Stance and swing phases of a gait cycle consists eight relevant phases shown in Figure 9.

**Figure 9.** Normal human gait phases [53].

The first phase starts when the heel contacts the ground and the waist is in its lowest position during the entire step. There is deceleration of the leg towards the horizontal axis as the velocity moves to zero. The zero velocity remains until the terminal stance phase where the foot is flat on the ground. The next phase is pre-swing where the toe is off the ground and starts forward movement demonstrating initial acceleration towards horizontal axis. The swing phase is when the heel moves off the ground. The acceleration interval corresponds to the change from the heel lift to the swing at the height point at mid-swing phase. Deceleration starts during the terminal swing phase from the highest point to the foot back flat on the ground. There is zero velocity again in the interval corresponding to the change from a flat foot to a heel lift. These different phases of gait cycle presented

in Figure 10 are identifiable from the IMU acceleration signal. The same phenomenon of human limb kinematic with accelerometer signal output during a typical walking cycle has been identified in the literature. Our gait cycle accelerometer signal *ATxyz* (Figure 10) is agreed with the signal pattern in [53,54]. The different phases of the gait cycle (Figure 9) with corresponding accelerometer signal are shown in Figure 10.

**Figure 10.** Eight different phases of a gait cycle from accelerometer data.

From Figure 10, we can observe that at the start and end of each stride, the walker's feet are stationary on the ground. As the IMU sensor is attached to the foot, the stance phase is stationary and swing phase is non-stationary. Due to the walker's movement towards the *x* axis, the acceleration shows its highest value in the swing phase. The output of the accelerometer signal will be different, if the sensor is placed on foot [54,55], waist [56] or different body locations [34,57].

Many algorithms [34] are available for stride event detection from IMU sensors. During human walking, a consistent sequence of motions is performed at each stride that results in a maximum peak value that lies in the mid-swing phase. This mid-swing phase appears when a user pushes off this foot and shortens the limb to clear ground thus releasing the foot from the ground until it again contacts with ground as shown in Figures 9 and 10. A particular threshold value is set to detect these characteristics for detecting stride [35,58]. One disadvantage of these algorithms is that any motion with a similar periodicity of walking will trigger for a false stride event. In addition, difficulty arises in finding the automatic selection of the threshold value which can vary between users, surfaces and shoes [59]. The variation in the peak magnitude gets larger for faster human waking velocities [57] and a window based threshold calculation [58] was used to obtain an acceptable level of accuracy for a larger window size. However increasing the window size may degrade the step detection accuracy during the translation of step mode because the threshold calculated from a larger window may not be able to effectively handle the variation in the recent statistics [57]. Due to peak magnitude variation, the threshold value also varies based on individuals walking style and even differs from left to right leg as shown in Figure 11.

**Figure 11.** Peaks magnitude variation from Figure 8.

The different threshold may result in a different output of detecting steps. Another important point is that when a subject begins walking from a standing state, stops walking for a turnaround or stops, there is poor acceleration and it is crucial to detect the gait cycle in these situations. For this reason the 1st stride is not considered for gait analysis by researchers [60]. We take this in consideration to address this point in this study. As the mid-swing phase in accelerometer data is a good indicator for performing a complete gait cycle, thus for counting the number of strides, the number of mid-swing phase in accelerometer data is analysed as walking strides are equal to the number of mid-swing phases. The highest peak is occurred at the push off phase starting from the terminal stance at the 4th to pre-swing at 5th phases shown in Figure 9 for gyroscope data. We apply threshold based algorithms obtaining low accuracy to detect the stride number for our collected accelerometer and gyroscope dataset. As the peaks at terminal stance phase are more prominent than the mid-swing phase, the threshold based algorithm detects two strides instead of single stride. To avoid this, a novel stride detection technique is proposed based on the local minimal prominence characteristics of strides associated with the time-varying magnitude of acceleration shown in Figure 12. The technique consists of designing a high-pass filter, computing the absolute value, designing a low-pass filter, shifting data to centroid and finding the strides using *findpeaks* [61] function.

**Figure 12.** The proposed step detection technique.

The accelerometer converts acceleration to an electrical signal and in the process, unwanted constant bias in acceleration becomes a linear error called drift. Thus the 2nd order Butterworth digital high-pass filter with a sampling rate *fs* = 50 Hz and cut off frequency *fc* = 1000 Hz is applied to *ATxyz* to remove the corrupted data and the DC component of the acceleration signal. The smoothness is achieved at the price of decreased roll off steepness. The output from the filter is then passed to the zero phase *filtfilt* delay filter. The *filtfilt* corrects for phase distortion introduced by a one-pass filter [62]. The output of *filtfilt* filter then is passed through a low-pass filter with *fc =* 5 Hz to obtain *ALP* which is shifted to centroid using Equation (3).

$$A\_{\mathbb{C}en} = A\_{LP} - mean(A\_{LP}) \tag{3}$$

To find the local minima prominences, *ACen* is passed through a *findpeaks* function which finds local peaks in the data vector and ignores small peaks that occur in the neighbourhood of a larger peak. A local peak is a data sample that is either greater than its two neighbouring samples or is equal to Inf. If a peak is flat, the function returns only the point with the lowest index. The *findpeaks* function detects the stationary periods when the foot touches the ground the point of minimal prominence during walking. The function returns two vectors containing the minimal local peaks *AStrides* and the locations *ATIME* at which the peaks occur. The number of strides is the same as the length of *AStrides* vector. Again, as each stride consists of stance and swing events, thus the initial contact and the transition between pre-swing and initial swing (4th and 5th phases in Figure 9) are detected using steps in Figure 13 to get stance and swing information.

**Figure 13.** Proposed stance and swing detection technique.

A *window* is prepared whose size is the difference between a pairwise consecutive strides from *AStrides*. Each *window* is then passed through *findpeaks* function as there is only one local maximum in each stride located between 4th and 5th phases in Figure 9. A loop from 1 to total detected strides number is used to find the stance and swing event for all strides. The detected *Start* (purple circle), *SS* (cyan triangle) and *End* (black rectangle) information of each stride are shown in Figure 14 for right and legs where the stance phase information is provided by the difference between *Start* and *SS*; and the swing information is the difference between *SS* and *End*.

**Figure 14.** Result of stride, stance and swing event detection using proposed method.

A step is the sequence of events between the contact of one foot and the next contact of the opposite foot. At the beginning of the stance phase, the initial contact of the foot contacts with ground of the one leg. The loading response begins at the initial contact and ends when the toe of opposite leg leaves the ground, midstance then begins and finishes when the center of gravity is over the same foot. The terminal stance begins when the center of gravity is over the supporting foot and ends when the opposite leg contacts the ground. The strides, stance and swing event are detected from right and left legs. The step event is then detected between the heel of two subsequent feet shown in Figure 15.

**Figure 15.** Result of step event detection using proposed method.

#### 2.6.5. Velocity and Distance Estimation

A pedestrian navigation system (PNS) has diverse applications in airports, theatres, underground parking, other indoor and outdoor places and modern cities. In order to estimate PNS in a virtual environment, a number of navigation methods [63] are available to derive pose estimates from electrical measurements of mechanical, inertial, acoustic, magnetic, optical, and radio frequency sensors. There are popular conventional approaches of IMU-based PNS known as Pedestrian Dead-Reckoning (PDR) solution and inertial navigation system (INS) [64]. In PDR, a constant step length is assumed on a relatively smooth surface, often usable in office-like environment and the average step length is integrated along with the orientation estimations obtained from IMU at each detected step to track the position. IMU based foot mounted PNS methods and systems have been proposed in [55,65]. The foot mounted method uses a double integral on horizontal acceleration to estimate distance and a gyroscope or magnetometer to measure the heading distance [64]. The disadvantage of PDR is that it assumed on relative smooth surfaces and a constant step length that needs tuning for individual users. Again, if the walking pattern is different from the predefined step length model, this may adversely affect the distance estimation [64]. The INS is adapted from the aerospace community in which IMU is used for tracking position, velocity and attitude [55,66,67]. Sensor drift is a well-known problem in INS system which depends on the precision of the sensor used and high end inertial sensors are very costly [66]. The IMU sensors have some small errors when estimating the distance and direction and signal noise can further exacerbate this problem described details in [64,66].

In this paper, we consider the walking constrains of a user with an IMU fitted on the both right and left legs. We apply appropriate methods to detect the movement of the leg, changes in position and compute its velocity and travelled distance from the initial location by means of the data collected from the accelerometers. The basic approach lies on the double integral of the accelerometer data where the first applying integration retrieves the current velocity and then the second applying integration computed on the velocity provides the distance travelled. Distance travelled is obtained principally from trapezoidal double integration [68] of the user movement signal on each stride detected in the direction of travel as mentioned in Section 2.6.4. However, there are two main problems for performing a double integration of the acceleration signal, unknown initial condition and drift. The unknown initial condition problem means integration requires a known initial condition. Drift means IMU sensors are subject to errors in acceleration that when integrated in to velocity and distance, leads to drastic integration error. This can be unbound over time if the acceleration signal is integrated without filtering [55,64–66,68]. The integration works properly with known initial conditions. Thus, to calculate the actual displacement, integration errors must be minimized. A method known as zero-velocity update (ZUPT) [64,66,67] is often used to correct for drift and is often used to aid in autonomous inertial pedestrian navigation. ZUPT uses the fact that during human walking time, one foot is always stationary on the ground. When a stationary period of the acceleration is detected the assumption is made that the foot is on the ground and the velocity at that time is set to 0. In this way, the drift is greatly reduced. However, ZUPT assumption implies that the angular rate is 0 as well and consequently if the accelerometer is moving at a constant velocity, the algorithm would misjudge the motion as stationary. ZUPT therefore cannot reduce all errors [50,66]. Based on our experience, an accelerometer is very sensitive to movement and walking is a complex course of acceleration and deceleration. The detection of zero velocity does not fail due to misjudgement, but adjusting the threshold value for motion detection plays an important role in that misjudgement when motion detection is not properly set (discussed and showed in Section 2.6.4).

In addition, this issue may not relevant to this study as the "foot stationary event" is already detected based on local minimal prominence as described in Section 2.6.4. The stationary period remains in the stance phase and the movement period remain in the swing phase shown in Figure 9. As IMU sensors are mounted on each foot, the acceleration is high in the swing phase due to the movement of the leg during walking. The zero-velocity in non-stationary period of stance phase is used in the ZUPT scheme to reduce the drift. The ZUPT based on local minimal prominence to detect the swing phase is shown in Figure 16.

**Figure 16.** Zero-velocity update (ZUPT) from & &*ATxyzi* & &.

Another concern regarding the double integration is that the displacement signal emphasizes the low frequency data more than the acceleration signal, a low-pass filter effect of the integrator. Therefore, the input data are passed through a high-pass filter to remove the direct component of the acceleration signal. Considering these issues, a double integral method shown in Figure 17 is proposed for calculating travelled distance.

**Figure 17.** Proposed method for estimating travelled distance.

In order to obtain the velocity and distance in time series, two stages of integration and two stages of high-pass filtering are applied. A stride window is prepared from *AStrides* and *AMxyz*. The swing and stance windows are brought out from the corresponding stride window using *Start*, *SS* and *End* mentioned in Section 2.6.4. A 1st order Butterworth high-pass filer is designed with *fs* = 50 Hz and *fc* = 1000 Hz. The 1st integral operation *cumtrapz* in Matlab is applied on *SWfi* with respect to time *t* that gives the *Vsi*(*t*) velocity for the 1st swing phase. The ZUPT is applied on stance phase to set the stationary velocity to 0. The non-stationary period of swing velocity and stationary period of stance zero velocity are then combined to obtain *Vi*(*t*) shown in Figure 18.

As the stationary period in stance phase velocity is set to zero, the integral constant from non-stationary period in swing phase exists in *Vi*(*t*). Therefore, it is important to remove the drift caused by integration from *Vi*(*t*). To remove the integral drift [50], the velocity difference between the initial and end of a non-stationary period is estimated. The velocity difference is then divided by the number of samples during this non-stationary period to get the drift rate. The drift rate is multiplied with the corresponding data index to estimate the drift value at that certain point. The drift value is then subtracted from the calculated velocity *Vi*(*t*) to obtain the error free velocity *Vdi*(*t*). *Vdi*(*t*) is then passed through the high-pass filter for the 2nd time and the distance *Di*(*t*) is estimated after

2nd integral operation. *Di*(*t*) consists of the distance towards *x*, *y* and *z* coordinates. Repeat the same procedure for all strides to calculate velocity and distance. Then estimate the travelled distance using Equation (4):

$$TD\_{xyz\_i} = \sqrt{D\_{x\_i}^{\ \ 2} + D\_{y\_i}^{\ \ 2} + D\_{z\_i}^{\ \ \ 2}} \tag{4}$$

Figure 19 shows the estimated distance *Di*(*t*) towards *x*, *y* and *z* and travelled distance *TDxyz*.

**Figure 18.** First integral operation to get velocity *Vi*(*t*).

**Figure 19.** 2nd integral operation to get distance *Di*(*t*).

#### 2.6.6. Selection of Gait Asymmetry Variables

GA in this study is considered as an indicator to show the difference between right and left leg walking which may serve as a diagnostic tool for clinicians. There is no commonly accepted superior guideline, preferred methodology or protocol for GA evaluation. The European GAITRite Network Group has developed Guidelines for Clinical Applications of Gait Analysis [69] to provide guidance to clinicians who implement spatiotemporal gait analysis to the clinic. Two issues are addressed in [69]: (1) Environmental measurement conditions and safety issues describe lighting, noise, visual distraction, clothing, footwear and safety; (2) Measurement procedures describe steady-state gait at different velocities, standardized walking instructions, assistive devices, stride-to-stride variability, gait analysis in association with simultaneous cognitive tasks and description of study population. To evaluate stride-to-stride variability, they recommend the highest possible number of gait cycles from a practical standpoint, with a minimum of three consecutive gait cycles for both left and right sides (i.e., a total of six gait cycles) [69]. Many issues relevant in GA assessment is reported in [69]. However there is no recommended systematic procedure for developing a GA assessment. General measurement of gait variability includes cadence, stride length and gait velocity [70], alone or in combination with other outcome measures such as stride to stride variability assessed by an accelerometer, gyroscope and magnetometer [24,36,39,40]. Researchers presented a set of 31 gait variables in [71] and 16 variables were investigated. Stride-to-stride variability [72] is commonly used to quantify walking consistency which is strongly associated with motor ability [73], mild cognitive impairment [74], dementia [75] and stroke [76]. In our study, a set of 24 commonly reported physical gait variables are initially considered for this analysis from both right and left legs. Figure 20 shows the GA variables.

**Figure 20.** Proposed variability monitoring for GA.

A stride length is the distance between the heel contacts of the same foot and a stride time is the interval between sequential initial heel contacts by the same limb. A step length is the distance and a step time is the interval from one foot strike to the other foot strike. The number of full steps taken within a minute is known as cadence. A stance length is the distance between the heel contact and pre-swing phases and a stance time is the interval of stance length. As the IMU sensor is placed at the foot location, the stance length is stationary. During the rest gait cycle, the foot is off the ground as the limb is swung forward to begin the next stride referred as swing phase. A swing length is the distance between the initial swing and terminal swing phases and a swing time is the interval of swing length. Figure 21 shows the stride, stance and swing information of older subject 1.

**Figure 21.** Stride, stance and swing information of right and left legs.

In the normal gait and the gait in patients after disease or injury, a certain GA level should be considered as normal [77]. Structural GA in human movement in limb length without underlying disease or injury is present in 90% of the population with an average magnitude of 5.2 mm and limb inequality of 20 mm must be present before considering clinically significant [78].

To quantify the temporal and spatial asymmetry of gait pattern, symmetry deviations (unaffected side − affected side, expressed as a fraction of the stride duration) [79], symmetric index (dividing the absolute difference of unaffected and affected by their average) [80], asymmetry ratios (1 − (affected/unaffected)) [81], Robinson symmetry index (2 × ((unaffected − affected)/(unaffected + affected)) × 100) [82], a log-transformed symmetry ratio (|100 × (*ln*(affected/unaffected))|) [83], and symmetry angles (([45◦ − arctan(affected/unaffected) × 100]/90) [84] were used. However quantifying indices have known limitations as described in [77]. A certain amount of GA is usually present in able-bodied individuals [77], but no agreement exists regarding the clinical criteria for quantifying of GA. In visual observation or self-reports of physical function, GA is frequently reported as present or not present which may not satisfy scientific criteria of reliability and validity [85]. Thus, an arbitrary cut-off value of 10% deviation from perfect symmetry was previously used as a criterion of asymmetry in gait assessment [76,82]. This was subsequently criticized due to its non-parameter specific nature [86]. Other previously used criteria to describe the presence or absence of GA included sensitivity and specificity [87], the use of 95% confidence intervals (GA within the limits of a 95% confidence interval (CI) obtained in a healthy population would define able-bodied gait, while GA outside the 95% CI would define pathologic gait) [86], and significant limbs difference [77]. In this study, we present gait variability quantities and validate the results with the ground truth for temporal GA monitoring. However, a parameter-specific criterion with optimal cut-off value that best discriminate GA from normal GA for each individual will be investigated in future study. The stride and step asymmetry information from older subject 1 is presented in Figures 22 and 23.

**Figure 22.** Stride asymmetry information of right and left legs.

**Figure 23.** Step asymmetry estimation of right and left legs.

#### 2.6.7. Statistical Analysis

Our estimated travelled distance and Qualisys results are tested for normality using Shapiro-Wilk [88]. The data is found to be normally distributed and a comparison of means is performed using the t-test for significance assuming equal variance. The estimated period and Qualisys results are found not normally distributed and a comparison of means were performed using the Wilcoxon Signed Ranks Test [89] for significance. The *p* values of <0.05 are considered to be significant for both analysis. Statistics are performed using SPSS Version 24 [90].

#### **3. Experimental Results**

Initial experimental results from older subject 1 (male, age 67, height 1.52 m and weight 68 kg) are presented. We extract automatic GA features based on the data collected from both feet.

#### *3.1. GA Results of Older Subject 1*

Table 1 shows the accuracy of the distance travelled and estimated, detecting stride and step number from both legs.


**Table 1.** Velocity, distance, stride and step information.

\* ActualValue, \*\* EstimatedValue.

$$Accuracy = \left(100 - \left|\frac{ActualValue - EestimatedValue}{ActualValue}\right| \times 100\right)\% \tag{5}$$

The accuracy is estimated using Equation (5). The actual distance travelled is 21.03 m measured using manual tape with 99.3 s walking time. The estimated both legs travelled distances are 20.59 m and 20.47 m. The actual and estimated distances are very close. Normal human walking velocity may vary from 1.5 to 2.5 m/s [91] and the walking velocity for this subject is 0.21 m/s which is slow. The accuracy of stride and step event detection are 100%. Table 2 shows the summery of average gait variability.

We can observe from Table 2 that the mean stride lengths of both legs are the same. Although, the standard deviations are low and the right leg's value is lower indicating that the left stride length has more variation compared to the right stride length. The highest stride length is found at the 15th stride (last stride) on the left leg which is before turning. The mean stride times are close for both legs. Although, the right and left leg stride length, time and velocity difference is low, Figure 24 shows that a little stride asymmetry is noticeable in right and left strides time and distance. The difference of other parameters between the legs is also low. However, it is noted from Figure 25 that step asymmetry is more prominent than stride asymmetry which may result in an inconsistent gait.


**Table 2.** Gait Asymmetry Variability.

Std = Standard Deviation, Var = Variance, Min = Minimum and Max = Maximum.

We validate our results using 10 young subjects (age average 27.55 ± 3.54) by conducting trials using the Qualisys Motion Capture System (Qualisys Medical AB, Gothenburg, Sweden) [92] and our IMU sensor concurrently. Applying our method to the collected data leads to the result in Table 3. The average accuracy of the result is 97.57% with 95% confidence interval 1.327 for the estimated distance and 99.01% with 95% confidence interval 0.266 for the Period.


**Table 3.** Validation our results with Qualisys.

Estimated = Estimated distance using our method; Period = Total time of travelling the distance.

Although the sample size is small, the significance of the test of normality for Qualisys and Estimated Distance are 0.83 and 0.37 using Shapiro-Wilk. The *t-*test shows that there is no difference in means (*p* = 0.094) between Qualisys (*μ*<sup>1</sup> = 7.67, σ<sup>1</sup> = 0.26) and Estimated Distance (*μ*<sup>2</sup> = 7.49, σ<sup>2</sup> = 0.39). There is a strong correlation (*r =* 0.81) present. The Wilcoxon Signed Ranks Test for Qualisys and Estimated Period shows no mean difference (*p* = 0.83).

#### *3.2. GA Results of Young and Older Subjects*

Table 4 shows the gait data from the 10 young subjects. Table 4 shows that the accuracy of estimating the total distance compared with the actual distance is also high for both legs. The detected stride and step number using the proposed method is excellent. For all young subjects, the accuracy of detecting stride number using proposed method is 100%. The accuracy of estimating travelled distance using proposed method is 97.73% for right and 98.82% for left legs.


**Table 4.** Velocity, distance, stride and step results for young subjects.

\* ActualValue, \*\* EstimatedValue.

Table 5 shows the details of both legs asymmetry variables information. The stride lengths of legs are the same for young subjects. The overall difference between legs is low for young subjects. In natural walking, the foot is on the ground for about 60% of the total gait cycle during stance phase

and 40% during swing phase shown in Figure 9. The ratio of stance and swing is found closest to the 60:40% split for average stride, stance and swing information (Table 5) for young subjects.


**Table 5.** Right and left legs asymmetry of young subjects.

Table 6 shows that the accuracy of estimating the total distance compared with the actual distance is high for both legs. For all older subjects the accuracy of detecting stride number using the proposed method is 92.67%. The accuracy of estimating the travelled distance using the proposed method is 88.71% for the right and 89.88% for the left legs. The detected stride and step number using the proposed method is also high. However, comparing to results in young subjects (Table 4), the accuracy is lower for older subjects. This is likely to be due to older people walking slowly resulting in a poorer signal output. Table 7 shows the details of both legs asymmetry variables for older subjects. Overall the stride lengths of both legs are similar. The overall difference between legs is very low.


**Table 6.** Velocity, distance, stride and step results for older subjects.

\* ActualValue, \*\* EstimatedValue.

**Table 7.** Right and left legs asymmetry of older subjects.


We check the data for statistical errors and assessed whether the estimated values are reasonable. Figure 24 shows the boxplot of travelled distance from young and older subjects. It is noted that the observations identified by the boxplots are not especially extreme. The young subjects' travelled distance for 30 strides has a wider range and is significantly different than older ones. On average young subjects travelled distance is 37.77 (95% CI ± 3.57) m and in older ones is 22.50 (95% CI ± 2.34) m. Similarly the legs stride and step variation is low for older ones than young ones. Older ones gait is slow and results in a low variation in walking comparing with young ones. The step length has more variation then stride length. Based on the total travelled distance, stride and step information, it can be seen that young and older subjects are distinguishable.

**Figure 24.** BoxPlot of stride and step asymmetry in distances from right and left legs.

Figure 25 shows a boxplot of total time from young and older subjects with their difference. The total time for performing a total of 30 strides is lower for young subjects than older ones. On average the young subjects travelled time is 51.85 (95% CI ± 3.08) s and older ones is 84.02 (95% CI ± 9.98) s. Young subjects show low leg variation with a lower range than older ones. Based on the total time, stride and step timing information, it can be seen that young subjects and older ones are distinguishable.

The detailed results of the 20 subjects are presented in the Supporting Information.

**Figure 25.** Box Plot of stride and step asymmetry in times from right and left legs.

#### *3.3. Discussion*

In this research we show that in a clinical setting outside of a gait laboratory it is possible to collect information about GA using IMU sensors. From Figure 10, we can see that our gait cycle accelerometer signal *ATxyz* is agreed with the signal pattern in [53,54]. We demonstrate the systematic steps of an automatic gait features extraction method that we deployed. Our research enriches the current literature in GA assessment. It is possible to evaluate walking distance using a multisensor approach. Current methods however rely on the threshold based detection of the spike [35,36,58,60]. Our method uses minimal prominence characteristics for detecting gait phases. The former relies on generating a movement of sufficient magnitude to generate the spike and therefore has limited utility in people with slow gait. Our method therefore has the potential for broader use as it can be used in people with slower gaits such as older adults. We demonstrate that our method can deliver accurate results of stride detection and distance travelled similar to accuracy levels demonstrated by other authors [36,60]. We believe that there are advantages to using the minimal prominence approach as it can be used in a wider population people with different gait patterns.

There are however a number of limitations. The number of subjects is still relatively small (20). There is the potential of a Type 1 error (false positive) in detecting an effect that is not there. IMU calibration is an essential part for distance estimation. Although in our methods we try to minimize errors, as gait features are intrinsically variable from person to person, any such algorithm should involve a degree of calibration and error in the measurements. Individual quirks, heel strike, significant body up-down movement and other factors can affect the results. However this can be considered a proof of concept study that has established our method for extracting automatic GA features. There are several other possible sources of errors [93] that may arise from the use of IMU sensors including errors of repeatability, stability and drift. Although IMU sensors performance has been ramped up dramatically, the errors in measurement are unavoidable, especially for miniature micro-electro-mechanical (MEMS) sensors. Future developments should focus on MEMS sensor error modelling and accommodation to further improve parameter estimation accuracy [94]. Other possible areas of error may arise from frictional noise and the relative movement of clothing and shoes to the sensor. However we compared the output from our sensor to a gold standard Qualisys motion capture system which shows good accuracy making the effect of such errors minimal.

To achieve our goal, data are collected from two sensors placed on the barefoot at the medial aspect of foot over the bony prominence of the first metatarsal. It is noted that the orientation of the sensor has a significant effect on output and placing the sensor in different locations gives a different pattern to the data. The position and orientation of the sensor are crucial as changes in position through human error may give different data patterns which might be difficult to interpret. This highlights the importance of properly fixing the sensor to the optimal location to avoid inaccuracies. The placing of sensors on foot locations requires other generic considerations such as battery life and android device that is BLE enabled to pick up sensor data.

To estimate the orientation of the IMU sensors, we apply the Madgwick technique [51] for our collected data but not the magnetic field parameter. The technique is developed assuming that the acceleration would only measure gravity. In practice, accelerations due to motion will result in an erroneous observed direction of gravity and the distortion will present for only short periods of time. Therefore, the magnitude of the filter gain *β* (Section 2.6.3) is chosen low enough that the divergence caused by the erroneous gravitational observations is reduced to an acceptance level over the period. In future, an investigation of dynamic values of gains *β* and *ς* will be conducted to reduce errors.

A threshold is used for detecting steps [36,60] and different value may result in a different output. It is crucial to detect the 1st and last strides of gait cycle when a person starts and stop walking. Thus, the 1st stride is not considered by researchers [60]. Our proposed method for detecting the stride information is based on the local minimal prominence which starts when the heel contacts the ground resulting in the stationary period (Section 2.6.4) and estimated the total number of strides. We also confirm these results obtained by counting the highest peak in the mid-swing phase as it also

is a good indicator for a complete gait cycle. From each stride, the local minimal prominence which is the transition between pre-swing and initial swing (4th and 5th phases in Figure 10) is detected. We find that that when turning or when stopping there is a poor acceleration signal. As gait of older subjects is much slower, it is crucial to detect strides, stance and swing phases from the gait cycle. However, the stationary stance phase is prominent for both young and older subjects. For this reason, we use the local minimal prominence characteristics to detect different events to avoid these crucial phases. We have shown that it is possible to detect, stride, stance and swing events but further analysis of the eight events including single and double support phases in a gait cycle is necessary to provide more accurate information for GA analysis.

In order to track the position in a virtual environment, several navigation methods [63] are available to derive pose estimates from electrical measurements of mechanical, inertial, acoustic, magnetic, optical, and radio frequency sensors. Each approach has advantages and limitations including modality-specific limitations related to the physical medium, measurement-specific limitations imposed by the devices, associated signal-processing electronics, and circumstantial limitations that arise in a specific application [95]. Our velocity and distance estimation is based on results of a double integral with ZUPT. We apply the high pass filter on acceleration data that removes linear trend from the signal and then remove drift to estimate distance. We use the simplest technique of trapezoidal rule for estimating distance for our collected data and our estimated distance results are close to the actual distance. There are many other types of numerical integration schemes available which are much more involved and with the potential for more accuracy. However, the trapezoidal rule is the simplest technique of an entire class of numerical integration schemes which are known as the Newton-Cotes formulas [96] and which we have adopted. Our future plan is to investigate other methods with our collected data.

The results show that our method is capable of extracting automatic gait features and has the potential to be used in GA assessment and gait change monitoring for home and clinical use. Gait with slow velocity is common in older adults [97] and an automatic system sensitive enough to detect gait features in these circumstances is required. Our low cost portable personalized proposed solution could bring out automatic GA features for monitoring longitudinal gait changes or abnormalities. In future work, we plan to use our automatic extracted GA features information to classify gait changes over time to identify abnormal gait patterns for the assessment of elderly fall risk, rehabitation and sports applications.

#### **4. Conclusions**

In the present work, two IMU sensors were placed at right and left metatarsal barefoot locations to collect accelerometer and gyroscope data. We designed and developed an android app to collect real time synchronous data from both sensors. We proposed a systematic method to extract automatic gait features for the GA assessment. We first applied the quaternion technique to raw data for estimating actual sensor orientation. We applied our proposed stride, stance, swing and step event detection technique and analysed for stride, step, cadence, step ratio, stance, and swing. We then estimated distance using double integration with drift removing from acceleration and analyzed for total velocity, distance and time. Our method was validated with the Qualisys motion capture system. We applied our method for 10 young and 10 older subjects. Our results show that it is possible to extract GA features automatically in a clinical setting outside of a gait laboratory. This has the potential to make the evaluation of GA widely available in clinical practice rather than being limited to gait laboratories.

**Supplementary Materials:** The supporting information is available online at http://www.mdpi.com/1424-8220/ 18/2/676/s1.

**Acknowledgments:** This work was supported by European Commission funding (ERASMUS MUNDUS FUSION project). The authors would like to thank all participants that participated in the study. The authors would like to thank Andrew Callaway for his help to collect data from Qualisys and all participants that participated in the study.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **An Affordable Insole-Sensor-Based Trans-Femoral Prosthesis for Normal Gait**

### **Srinivas Pandit 1, Anoop Kant Godiyal 1, Amit Kumar Vimal 1, Upinderpal Singh 2, Deepak Joshi 1,3,\* and Dinesh Kalyanasundaram 1,3**


Received: 18 January 2018; Accepted: 19 February 2018; Published: 27 February 2018

**Abstract:** This paper proposes a novel and an affordable lower limb prosthesis to enable normal gait kinematics for trans-femoral amputees. The paper details the design of a passive prosthesis with magneto-rheological (MR) damping system and electronic control. A new control approach based on plantar insole feedback was employed here. Strategically placed sensors on the plantar insole provide required information about gait cycle to a finite state controller for suitable action. A proportional integral (PI) based current controller controls the required current for necessary damping during gait. The prosthesis was designed and developed locally in India keeping in view the cost, functionality, socio-economic, and aesthetic requirements. The prototype was experimentally tested on a trans-femoral amputee and the results are presented in this work. The implementation of the proposed design and control scheme in the prototype successfully realizes the notion that normal gait kinematics can be achieved at a low cost comparable to passive prostheses. The incurring cost and power expenditure of the proposed prosthesis are evaluated against passive and active prostheses, respectively. The commercial implications for the prosthesis were explored on the basis of recommendations of ISPO Consensus Conference on Appropriate Prosthetic Technology in Developing Countries. The key objective of this work is to enable lucid design for development of an affordable prosthesis in a low-resource setting.

**Keywords:** lower limb prosthesis; trans-femoral amputee; MR damper; knee damping control

#### **1. Introduction**

About 30 million amputees are residing in low income countries [1,2]. Among those, only 5% to 15% are capable to receive a much-needed prosthesis [3]. Although the popular state-of-the art prostheses used in developed countries offer several advanced features, its application is hindered in low income developing countries primarily due to high cost, different functional demands, cultural issues, and unavailability of components locally [4,5].

Gait parameters of trans-femoral amputees including step length, inter-leg symmetry, hip exertion, and involvement of upper body are highly affected [6]. Over the years, there have been considerable efforts to develop affordable prostheses for trans-femoral amputees in developing countries [6–9]. These efforts have resulted in a variety of lower limb prosthetic designs at reduced cost. However, these prosthetic designs include low-end technologies like manual locking knee joints, the weight activated braking mechanism, and polycentric knee joint mechanisms [2,6,10–12]. In manual-locking-knee joints, knee stability is improved by providing more stiffness which renders lower speed and increased

energy expenditure during gait [10]. Problems associated with stiff-legged gait were overcome by incorporating weight activated brakes with a knee joint. A typical weight activated arrangement suffers from limitations of abnormal gait and delayed initiation of swing phase [10]. Polycentric knee joint mechanisms were quite useful in developed countries and now, with some modifications, are being developed to suit the needs of low-income countries. The polycentric knee joint linkages can be configured to improve the stability in the stance phase with fair swing phase initiation. It does impose a limitation. As the extent of stability is increased in the stance phase, the swing phase initiation gets delayed proportionally [8,13]. In this regard, some recent prostheses such as Le Torneau polycentric knee (Limbs International, El Paso, TX, USA), Stanford-Jaipur leg (Stanford University, Stanford, CA, USA), Remotion Knee (D-Rev), SASPL knee joint LCKnee (Andrysek et al.) and a three axes knee (Arelekatti et al.) at MIT [2,10–12] have been developed.

Le Torneau knee and Stanford-Jaipur knee joints have a similar polycentric design of four-bar linkage with little to offer other than stance phase stability. On the other hand, Remotion Knee is the advanced version of Stanford-Jaipur knee with curvature mimicking to address late stance phase initiation [2,10]. Studies have revealed that such prosthetic designs though provide a low-cost solution to a grieving population, it does not take into account the need for an early stance flexion-extension and properly timed late stance flexion [11,12]. The basic biomechanical aspects, when overlooked, can result in many psychological, physiological, and socio-economic problems [14]. Abnormal and sub-normative gait puts immediate and extended costs in terms of psychological stigma and socio-economic distress, whereas in the longer term it results in musculoskeletal impairments and adversely affects the physiological equilibrium [15–18]. It suggests that the excessive focus on cost reduction alone cannot serve the desired purpose and low-cost prostheses must also have the ability to provide a gait kinematics closer to that of a healthy person.

Also, Andrysek et al. and Arelekatti et al. addressed the problem of abnormal gait in low-cost passive prostheses. Andrysek et al. SASPL engages or disengages depending on the loading of the prosthetic limb during weight bearing but is not able to provide early stance flexion-extension and optimal swing phase damping [11]. Arelekatti et al. incorporated able-bodied gait kinematics using an automatic early stance lock for stability, a linear spring for early stance flexion-extension and swing control. The design suffers from few limitations such as the necessity of full knee extension at the end of swing phase and failure to lock the knee before stance phase that can lead to instability and risk of fall. Further, due to the lack of any clinical gait analysis, the study does not establish the extent of improvement in amputee kinematics towards normative gait [12]. It offers sufficient motivation to look for a new prosthesis design and control scheme that can enable able-bodied gait kinematics in trans-femoral amputees at an affordable cost in developing countries. So far, all developed prostheses designed for developing countries are primarily passive in nature, the authors believe that a variable damping can provide the subject with comfortable walking, thereby yielding the desired results at a low cost.

In this paper, a novel low cost lower limb prosthesis was developed and its clinical testing and analysis are presented. A new foot plantar insole feedback-based control strategy was developed and implemented to provide necessary damping using an MR damper. The prosthesis enables the amputee to realize the able-bodied normal gait kinematics and meets the functional, socio-economic, and aesthetic needs of a trans-femoral amputee and is affordable even in low-income economies. The paper also focuses on essential design features and requirements of a prosthesis to assess its commercial viability in these countries such as low cost, light-weight, functionality, biomechanically appropriate, durable, using locally available materials etc., based on the recommendations of ISPO Consensus Conference on Appropriate Prosthetic Technology in Developing Countries and elsewhere [19–22]. The details are discussed later in Section 5.7. The key objective of this work is to enable lucid design for development of an affordable prosthesis in low-income economies.

#### **2. Prosthesis Design**

This rationale of the prosthesis design is presented in this section. The trans-femoral prosthesis shall serve the major function of providing stability during stance phase and controlled flex-extension during the swing phase of a gait cycle. In the stance phase, the stability was ensured by providing suitable resistance to knee joint flexion while the user transfers body weight from the sound/healthy limb to the prosthetic limb. Similarly, the swing phase flex-extension was controlled by regulating walking speed [22]. Both the functions were achieved by including springs and dampers in the design [3]. Either a variable damping system or a system with multiple dampers is required for a variable damping system incorporated in state of art active prostheses, whereas multiple dampers are used in recent advanced passive prostheses. The multiple-damper approach inherently puts the design at disadvantage in terms of serviceability [10,11].

A study conducted by Herr and Wilkenfeld [23] serves as the theoretical background for our design to realize able-bodied gait kinematics. The authors used an MR knee prosthesis to adapt knee damping using local sensing of knee force, torque, and position. A few recent studies conducted by Xu et al. [24], Park et al. [25], and Fu et al. [26] have also attempted to design a lower limb prosthesis based on an MR damper. Complex sensing, data-driven AI control and lack of local climatic adaptations as well as increased cost of the prosthesis arising from various factors ranging from non-inclusion of off-the-shelf components or locally manufactured components increased the inherent cost, making it unsuitable for applications in low-income economies. Further, the control architecture based on stiff tracking of the knee angle should be avoided in order to enable the amputee to use the prosthesis more interactively, rather than reacting to it [24]. Therefore, the potential of prostheses based on MR damper remain untapped in such low-income economies due to the unavailability, complexity, and incurring cost of design and manufacturing [20,23,27].

It is proposed that a single damper with a controllable damping capacity, such as an MR damper, if used with suitably designed sensing modality, driving circuitry, and control architecture, can provide able-bodied gait kinematics yet at a cost below par with advanced passive prostheses. Here, strategically placed sensors on the plantar insole provide sufficient information for normal gait kinematics, thereby enabling the control of the current in the MR damper during the different phases of a gait cycle. This forms the basis for the design of a simple MR damper-based knee joint mechanism and its control with plantar insole feedback.

#### *2.1. Mechanical Design*

The design of the prosthesis' mechanical structure is shown in Figure 1. The design includes an MR damper, a hinge-type knee joint, and braces to support the leg assembly. In the MR damper, modulation of magnetic field strength can offer variable resistance to the flow of MR fluid. Placement of electromagnets around the piston rod and MR fluid as a working medium inside a hydraulic cylinder can result in a variable damping [20,23,27]. Damping can be modulated by changing the current flowing in the coils of the electromagnet. A variety of MR dampers are available to suit lower limb prostheses. In this study, a commercially available MR damper (Make: Lord Corporation, Cary, NC, USA; Model RD-8041 with a stroke length of 74 mm, and extended length of 248 mm) was employed in the prototype.

The piston rod and outer cylinder of the MR damper were fixed to the knee joint and the shank individually with swing link and lower-link rods. The upper-link and lower-link rods are attached to the braces. The upper link rod and swing-link rod form the knee axis and the control axis of the prosthesis, respectively; the upper link rod along with knee body articulates the knee braces via the knee axis. The control axis primarily responds to the ground reaction forces that cause relative motion between the knee and the braces. The swing-link rod transfers the forces generated by the MR damper during locomotion. The biomechanical mechanism of the control axis has been extensively reported elsewhere [26]. The control of the prosthesis in both the stance and swing phases was achieved by the same mechanical assembly and no separate mechanical component was used to lock the knee. The upper part of the knee joint (knee body) was connected to a socket connector to fasten the socket properly. The braces along with MR damper form the shank of the leg. It was connected to a Sach foot with a shank connector. The socket and shank connectors are connected to the knee and ankle adapters used by prosthetists for proper alignment of the prosthesis to the individual amputee. Knee Movement has been restricted to 90 degrees with protective bars (stoppers) placed in the forward direction of the damper. Custom designed components were used for link rods, plates, braces, and knee joint to meet the design specifications. Proper motion of the knee joint was ensured by simulating the solid model design in Solid Works®. During the simulation along the entire knee range, there was no interference between the joint links, bars, plates, and braces. With proper machining of all the custom components, the mechanism was assembled by bolting the braces to the knee joint. The MR damper was shoulder-bolted with the links of the assembly. Bearings were employed for smooth motion of joint between knee body plates and link rods. Off-the-shelf components such as connectors, adaptors, and the Sach foot were employed to extend the design with existing prosthetic components.

**Figure 1.** Three-dimensional structure and mechanical design of the proposed prosthesis.

#### *2.2. Sensing Mechanism*

Plantar insoles are extensively used for gait analysis of many locomotion impairments [28,29]. Although plantar insoles provide information during only the stance phase, it is sufficient for most of the rehabilitation support systems, posture and balance control [28]. Most of the plantar insoles employ a discrete number of sensors to obtain the required information; the minimum number of sensors vary from 2 to 16, or more depending upon the need [28]. Taking into account the advantages of foot plantar insoles, a novel sensing method was developed to control the prosthesis both in the stance and swing phases. It is composed of 24 switches strategically placed on the heel (S1), mid-foot (S2), metatarsal (S3), and toe (S4) of the plantar insole as shown in Figure 2. The sensor placement was determined by considering dominant pressure points of the plantar surface during walking as reported by Razak et al. [28]. The rationale for considering sensors on plantar insole lies in the wheel-like-behavior of plantar foot and its subsequently resulting roll-over shapes (ROS) [30,31]. Such ROSs are formed for various biomechanical events corresponding to different phases of a gait cycle of human locomotion i.e., loading response, mid-stance, terminal stance, pre-swing, and swing [32]. The progression of a typical gait cycle of a healthy individual along with subsequent phases and events are depicted in Figure 3. As the five segments are sequential in nature, the phases are distinctly noticed

due to the formation of ROSs based on the combination of states of four groups or sensors placed on the foot plantar insole. The grouping of the 24 sensors into 4 groups is shown in Figure 2b. The stance phase events can be related to different combinations of the states of four sensor groups as tabulated in Table 1. If the state of any of the sensors in a group is high, then the state of the same group is assigned the value of 1.

**Figure 2.** (**a**) Photograph of plantar insole and (**b**) sensor grouping.

The circuitry and mechanical dimensions of the insole were designed using Coral Draw®. The insole was printed using the silk screen printing process (Keywell Industrial Co., Noida, UP, India). In this process, silver-conducting ink and dielectric was used to print the circuit. While designing the insole, it was important to determine the minimum thickness of the dielectric layer and the amount of pressure that the insole may undergo during walking. It enables the designer to use an optimum thickness of the dielectric layer that is sufficient to trigger the events. For the initial prototype, this trade-off between minimum dielectric layer thickness and maximum loading was made using the ground reaction force (GRF) profile of the amputee on a force plate.

**Figure 3.** Typical gait cycle of a healthy person [33] and its classification [30].


**Table 1.** Sensor states during different phases of gait.

#### *2.3. Embedded System*

To feed the desired current to the MR damper at the given time instant, an electronic circuit that uses sensor readings was designed to regulate the current. The MR damper requires variable current to provide variable damping; the MR damper current can be varied by varying the input DC voltage across the load. It was achieved by a step-down DC/DC converter, controlled by a pulse width modulation (PWM) scheme. The duty cycle of the PWM signal was varied to reduce the error in the reference and actual feedback current. Error in reference and feedback current can be reduced to a permissible extent in desirable time by employing a proportional integral (PI) controller. The actual feedback current was measured with a Hall current sensor ACS712. The power circuit includes a battery bank rated at 9 V and 2.4 Ah, a power MOSFET, freewheeling diode, low pass LC filter, and MR damper.

The power and control circuits were isolated with an optocoupler and gate-driving circuit to ensure smooth operation. The control circuit includes the sensing mechanism along with a control unit as shown in Figure 4. A microcontroller is the central component of the control unit, that in-turn is the brain of the total signal processing system. The primary objective of the microcontroller is to control the power driving unit when required to change the damping. After careful analysis of the requirements, the Atmel mega (ATmega) 329P microcontroller was chosen for this work for its low cost, versatile characteristics, and more advanced features. The necessary processing of sensor signals and implementation of control algorithm was achieved with the firmware. The details of the components of the embedded system are given in Table 2.

**Figure 4.** Embedded system layout.


**Table 2.** Details of the embedded system components.

#### *2.4. Control Architecture*

A two-level control scheme was employed to achieve able-bodied gait kinematics as depicted in Figure 5. The control approach involves a finite state controller as a secondary controller and a conventional PI controller as a primary controller. The secondary controller generates the required current references for the MR damper using a finite state machine that ultimately modulates the impedance of the gait, depending on the gait event segment. The primary controller is a closed loop PI controller for MR damper current, which compensates for the load transfer dynamics, thus enabling the faithful tracking of current references with a higher bandwidth and accuracy as compared to its open loop counterparts. Prosthetic control is different from robotic control where the oscillatory response was avoided and the system was tasked to follow a stiff trajectory. PD, P, PID, sliding mode, and hysteresis controllers are inherently suitable for stiff tracking, whereas a PI controller offers a more suitable response. Furthermore, a PI controller is the most widely used current controller in industrial applications, and its incorporation would attract more commercial acceptability.

**Figure 5.** Block diagram of the control scheme.

This ordered behavior of state machines are utilized in neuro-prostheses with event triggered control [34,35]. The states of the finite state machine were determined by considering various demanded requirements [36]; here we intend to achieve the flexion-extension of the knee during the stance phase with a high degree of stability and a suitably damped swing phase to adjust the walking speed. In order to meet the stated requirements, a gait cycle with five segments/phases as described earlier were considered i.e., loading response, mid-stance, terminal stance, pre-swing, and swing. The above states are illustrated in Figure 6. Though there could be any number of states that can be used to form a state machine, as previous studies have reported, the introduction of a foot plantar insole with four sensor groups provides a total of 16 states, out of which 15 belong to stance phase and one to swing phase. Due to the ROS trajectory during level walking, only 7 states of stance can be assigned to four phases as tabulated in Table 1, whereas the remaining 8 states of stance do not represent any of the ROS segments. The available literature suggests that consideration of 4 states have successfully resulted in desired kinematics in the stance phase [27], whereas treatment of swing phase as a single state does not deviate much from the same [37].

In state 0, the knee flexes near to the maximum stance flexion. A relatively high damping was applied during this state to prevent buckling at the knee due to the user's weight. During state 1, the knee begins to extend after maximum flexion. The rate of extension was lower than that of flexion during state 0, thereby requiring a high damping. The level of damping shall account for slow extension in order to avoid potential slamming of rotating parts to the knee housing and stoppers. State 2 involves the flexion-extension while the heel is off the ground and the sound limb is sharing the body weight, a moderate damping shall serve the purpose here. State 3 encounters the extension with a high rate of change and low body weight bearing; a low damping will provide the desired profile. The final state (state 4) represents a large knee flexion and associated extension after flexing to its maximum. The rate of change of flexion-extension suggests a very small damping for the same. All states can be tuned to suitable current references which can be accurately followed by a well-designed PI controller. The response of the PI controller must be fast and accurate as it lies in the inner loop of the control architecture.

**Figure 6.** Finite state machine diagram.

#### **3. Methods**

The developed prosthesis was tested on a 24-year-old male (1.63 m, 60 kg) unilateral trans-femoral amputee in India as part of the target population. The subject was using a commercially available passive prosthesis (single axis extension assisting prosthetic knee joint) for the last three years post traumatic amputation. The amputee mobility predictor (AMP) score of the amputee was more than 35, i.e., above the K3 activity level. Table 3 shows the measurement of amputee and component used in the proposed prosthesis. A custom made Ischial containment socket was designed for the amputee by a licensed professional prosthetist. The prosthesis was assembled according to the measurements of the amputee.

**Table 3.** Details of the amputation and prosthesis.


The dynamic alignment of the prosthesis was performed by the trained prosthetist who is one of the co-authors of the current work. The participant did not have any musculoskeletal injuries, sensory or neurological impairment or related disabilities other than trans-femoral amputation. The study was approved by the All India Institute of Medical Sciences ethics committee (Ref. No. IEC-35/09.02.17). The participant was explained the experiment protocols before starting the experiment and written consent was obtained.

#### *3.1. Parameter Tuning and Training*

The controller parameters were tuned sequentially considering one controller at a time. Firstly, the innermost current controller was designed for the electrical time constant of the MR damper, so that a maximum current of 1A could be attained with a desired speed and accuracy. The gain parameters were tuned separately, and the speed was optimized after attaining a fair accuracy. It was achieved by the Ziegler–Nicholas method of tuning the PI controller. After tuning of the PI controller, i.e., the current controller, the prosthesis was fitted on the subject for familiarization. When the subject was able to walk properly with certain level of comfort, the prosthesis was then considered ready to be tuned for the secondary level of control. As stated in Section 2.2, a gait cycle was segmented into 5 states (shown in Figure 7), and for each state one current reference was to be set, thereby a total of 5 parameters of the finite state controller needed to be tuned. Figure 7e,f depict the same damping as shown in Figure 8, however, for the sake of visualization, they are shown separately. It was usually straight-forward to make an initial guess of only 5 parameters. However, to avoid unnecessary discomfort to the subject, the parameters were tuned sequentially in each state depending upon the feedback from the subject and the prosthetist.

**Figure 7.** Photograph of subject with developed prosthesis during training in segment (**a**) loading response, (**b**) mid-stance, (**c**) terminal-stance, (**d**) pre-swing, (**e**) extended knee in swing, and (**f**) fully flexed knee in swing.

After obtaining an initial set of parameters, all the values were iteratively fine-tuned to attain a gait characteristic closer to that of a healthy individual. The knee angle was used as an essential quantitative measure to observe the performance. Subject feedback along with that of the trained prosthetist was considered primary qualitative criteria to tune the parameters. This type of tuning avoids unnecessary discomfort to the subject. The final set of current references are depicted in Figure 8.

**Figure 8.** Final set of current references after tuning.

Once the prosthesis was tuned for level walking, the subject was given sufficient training time to get proper acclimatization to it. A number of training sessions were conducted for the subject in accordance with the training methods reported in several studies [38–41]. The training was usually specific to each subject, thus the sessions were planned by a trained prosthetist keeping in view that the subject had been using a passive prosthesis for a long duration. When the subject actively started to walk on level ground, the experimental data was collected for analysis.

#### *3.2. Experimental Setup*

The subject was asked to walk on a 24.38 m (80 feet)-long corridor at a uniform subject comfortable speed. The initial walk of 3.05 m (10 feet) and 3.05 m of walk towards the end were excluded to avoid any deviation in biomechanical data that may arise due to acceleration and deceleration during beginning and ending, respectively. Hence the walking variables were measured over an active length of corridor of 18.29 m (60 feet) which was assumed to be uniform.

For the subject (fitted with the prosthesis as shown in Figure 9), a total of 10 trials were conducted; one trial included initial standing of 10 s and then level walking of 24.38 m (80 feet) followed by standing for 10 s. The starting, stopping, and other necessary instructions were provided verbally to the subject. After each trial, the subject was asked to take a minimum of 2 min rest which could be increased based on subject's requirement. In order to evaluate the performance of the developed prosthesis, the data obtained from the knee goniometer, foot plantar insole, on-board clock, and the voltage and current sensors were transmitted wirelessly with a Bluetooth module to a remote desktop.

A freeware serial port terminal application (CoolTerm, version 1.4.7) was used to record the data with a sampling frequency of 100 Hz. The data was filtered in MATLAB® R2015b using a low pass Butterworth filter of 4th order with a cut-off frequency of 10 Hz. The cut-off frequency of 10 Hz was preferred to filter the data as the frequency of human motion was typically less than 5 Hz. The signal was further smoothened using a 5-point moving average filter.

**Figure 9.** Photograph of experimental setup and subject during testing of prosthesis.

#### **4. Results**

#### *4.1. Gait Kinematics*

Knee joint angle is the most established parameter to evaluate the kinematic performance of a prosthesis [37]. The measured knee angle from the on-board goniometer during level walk is shown in Figure 10. The prosthesis was able to demonstrate an early stance phase knee flexion-extension; at a self-selected walking speed of a gait cycle, the knee joint in early stance flexed up to 15 ± 5 degrees. The swing phase flexion was constrained up to 45 ± 7 degrees, which is below the biologically acceptable limit of 70 degrees. Combining both, the prosthesis realizes a knee angle trajectory that is closer to that of able-bodied or normal gait kinematics.

**Figure 10.** Gait kinematics measured during experiment.

#### *4.2. Power Consumption*

External power requirement is the major concern for both active or semi-active prostheses. To evaluate the performance of the prosthesis, the assessment of power consumption was carried out by measuring the average current through the battery. In this study, the average electrical power provided by battery bank and the consumption by MR dampers were measured by Hall Effect current sensors (ACS712). The input electrical power consumed in a gait cycle can be evaluated from Figure 11

by considering supply voltage of 9 V. In-line with the expectation, the peak power was perceived during mid-stance as the prosthesis supports entire body weight at this time. The power consumption during loading response was less than the peak power due to double support. In terminal stance, the body transfers the load to the sound limb. Hence, a moderate power was consumed during this phase. During the swing and pre-swing phases, the prosthesis consumes very little or almost no power due to very little load of prosthesis itself. An average power of 2.25 W was consumed by the MR damper during a typical gait cycle of over-ground level walking.

**Figure 11.** Actual current flowing in the MR damper during a typical gait cycle.

#### **5. Discussion**

#### *5.1. Gait Biomechanics*

The response of developed prosthesis is seen in Figure 10. It can be noticed that the gait kinematics are more similar to that of a healthy person performing a level walk. On comparison of measured knee angle to a typical gait data [29] (as shown in Figure 3), the response is quite similar to that of normal gait. Contrary to its passive counterparts, it enables the amputee to achieve knee flexion-extension during heel strike and mid stance with the required degree of stability. Such an extent of stance flexion and extension is absent in the typical gait cycle of an amputee using any passive prosthesis [38–40]. Controlled swing phase knee flexion-extension is also achieved; thus meeting the two necessary requirements of a prosthesis to provide able-bodied gait kinematics [35]. Though the kinematic data of healthy individuals and amputees can be compared, one should avoid trying to find a precise match between the two. The healthy gait data serves as a reference for essential characteristics only. The results presented in this study can best compared with that in the study conducted by Herr and Wilkenfeld on four unilateral transfemoral amputees using MR damper knee joints [23]. The knee joint angle of the proposed prosthesis exhibits all the characteristics similar to that of the knee angle trajectory obtained in [23].

The lack of smooth stance-to-swing transition is a major problem witnessed in many of the previous prostheses [11,37]. This problem is usually associated with hyper-stabilizing stance phase knee joints, such as polycentric knee joints, causing delayed initiation of swing phase. The developed prosthesis, due to its single axis design and control strategy, effectively overcomes the stance to swing transition problems. The design and control of the prosthesis also eliminates the necessity of full knee extension at the end of the swing phase as observed by Arelekatti et al. [16].

#### *5.2. Control with Plantar Insole*

The control of a semi-active or active prosthesis is usually achieved by highly sophisticated methods including complex sensing, data driven advanced AI, and state of the art embedded systems [23,26]. All these methods require computationally efficient software and hardware platforms associated with delicate peripherals and sophisticated electronic circuits, thereby making it a costly affair for low-income economies despite its huge potential to realize normal gait kinematics. The novel control method introduced in this study provides a low cost, effective, and efficient alternative to existing prostheses. The simple form of digital information requires almost no processing and does away with the unnecessary delay caused by Analog to Digital Convertors. Hence, it enables the designer to select a simple, low cost and low power microcontroller, a major component contributing to higher cost.

The study establishes that a foot plantar insole possesses significant information required to control the prosthesis. The basic principle of extraction of information from a plantar insole lies in ROS created by the wheel-like mechanism of the foot during locomotion, as also reported earlier [30,31]. Different ROSs are caused by different kinematic patterns. However, the kinematic data is also subject dependent but has well defined characteristic trajectory. The patterns usually vary with age, type of locomotion, and kind of locomotion impairment. This gives the opportunity to extend to other fields, such as orthoses for lower limbs. The electronically controlled prostheses are either joint angle controlled or torque controlled. The angle control approach is inherently not suitable for prostheses due to biomechanical reasons, as reported by Sup et al. [37]. On the other hand, the torque control or tracking with state impedance control schemes, have already established its suitability in active prostheses [42], but its application in semi-active prostheses [43] may not be cautious enough as semi-active prostheses reported in [24–26] using an MR damper do not contribute to net propulsive power during walking. Its function still remains as a passive device with impedance or damping modulation despite torque control. Hence, torque as a kinetic variable representing work and energy interaction is not suitable and is an over-applied idea for semi-active prostheses. Damping modulation can be achieved by simpler yet efficient control methods based on variables representing sufficient information of a gait cycle.

#### *5.3. Cost and Weight*

The developed prosthesis was designed in-house, fabricated, and assembled in the lab in India. Other than the MR damper, all the components were customized and locally manufactured with locally available materials. In an initial attempt to test the feasibility of the control scheme in less time, the fabrication of the MR damper was avoided and the most widely used MR damper (Lord Corporation Model RD-8041) was imported. The total cost of developed prosthesis is 17,000 INR excluding the damper; a suitably customized, locally designed, and manufactured MR damper is estimated to incur 5000 INR more, amounting the total cost to 22,000 INR. This cost is comparable to commonly available low cost passive prostheses which typically cost ranging from 6000 to 150,000 INR [13]. The total weight of developed prosthesis is 1.77 kilograms (kg) excluding damper and is expected to add 0.5 kg after attachment of customized locally manufactured damper. A total weight of 2.27 kg is quite lower than that of reported active and semi-active prostheses [12,23]. The component-wise manufacturing/production cost and weight is summarized in Table 4.


**Table 4.** Manufacturing cost and weight of various components.

<sup>a</sup> Excluding MR damper.

#### *5.4. Battery Life and Power Consumption*

While importing external power to the prosthesis, the power consumption assessment is vital to its commercial acceptance. The average power consumption during a gait cycle was measured to be less than 2.5 W during over-ground level walking, whereas that of a typical active prostheses is in the range of 60–100 W for the same [37,42]. Hence, (i) the power consumption of the developed prosthesis is almost 30 times less than its active commercial counterparts, and (ii) a reduction in power requirement can enable further reduction in weight and cost. For a battery bank of 9 V and 2.4 Ahr, the battery can deliver up to 8 h of continuous supply to the developed prosthesis on a level walk. With this power arrangement, considering stride time of 2 s and stride length of 70 cm, the amputee will still be able walk 14,400 steps or a distance of around 10 km. As suggested by references, an adult walks 4000 to 18,000 steps per day [44,45]. Hence, the prosthesis is capable of operating an entire day with one time charging in a day. This is a significant advantage over active prostheses that requires ~115 Wh (Equivalent to a 9 V, 12 Ahr battery) to work all day [37]. In low-income/developing countries, where access to usual electric charging may not be available at all times, a customized provision of charging with renewable sources such as PV panel, wind etc. can be made at a minimal additional cost.

#### *5.5. Prosthesis Utilization*

The level of utilization is one of the parameters to evaluate the usability of any system. A system usability score (SUS) [46] has been calculated based on the questionnaire response of the subject after using the developed prosthesis. The same questionnaire has been used in the past on several rehabilitation studies [47]. The SUS score of 85 shows a promising scope for high usability of the prosthesis in Indian scenario (Table 5).



<sup>a</sup> The SUS scores range from 1 ("strongly disagree") to 5 ("strongly agree"). The SUS score was calculated based on the responses following Brooke [46].

#### *5.6. Implications for Commercial Applications*

There are a variety of factors that affect the applicability, commercial viability, and social acceptability of a prosthesis in many low-income economies. The prosthetic needs in these countries differ based on functional demand, local availability of materials, cost, and cultural barriers [4]. As suggested, some points of consideration of appropriate technologies for prosthetics in developing countries like India primarily include low cost, local availability, consideration of local climate, suitability for local manufacturing, and durability [5,21,45,48,49]. Similar recommendations were made at the ISPO Consensus Conference on Appropriate Prosthetic Technology in developing Countries [21]. The developed prosthesis complies with most of the recommendations, and these suggested features, along with description, are tabulated in Table 6. The cost of the passive prosthesis ranges from US \$125 to 1875 as per Sam et al. [9].

**Table 6.** Technical features and descriptions based on recommendations of ISPO consensus conference on appropriate prosthetic technology in developing countries [5,20,21].


#### *5.7. Parameter Tuning*

It may appear that there are seven parameters to tune, but usually in a scenario where the prosthesis is manufactured at a facility and the patient is trained by prosthetist, the current controller parameters can be tuned by the manufacturers and the remaining five can be tuned by the prosthetist. It is justified as the prosthetist is trained and experienced at analyzing gait patterns, and therefore it is easier for he/she to judge the suitable value of the parameters. The highlight of the proposed design is the ease of fabrication in low-resource settings, and thereby can help amputees in locomotion.

#### *5.8. Limitations and Future Work*

Though the prosthesis achieves its objective in terms of kinematic performance and other parameters explained earlier, there is ample scope to take it to the next level, as this work is a pilot study. Factors such as heat dissipation, the reliability of the insole sensors, and security of control to prevent falls need further attention. The developed prosthesis design is in its initial stages and further modifications need to be made to bring the design out of the lab setting. A locally manufactured MR damper needs to be fixed to the prosthesis in order to extensively evaluate the performance of the developed prosthesis on a large number of the targeted population as the prosthesis has been tested on single amputee only. A more aesthetic make over with suitable casing and fitting could unveil its humanoid appearance. The present design and control method addresses level walking with a single speed only, a variable speed adaptation would enhance its maneuverability and user friendliness. The future work would include the implementation of the above left-over points with design, development of control scheme for variable speed adaptation, and thorough testing of the modified design on more subjects.

#### **6. Conclusions**

A novel low-cost lower limb prosthesis based on sensors placed in a plantar insole was presented in this work. The hardware with suitable control architecture enabled the trans-femoral amputee to attain able-bodied gait kinematics. The results are encouraging and provide a new dimension to the control philosophy of proposed prostheses. The study presents the initial prototype and hence may need thorough testing and related modifications before commercialization. The developed prosthesis is an optimum choice for lower limb amputees in low-income or low-resource settings as well as in developing countries. The attractive features of low cost, durability, and able-bodied gait performance are expected to expand its commercial reach beyond India to other low and medium income developing countries.

**Acknowledgments:** The authors gratefully acknowledge the active participation of Vipin Kumar as a subject in this study. His patient support during training and testing of prosthesis is highly appreciable. We thank Indian Council of Medical Research (ICMR) New Delhi, Govt. of India for funding this research via research grant No.5/20/13/Bio/2011-NCD-I.

**Author Contributions:** D.J. conceived and suggested the development; S.P., A.K.G., D.J., and A.K.V. contributed in the development and experiment; A.K.V. and U.S. provided training; D.K. contributed in mechanical structure development; D.J., U.S., and D.K. wrote the paper.

**Conflicts of Interest:** The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Inertial Measurement Units for Clinical Movement Analysis: Reliability and Concurrent Validity**

**Mohammad Al-Amri 1,2,\*, Kevin Nicholas 1,2,3, Kate Button 1,2,3, Valerie Sparkes 1,2, Liba Sheeran 1,2 and Jennifer L Davies 1,2**


Received: 15 January 2018; Accepted: 26 February 2018; Published: 28 February 2018

**Abstract:** The aim of this study was to investigate the reliability and concurrent validity of a commercially available Xsens MVN BIOMECH inertial-sensor-based motion capture system during clinically relevant functional activities. A clinician with no prior experience of motion capture technologies and an experienced clinical movement scientist each assessed 26 healthy participants within each of two sessions using a camera-based motion capture system and the MVN BIOMECH system. Participants performed overground walking, squatting, and jumping. Sessions were separated by 4 ± 3 days. Reliability was evaluated using intraclass correlation coefficient and standard error of measurement, and validity was evaluated using the coefficient of multiple correlation and the linear fit method. Day-to-day reliability was generally fair-to-excellent in all three planes for hip, knee, and ankle joint angles in all three tasks. Within-day (between-rater) reliability was fair-to-excellent in all three planes during walking and squatting, and poor-to-high during jumping. Validity was excellent in the sagittal plane for hip, knee, and ankle joint angles in all three tasks and acceptable in frontal and transverse planes in squat and jump activity across joints. Our results suggest that the MVN BIOMECH system can be used by a clinician to quantify lower-limb joint angles in clinically relevant movements.

**Keywords:** inertial measurement units; motion analysis; kinematics; gait; functional activity; repeatability; reliability; biomechanics

#### **1. Introduction**

Assessment of movement patterns during functional activities such as walking and squatting, and during sporting manoeuvres such as jumping, is a cornerstone of musculoskeletal physiotherapy. Commonly, the physical examination involves observation by the clinician and completion of clinicianor patient-rated scales [1]. However, it is challenging to accurately evaluate multiple joints of both legs in multiple planes of movement when an individual is performing a dynamic functional activity, which often occurs at speed. Three-dimensional optoelectronic (camera-based) motion systems can be used to provide comprehensive, objective measurements [2], but this typically requires the patient to attend a specialised movement analysis laboratory. The equipment within these laboratories is expensive, non-portable, and requires a high level of technical expertise and a lengthy calibration process. The use of these systems is therefore not widespread in clinical practice, and clinicians typically do not have access to objective biomechanical information for assessing patient performance. A more rigorous approach to quantifying joint movement in the clinic is required.

A potential solution to this problem is inertial measurement units (IMUs), which could be used in clinical settings to objectively measure movement patterns during functional activities. An IMU is comprised of accelerometers, gyroscopes, and magnetic sensors combined with a fusion algorithm, for example a Kalman filter [3–5]. It can be attached to a body segment to estimate the movement of that segment in space. When combined with other IMUs on adjacent body segments, the kinematics of movements can be calculated [6–12]. IMU motion-capture systems are portable and less expensive than traditional camera-based motion-capture systems. The validity of joint kinematics calculated with IMU systems has been confirmed with respect to optoelectronic systems [10,11,13].

Commercial IMU systems are increasingly available, from companies such as Xsens Technologies B.V., (Enschede, The Netherlands), Shimmer Sensing (Dublin, Ireland), BioSyn Systems (Surrey, BC, Canada), I Measure U (Auckland, New Zealand), and APDM Wearable Technologies (Portland, OR, USA). The Xsens MVN BIOMECH system is composed of wireless motion IMUs (called MTw2 sensors) and native biomechanical protocols, and estimates three-dimensional joint kinematics [14]. The accuracy of the orientation information provided by an individual IMU [15–17] and the joint angles provided by MVN BIOMECH software [18,19] have been reported. Robert-Lachaine et al. [18] and Zhang et al. [19] compared joint angles obtained from MVN BIOMECH software against those obtained from a biomechanical model associated with the Optotrak camera-based motion analysis system. During walking and stair climbing, the mean differences in lower-limb joint angles between the two systems were between 1.4◦ and 6.7◦, and the coefficient of multiple correlation (CMC) was between 0.39 and 0.99 [19]. In more complex manual handling tasks, errors due to technology were < 5◦ and CMC was 0.79 to 0.97 [18]. This suggests that the MVN BIOMECH system is able to quantify joint kinematics during complex functional movements. In these studies, the motion analysis systems were operated by individuals who were experienced in performing movement analysis experiments, and not by clinicians. However, if an IMU system is to be used in a clinical setting, it is most likely to be used by a clinician with little or no experience of collecting biomechanical data. At present it is not clear if undergoing brief training using only the guides provided by the manufacturer is sufficient to enable a clinician to collect accurate and reliable biomechanical information using a commercially available IMU system. Additionally, the tasks evaluated in previous studies were limited to walking and stair climbing [19], and manual handling tasks [18]. To facilitate transfer of this technology to the clinic it is necessary to quantify the validity and reliability of the Xsens MVN BIOMECH system when used by a clinician, and for clinically relevant functional movements that may involve a greater range of motion or dynamic movements, such as squat and jump.

The primary aim of this study was to determine the reliability of joint angular kinematics provided by the Xsens MVN BIOMECH system when used by a clinically experienced musculoskeletal physiotherapist (MSKP) with no experience of using IMUs across different sessions (i.e., within-rater, between-session repeatability). Secondary aims were (1) to determine the reliability of joint angular kinematics provided the Xsens MVN BIOMECH system when used by an experienced MSKP and an experienced clinical movement scientist (i.e., between-rater, within-session repeatability); and (2) to determine the concurrent validity (agreement) of joint angular kinematics provided the Xsens MVN BIOMECH system when used by an experienced MSKP against data collected using a gold-standard camera-based motion capture system (VICON motion analysis system, Oxford Metrics Group Ltd., Oxford, UK). Validity and reliability were evaluated during walking, squatting and jumping to cover a larger range of motion (squat, jump) and a more dynamic movement (jump) than previously investigated.

#### **2. Materials and Methods**

#### *2.1. Research Participants and Setting*

The study was carried out in the Research Centre for Clinical Kinaesiology at Cardiff University. The study was conducted as part of the Arthritis Research UK Biomechanics and Bioengineering Centre at Cardiff University, and was approved by the Wales Research Ethics Committee 3 (10/MRE09/28). The required sample size (*n* = 16) was determined according to the recommendation of Walter et al. (Table II, [20]), using *α* = 0.05, *β* = 0.2, *ρ*<sup>0</sup> =0.2 and *ρ*<sup>1</sup> = 0.7. To account for an anticipated attrition of 40% over the repeated sessions, the target sample size was increased to 27. A convenience sample of healthy participants was recruited using the following criteria: age between 18 and 60 years; healthy with no known neurological, cardiovascular, or musculoskeletal condition. Written informed consent was obtained prior to participation.

#### *2.2. Raters*

The Xsens MVN BIOMECH system was used by two raters. Rater 1 (MSKP) had 13 years experience as a qualified physiotherapist and 5 years experience as a specialist knee physiotherapist, but no prior experience of biomechanical motion capture. Rater 2 (clinical movement scientist) had 10 years experience performing biomechanical motion capture studies in laboratory settings. The two raters received the same information and instructions for use of the Xsens MVN BIOMECH system. This consisted of self-training using the MVN user manual [21] and online tutorials (https://tutorial. xsens.com/) and a half-day training session provided by Xsens Technologies at Cardiff University.

#### *2.3. Experimental Protocol*

Each participant underwent two movement analysis sessions, approximately 1 week apart (mean ± standard deviation, 4 ± 3 days). In the first session, anthropometric measurements were taken by the MSKP from the right lower limb with the participant in a standing posture. Body mass was recorded using SECA scales (Birmingham, UK). These measurements were only taken in the first session. Apart from this, the two sessions were identical and proceeded as follows:

At the start of each session, 16 retroreflective markers (14-mm diameter) were placed on the participant according to the Plug-in-Gait lower-body model (Vicon Motion Systems, Oxford Metrics Group Ltd.). In every session, retroreflective markers were placed by the MSKP using standardised palpation methods on anatomical landmarks [22]. Retroreflective markers were always attached prior to placement of any IMU, and were held in position with medical grade double-sided adhesive tape.

Seven MTw2 trackers (Xsens Technologies) were then placed in accordance with Xsens instructions [21]. MTw2 trackers were placed by either rater 1 (MSKP) or rater 2 (clinical movement scientist). The order of the raters was randomised across participants, but consistent within participants across sessions. MTw2 trackers were secured using elasticated Velcro straps on each upper thigh (centrally and halfway between the greater trochanter and lateral epicondyle of the knee), each lower leg (proximal medial surface of the tibia), the dorsum of each foot and one centrally over the sacrum. Each lower-limb tracker was placed between the two outermost layers of the strap and attached to the Velcro of the inner layer to secure its position and minimise any movement. The sacral tracker was placed directly over the sacrum with the upper border of the sensor aligned centrally between the two posterior superior iliac spines. The sacral sensor was held in position with medical grade double-sided adhesive tape. Care was taken not to interfere with the reflective markers.

Each participant performed eight repetitions of each of the following three activities: over-ground walking, squatting, and vertical jumping. Prior to performing each activity the participant was provided a demonstration by the MSKP, and was allowed to ask any questions. All activities were performed at a self-selected speed to best mimic performance of these activities in a clinical setting. The order of the activities was randomised across participants, but consistent within participants and across sessions. Each walking trial consisted of a walk in a straight line across the laboratory

(approximately 8 m). Squat depth was standardised to prevent occlusion of the reflective markers on the anterior superior iliac spines at low squat depths. This was done using a plinth placed behind the participant at the height of the knee joint line plus 5 cm, which prevented the participant from squatting past this depth. Once all activities were completed, the MTw2 trackers were removed by the same rater who had placed them. The participant rested for at least 10 min to allow any strap marks on the skin to disappear, and then the MTw2 trackers were placed by the other rater. The retroreflective markers were not moved during this rest period. The participant then repeated the activities in the same order as in the first half of the session.

#### *2.4. Data Collection*

Kinematic data were collected at 120 Hz using a ten-camera VICON MX3+ motion analysis system (Oxford Metrics Group Ltd.) and 60 Hz using the Xsens MVN BIOMECH system (Xsens Technologies). The systems were synchronised using a trigger sent from the MVN BIOMECH studio software (Version 4.4) to VICON Nexus software (Version 1.8). Prior to beginning the tasks, the participant was asked to stand in a static N-pose, as per the instructions in the MVN BIOMECH user manual [21]. This was maintained for ~30 s. At the start of this period of quiet stance, the MTw2 trackers were calibrated within the MVN BIOMECH software. During this process the software establishes the relation between segment and tracker orientations [23,24]. Once this was complete (~10 s), the remaining duration of the quiet stance was synchronously recorded by the two systems. These data were used offline in VICON Nexus software as the static anatomical calibration trial for the Plug-In-Gait model.

#### *2.5. Data Processing*

Data collected using the MVN BIOMECH system were exported as an mvnx file. For data collected using the VICON system, reconstruction and auto-labelling of marker trajectories was first performed in VICON Nexus software. Each trial was then visually inspected and unmarked trajectories were manually labelled. Gaps in trajectories of up to 10 samples were joined with linear interpolation filtered with a quintic spline filter (Woltring; mean square error of 15). The Plug-in Gait pipeline was then implemented in VICON Nexus software, and resulting kinematic data were exported as a c3d file. Plug-in Gait is the commercial name for the implementation of what is commonly called the conventional gait model [22,25–27].

Custom analysis scripts were created in Matlab software (version 2015a; The MathWorks Inc., Natick, MA, USA) and used to perform subsequent analyses. Hip, knee and ankle joint angles calculated by MVN BIOMECH software and VICON Nexus software were extracted from the mvnx and c3d files, respectively. In VICON, positive joint values indicate flexion, adduction and internal rotation in the sagittal, frontal and transverse planes, respectively. In MVN BIOMECH, positive joint angles indicate flexion, abduction and internal rotation, respectively. Frontal plane joint angles from MVN BIOMECH were therefore inverted before any further analysis. Data were filtered with a 6-Hz low-pass fourth-order Butterworth filter. Vicon data were resampled to 60 Hz using the resample function in Matlab. No other post-processing was performed on the joint angles provided by the two systems. Although the systems were synchronised, we observed discrepancies in the duration of the files recorded by the two systems. Due to the uncertainty as to whether this failure of synchronisation occurred at the start or the end of the recording, the mvnx and c3d files were aligned using the 'aligndata' function within Matlab performed on the sagittal plane knee angle.

Movement cycles were defined using data from VICON, as described in Table 1. For each movement cycle, the following variables were quantified for hip, knee and ankle joints on both sides of the body in the sagittal, frontal and transverse planes: minimum angle, maximum angle, range of motion. For walking trials, the joint angle at heel strike was also identified. The minimum and maximum angle, the range of motion and the angle at heel strike represent discrete, clinically relevant events in the movement cycles [28–31]. If the MVN BIOMECH system is to be transferred into the clinic it is important that it can accurately and reliably measure these angles at these time points.


**Table 1.** Definition of start and end of movement cycles for each activity.

\* Heel strike was determined as described by Zeni et al. [32], as the local minima in the anterior-posterior position of the heel relative to the sacrum. The sacrum was defined as the average position of left and right posterior superior iliac spines.

#### *2.6. Data Analysis*

Data analysis was carried out in Matlab software for each activity (walk, jump, squat), each joint (hip, knee, ankle), each plane of movement (sagittal, frontal, transverse) and each side of the body (left, right). For each walk trial, a single stride from the middle of the trial was used for analysis. This gave eight strides for analysis, alongside eight repetitions of squat and jump. The descriptive analysis included mean, standard deviation, difference, and 95% confidence interval of difference of each outcome measure. These were calculated for each participant as an average of all repetitions, before being averaged across all participants. Reliability and validity were quantified for each activity, each joint, and each plane according to the COnsensus-based Standards for the selection of health Measurement INstrument (COSMIN) standards [33,34] as detailed below.

#### 2.6.1. Reliability

Within-rater, between-session reliability was evaluated for rater 1 (MSKP) by comparing data from session 1 and session 2. Between-rater (within-session) reliability was evaluated for session 1 by comparing data from rater 1 (MSKP) and rater 2 (clinical movement scientist). Reliability was quantified using single ICC. For this calculation, the two-way random effects model was used, with a confidence of 95%. ICC was interpreted according to Shrout and Fleiss [35] where ICC ≥ 0.75 indicates excellent repeatability, ICC 0.4–0.74 indicates fair-to-high repeatability, and ICC ≤ 0.39 indicates poor repeatability. Standard error of measurement (SEM) was also calculated using Equation (1) [36]:

$$\text{SEM} = \text{SD} \times \sqrt{1 - \text{ICC}} \tag{1}$$

where SD refers to the standard deviation. SEM was used to evaluate absolute reliability and provide information on variability over repeated measurements [37]. Absolute reliability indicates the reliability of scores within individual participants on different occasions, and was considered excellent if SEM < 3.0◦ and acceptable if SEM < 5◦ [38].

#### 2.6.2. Concurrent Validity

Concurrent validity was evaluated for rater 1 (MSKP) using data from session 1. For each activity (walk, jump, squat), each joint (hip, knee, ankle), each plane of movement (sagittal, frontal, transverse) and each side of the body (left, right), MVN BIOMECH and VICON data were compared using the difference in minimum angle, maximum angle, and range of motion (Δmin *θ*, Δmax *θ*, and ΔROM, respectively), the coefficient of multiple correlation (CMC [10,25,39,40]), and the linear fit method (LFM [41]). In addition, for walking trials, MVN BIOMECH and VICON data were compared using the difference in the angle at heel strike (Δ@HS).

Delta values indicate the similarity of the two signals at clinically relevant points in the movement cycle. The CMC and LFM indicate the similarity of the two signals across the full movement cycle. The CMC was calculated using the formula provided in Ferarri et al. [10]. When the range of motion of the two waveforms is comparable to the offset between them, the CMC is not a real number. Following Ferarri et al. [10,39], CMC was calculated before and after the offset was removed from the two signals. For each signal (MVN BIOMECH and VICON), offset was calculated as the mean of the signal over the entire movement cycle. This value was then subtracted from the signal, giving a signal with zero offset. The CMC before and after offset removal is referred to as CMC1 and CMC2 respectively. CMC is reported only for joints and planes where all values are real numbers [10]. The LFM gives three parameters: α1 indicates the scaling factor, α indicates the scalar addition (i.e., shift or offset) and *R*<sup>2</sup> indicates the strength of the linear relation between the two signals [41]. Concurrent validity was considered excellent if CMC and *R*<sup>2</sup> > 0.75, fair-to-high if CMC and *R*<sup>2</sup> 0.4–0.74, and poor if CMC and *R*<sup>2</sup> < 0.39.

#### **3. Results**

#### *3.1. Demographics and Descriptive*

Twenty-six participants (mean ± standard deviation age: 35.2 ± 8.4 years; height: 162.0 ± 32.9 cm; body mass: 71.6 ± 12.8 kg; body mass index: 25.3 ± 3.8 kg/m2) were enrolled in the study and completed the first session. These individuals were all recruited in the first wave of recruitment. Although this was one below our target recruitment of 27, at the end of the first wave of our recruitment it was clear that our attrition of 4% was well below the allowed 40%, and so a 27th participant was not sought. One participant (4%) did not return for the second session. For the remaining 25 participants, the duration between sessions was 4 ± 3 days. For a second participant, data were not processed for session 1 and for a third participant, data were not processed for the jump activity for session 1, rater 1. In both cases this was because clothing obscured the right thigh marker. Between-session reliability of MVN BIOMECH data was therefore evaluated for 24 participants, apart from the jump where it was evaluated for 23 participants. Similarity of MVN BIOMECH and VICON data was evaluated using data from session 1, and was therefore evaluated for 25 participants. Results were similar for left and right sides of the body, and only results for the left side of the body are presented. The mean ± SD speed of walking was 1.40 ± 0.14 m/s in session 1 and 1.39 ± 0.13 m/s in session 2. The mean ± SD duration of squat was 2.25 ± 0.60 s in session 1 and 2.18 ± 0.62 s in session 2. The mean ± SD duration of jump was 1.21 ± 0.55 s in session 1 and 1.13 ± 0.38 s in session 2.

#### *3.2. Reliability*

The mean and standard deviation of each measure, and the mean difference of each measure between days and raters are presented in Appendix A. For all activities and all joints, the between-session (within-rater) reliability of sagittal plane joint angles provided by MVN BIOMECH was high, with ICC between 0.6 and 0.95 (Figure 1). The between-session (within-rater) reliability of the frontal plane knee angle and transverse plane ankle angle at the heel strike was poor (Figure 1). The absolute between-session (within-rater) reliability of the MVN BIOMECH data was generally acceptable (SEM < 5◦; Figure 2).

**Figure 1.** Between-session (within-rater) intraclass correlation coefficient of MVN data collected by rater 1 (musculoskeletal physiotherapist) in sessions 1 and 2. Intraclass correlation coefficient is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; **bottom row**) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

**Figure 2.** Between-session (within-rater) standard error of measurement of MVN data collected by rater 1 (musculoskeletal physiotherapist) in sessions 1 and 2. Standard error of measurement is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; bottom row) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

For walking and squatting, the between-rater (within-session) reliability of all joint angles provided by MVN BIOMECH was acceptable, with ICC > 0.6 (Figure 3) and SEM < 5◦ (Figure 4) across all planes. For jumping, the between-rater (within-session) reliability of joint angles ranged from poor to excellent (Figures 3 and 4).

**Figure 3.** Between-rater intraclass correlation coefficient of MVN data collected by rater 1 (musculoskeletal physiotherapist) and rater 2 (clinical movement scientist) in session 1. Intraclass correlation coefficient is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; **bottom row**) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

**Figure 4.** Between-rater standard error of measurement of MVN data collected by rater 1 (musculoskeletal physiotherapist) and rater 2 (clinical movement scientist) in session 1. Standard error of measurement is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; **bottom row**) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

#### *3.3. Concurrent Validity*

Figures 5–7 present the joint angle time series' obtained from MVN BIOMECH and VICON systems for each participant during walk, squat, and jump. Across all three activities there appears to be good similarity between the pattern of sagittal and frontal joint angles provided by the two systems. However, an offset in the absolute values provided by the two systems is evident, especially for the hip in the sagittal plane (Figure 5) and the knee and ankle in the transverse plane (Figure 7).

**Figure 5.** Sagittal plane joint angles throughout the movement cycle for each participant. Time series of hip (**top row**), knee (**middle row**) and ankle (**bottom row**) joint angles obtained from VICON (black) and MVN BIOMECH (red) systems for each participant during walk (**left**), squat (**centre**), and jump (**right**). Y-axis represents joint angles in degrees and X-axis represent the movement cycle in percentage.

**Figure 6.** Frontal plane joint angles throughout the movement cycle for each participant. Time series of hip (**top row**), knee (**middle row**) and ankle (**bottom row**) joint angles obtained from VICON (black) and MVN BIOMECH (red) systems for each participant during walk (**left**), squat (**centre**), and jump (**right**). Y-axis represents joint angles in degrees and X-axis represent the movement cycle in percentage. Y-axis scale is the same as in Figure 5 to allow comparison.

**Figure 7.** Transverse plane joint angles throughout the movement cycle for each participant. Time series of hip (**top row**), knee (**middle row**) and ankle (**bottom row**) joint angles obtained from VICON (black) and MVN BIOMECH (red) systems for each participant during walk (**left**), squat (**centre**), and jump (**right**). *Y*-axis represents joint angles in degrees and *X*-axis represent the movement cycle in percentage. *Y*-axis scale is the same as in Figures 5 and 6 to allow comparison.

The difference in joint angle provided by the MVN BIOMECH and VICON systems at discrete, time points is shown in Figure 8.

**Figure 8.** *Cont*.

**Figure 8.** Difference between MVN BIOMECH and VICON data at discrete, clinically relevant events. The difference in minimum angle (Δmin; **top row**), maximum angle (Δmax; **second row**), range of motion (ΔROM; **third row**) and angle at heel strike (Δ@HS; **bottom row**) between the MVN BIOMECH and VICON systems for the hip (**left**), knee (**centre**) and ankle (**right**) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Positive values indicate that MVN BIOMECH angle was larger than VICON angle. Error bars indicate 95% confidence interval of the difference.

The similarity between the joint angle waveforms obtained from the MVN BIOMECH and VICON systems during walk, squat, and jump was evaluated using CMC and *R*<sup>2</sup> provided by the LFM. CMC indicated excellent similarity between MVN BIOMECH and VICON systems (CMC2 > 0.9) for all three joints in the sagittal plane in all functional activities, and excellent similarity for the hip joint in the frontal plane during walking (Figure 9). CMC is not reported for other planes and activities because of the presence of non-real numbers (see Section 2.6.2).

**Figure 9.** The coefficient of multiple correlation (CMC) between MVN BIOMECH and VICON data before (CMC1; **top row**) and after (CMC2; **bottom row**) offset removal for the hip (**left**), knee (**centre**) and ankle (**right**) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. CMC is reported only for joints and planes where all values are real numbers [39]. Where there are three data points in each plane they correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Where there is only one data point it corresponds to the walk task. Error bars indicate 95% confidence intervals.

*R*<sup>2</sup> provided by the LFM indicated excellent similarity between MVN BIOMECH and VICON systems (*R*<sup>2</sup> > 0.8) for sagittal plane angles for all joints across all activities, and fair-to-good similarity (*R*<sup>2</sup> 0.4–0.8) for transverse and frontal plane angles for all joints during squat and jump and knee joint during walking (Figure 10). The similarity between the two systems was poor for transverse plane hip and ankle joint angles during walking and frontal plane ankle angle during walking.

**Figure 10.** The outcome parameters of linear fit method comparing MVN BIOMECH and VICON data. Outcome parameters are *α*1 (scaling factor; **top row**), *α*0 (scalar addition; **middle row**) and *R*<sup>2</sup> (strength of the linear relation between the two signals; bottom row) for the hip (**left**), knee (**centre**) and ankle (**right**) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Error bars indicate 95% confidence intervals.

#### **4. Discussion**

This study aimed to quantify the within- and between-rater reliability and the concurrent validity of joint angles provided by the commercially available Xsens MVN BIOMECH system during three clinically relevant functional activities. Within-rater reliability was fair-to-excellent in all three planes for hip, knee, and ankle joint angles in all three tasks when comparing the kinematics obtained on two different testing days, i.e., between sessions. Between-rater reliability (between a MSKP and a clinical movement scientist) was fair-to-excellent in all three planes during walking and squatting, and poor-to-high during jumping. Concurrent validity (agreement between the Xsens MVN BIOMECH system and VICON Plug-in Gait, a camera-based motion capture system) was excellent in the sagittal plane for hip, knee, and ankle joint angles in all three tasks and acceptable in frontal and transverse planes in squat and jump. These results indicate that, in the present study, the MVN BIOMECH

system was reliably used by an experienced clinician with no prior experience of biomechanical motion capture to quantify lower-limb joint angles in clinically relevant movements. The implications of this for transition of this technology to the clinic are discussed below.

#### *4.1. Reliability*

For walking, the between- and within-rater reliability of discrete kinematic parameters provided by the MVN BIOMECH system was fair to excellent. The ICC and SEM values are similar to published values for camera-based systems [38,42–45]. Most reports of camera-based systems have not reported the reliability of ankle angles in the transverse and frontal planes; however, our results for ankle ROM and maximum ankle angle in the transverse plane during walking are comparable to those of Meldrum et al. [46]. According to a systematic review [38], between-rater ICC values for kinematic variables obtained from camera-based systems across joints in sagittal and frontal planes ranged from 0.5 to 0.99. We report ICC values for the MVN BIOMECH system of between 0.65 and 0.99, with a relatively small SEM (<3.0◦) for all joints, planes and tasks, with the exception of minimum hip angle in the sagittal plane (ICC 0.39 and SEM 3.0◦). This suggests that between- and within-rater reliability of kinematic variables obtained from the MVN BIOMECH system across joints and planes is comparable to or better than those obtained from optoelectronic motion capture systems [38,42–46].

For squatting, the between-rater reliability of kinematic parameters was fair-to-excellent for all parameters across planes. The within-rater reliability was good-to-excellent for all parameters except maximum hip angle in the sagittal plane, where SEM reached 5.5◦. For sagittal plane angles, this within-rater reliability compares well to published values for a camera-based system [47]. Between- and within-rater reliability of joint angles in the frontal and transverse planes have not been reported for camera-based systems.

For jumping, the within-rater reliability of kinematic parameters in the sagittal and frontal planes was good-to-excellent, except for maximum hip angle in the sagittal plane, where SEM reached 7◦, and the within-rater reliability of kinematic parameters in the transverse plane was fair-to-excellent. The within-rater reliability was slightly higher than that reported for drop jumps captured with a camera-based system [48,49].

For jumping, the between-rater reliability of kinematic parameters was fair for the ankle and hip joints in sagittal, frontal, and transverse planes, and slightly worse for the knee joint. The performance of a dynamic task like jumping may have considerable variability across trials [50], and the poor between-rater reliability across planes and joints observed in the jump might be a true reflection of changes in performance, rather than any deficit of the motion capture system. To help us evaluate whether the poor between-rater reliability was due to the motion capture system or true variability in performance we evaluated the reliability of joint angles obtained from the camera-based motion capture system (VICON) in the first and second half of the session. The markers were not moved between the first and second half of the session; however, ICC was between 0.25 and 0.75 (Appendix B, Figures 1 and 2). This indicates that low between-rater reliability observed with the MVN BIOMECH system is likely due to true variability in performance, and is not a deficit of the motion capture system.

#### *4.2. Concurrent Validity*

The high CMC2 and LFM *R*<sup>2</sup> values in the sagittal plane indicate that, in this plane of movement, the joint angle waveforms obtained from the MVN BIOMECH system were similar to those obtained from the VICON system. This supports previous work in healthy participants, which also found a high similarity of joint angle waveforms in the sagittal plane between MVN BIOMECH and an Optotrak motion capture system [18,19], and studies that compared kinematic data estimated from Xsens IMUs using custom-developed techniques to an optoelectronic motion capture system [10,13]. Our results extend these previous findings by including a more challenging dynamic task; the vertical jump. The high similarity between MVN BIOMECH and VICON joint angle waveforms in the jump task indicates that the impact that occurs upon initial contact with the ground when landing from the jump

did not adversely affect the kinematics provided by the MVN BIOMECH system, and suggests that the MVN BIOMECH system can be used in a dynamic task such as jumping.

Previous studies have reported CMC values for data in frontal and transverse planes [10,18,19], but for the majority of joints and activities we were not able to report CMC in these planes. In line with Ferrari et al. [10], we only report CMC values for joints and planes where all values were real numbers. In the frontal and transverse planes, CMC was a non-real number for at least one participant for most joints and activities. This is likely because of small range of motion in these planes for the activities studied. An alternative measure of similarity between waveforms can be obtained from the LFM *R*<sup>2</sup> value [41]. There was a good-to-strong relation between frontal plane data from the two systems for the hip during walking and the ankle during squatting, and a moderate relation in frontal and transverse planes for the hip during squatting and jumping, the knee during all activities, and the ankle during jumping. These results are in line with previous reports of a moderate-to-strong relation between waveforms obtained from the MVN BIOMECH system and an optoelectronic system for hip and knee angles during walking [19], stair ascent and descent [19], and manual handling tasks [18]. We found a poor relation (*R*<sup>2</sup> = 0.1) for the ankle in the transverse and frontal planes during walking, which may be due to the small range of motion in these planes or differences in the anatomical biomechanical definitions between the two systems.

A challenge when validating any new technology is the choice of 'gold-standard' against which to validate. This is particularly relevant in movement analysis where there are several biomechanical models available. We chose to validate against the conventional gait model, which is commercially implemented as Plug-in Gait in VICON Nexus software. One limitation of this model is the effect of errors in markers placement on calculated joint angles [51–55]. For example, a small error in the placement of the thigh maker, which is used to define the internal/external rotation of the femur about the line between the hip joint centre, causes appreciable errors in knee joint kinematics, particularly in the frontal and transverse planes (cf. Figure 11.3; [55]). The high values of knee external rotation obtained from Plug-in Gait in this study (see Figure 7) appear non-physiological, and likely indicate errors in thigh marker placement. There are similar potential effects of errors in shank marker placement on the ankle joint kinematics. The validity results, particularly in the frontal and transverse planes, must be interpreted in light of this limitation. The influence of errors in thigh marker placement can be minimised by using a medial epicondyle marker or knee alignment device, and those of errors in the shank marker placement by using a medial malleolus marker, and future studies may consider these techniques. The CMC values for ankle and hip joint angles in frontal and transverse planes during walking in the current study are lower than those reported by Ferrari et al. [10] and Zhang et al. [19]. In addition to the potential influence marker placement errors, this may be impacted by the biomechanical models used by the systems under comparison. Ferrari et al. [10] compared joint angles obtained from IMUs used with a biomechanical model called Outwalk to those obtained from an optoelectronic system (VICON) used with the Calibration Anatomical System Technique [56]. Our comparison for IMU data was VICON Plug-in Gait. Plug-in Gait estimates joint angles based on Kadaba et al. [26], using segment frames that are constructed from anatomical locations identified by optical markers. By contrast, MVN BIOMECH determines segment frames and changes in body posture based on the segment orientations related to an initial neutral pose (the calibration posture), and calculates joint angles directly from the measured segment orientations. The similarity between biomechanical models used by VICON Plug-In-Gait and MVN BIOMECH is lower than that between the two systems used by Ferrari et al. [10], and this may contribute to the lower waveform similarity.

In the current study, there was a difference (offset) in joint angles provided by the MVN BIOMECH and VICON systems that was particularly noticeable for hip flexion/extension and knee and ankle internal/external rotation. This was constant across participants and activities (see Figures 5 and 7 and small confidence intervals in Figure 8), indicating that it is more likely attributable to differences in the biomechanical models, errors in marker placement or deviation of the body position (posture) during calibration from that required by the system than technical issues such as magnetic field. Another possible explanation is that as calibration data for the two systems were collected at different points in the quiet stance trial (see Section 2.4), the posture of the participant may have been slightly different. However, as the data were collected in the same trial, and the participant did not visibly move throughout this trial, we believe that any difference in posture would have been small.

In summary, there was good similarity in joint angle waveforms between the MVN BIOMECH system and VICON Plug-In Gait for all activities, across all joints and planes of motion, with the exception of the ankle joint during walking in the frontal and transverse planes. Despite the similarity in waveforms, the absolute values reported at discrete time points were not similar, particularly at hip and ankle joints. This is because an offset exists between the systems, likely as a result of different biomechanical models employed by the two systems and/or errors in marker placement. This means that the kinematic outputs from the two systems cannot be used interchangeably; for example, hip flexion cannot be compared between one individual who was measured with the MVN BIOMECH system and another individual who was measured with the VICON Plug-in Gait system.

#### *4.3. Limitations*

Participants in this study were able-bodied adults who attended a single laboratory. Further research is needed to assess the within- and between-rater reliability of the MVN BIOMECH system in pathological populations. The retroreflective markers were placed before the MTw2 trackers, and were not removed until the end of the session (see Section 2.3). This means that the markers were present when both raters placed the MTw2 trackers. The instructions for placement of the MTw2 trackers do not reference anatomical landmarks, therefore we considered the potential impact of the markers being present to be minimal; however, it remains possible that the markers influenced the attachment of the trackers, and thus influenced the inter-rater reliability. We did not use either a medial knee marker or knee alignment device to minimise errors in knee joint angles caused by possible inaccurate thigh marker placement. Our results must be interpreted in with this in mind. Finally, to provide a measure of how repeatable the system is in a clinical environment, participants were not provided with any specific instructions on how to perform the task. This resulted in performance strategy that varied across trials, particularly for the jump. Low reliability in this task is likely influenced by true variations in performance rather than being a reflection of the motion capture system used. This is a clinical strength as the data provide reliability of the system when performance of the task is not constrained.

#### *4.4. Clinical Implications and Future Research Directions*

Although the reliability and criterion validity of any biomechanical motion analysis system is important, it is not sufficient to recommend transfer of the technology into the clinic. Future work should identify what information is relevant to clinicians, the type of motion capture system that is preferable to clinicians, how the information provided by quantitative motion analysis can best be presented to clinicians, and if such information has any impact on diagnosis or clinical decision making.

#### **5. Conclusions**

The good between-rater reliability for walk and squat demonstrate that the MVN BIOMECH system provided joint kinematics that were independent of the rater. This is important as the two raters had different backgrounds and expertise: one was a clinically experienced MSKP with no experience of biomechanical motion analysis and the other was an experienced clinical movement scientist. Our results therefore indicate that the data provided by the MVN BIOMECH system are independent of user expertise and the system does not need to be used by an experienced movement scientist. The good within-rater reliability supports the use of this system across multiple participants or sessions. There was excellent similarity between joint angle waveforms obtained from the MVN BIOMECH and VICON systems in the sagittal plane, and acceptable similarity in the frontal and transverse planes in all three tasks. This extends previous reports and indicates that MVN BIOMECH system can be used in a dynamic task such as a jump. However, it must be noted that the MVN BIOMECH and VICON Plug-in Gait systems cannot be used interchangeably. Together, these results indicate that the commercially available MVN BIOMECH system is suited for clinical movement analysis in clinical practice. Future work should evaluate reliability across centres and in pathological populations, and explore the utility of the information that can be provided by this system to clinicians.

**Acknowledgments:** The authors would like to thank the volunteers for their time, enthusiasm, and feedback. J.D. and K.N. are funded by Arthritis Research UK (Grant No. 18461).

**Author Contributions:** M.A.-A. conceived the initial study design and all authors contributed to the final study design. M.A.-A. and K.N. recruited and enrolled participants and collected data. M.A.-A., K.N. and J.D. processed the data. M.A.-A. and J.L.D. analysed the data and produced figures and tables. M.A.-A., K.N., K.B., L.S. and J.L.D. interpreted the data. M.A.-A. and J.L.D. drafted the manuscript. All authors read and approved the final manuscript.

**Conflicts of Interest:** The MTw2 hardware and MVN BIOMECH software were provided on a temporary basis at no cost by Xsens Technologies (B.V., The Netherlands). The authors received no financial contribution from Xsens Technologies. Training on how to use the MTw2 hardware and MVN BIOMECH software was provided by Xsens Technologies. The final draft of this article was reviewed by Monique Paulich (Senior Product Specialist in Biomechanics, Xsens Technologies) to ensure accuracy of technical details provided on the MVN BIOMECH system. Xsens Technologies had no input on the data interpretation, data analysis, or manuscript writing. None of the authors have any financial interest in Xsens Technologies.


**Table A1.** Within-rater (between-session) reliability of sagittal plane kinematic parameters (left side of the body). Data for each session are mean (standard deviation).

**Appendix**

 **A**

#### *Sensors* **2018** , *18*, 719


 **A2.**Within-rater (between-session) reliability of frontal plane kinematic parameters (left side of the

 body).

**Table**

Data for each session are mean (standard deviation). Data for the difference are difference [95% confidence interval].



**Table**

#### *Sensors* **2018**, *18*, 719


 **A4.** Between-rater (within-session) reliability of sagittal plane kinematic parameters (left side of the

 body).

**Table**



 body).

**Table**

#### *Sensors* **2018** , *18*, 719


**A6.** Between-rater (within-session) reliability of transverse plane kinematic parameters (left side of the

 body).

**Table** 

**Figure 1.** Within-session intraclass correlation coefficient of VICON data collected in session 1. Intraclass correlation coefficient is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; **bottom row**) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

**Figure 2.** Within-session standard error of measurement of VICON data collected in session 1. Standard error of measurement is shown for the minimum angle (min; **top row**), maximum angle (max; **second row**), range of motion (ROM; **third row**) and angle at heel strike (@HS; **bottom row**) for the hip (left-most column), knee (centre column) and ankle (right-most column) joints in the sagittal (Sag), frontal (Frnt) and transverse (Tran) planes of movement. The three data points in each plane correspond to the walk (left-most point; circle), squat (centre point; square) and jump (right-most point; diamond) tasks. Grey shading indicates values considered to indicate excellent reliability.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Motor Planning Error: Toward Measuring Cognitive Frailty in Older Adults Using Wearables**

### **He Zhou 1, Hyoki Lee 1, Jessica Lee 1, Michael Schwenk <sup>2</sup> and Bijan Najafi 1,\***


Received: 18 January 2018; Accepted: 16 March 2018; Published: 20 March 2018

**Abstract:** Practical tools which can be quickly administered are needed for measuring subtle changes in cognitive–motor performance over time. Frailty together with cognitive impairment, or 'cognitive frailty', are shown to be strong and independent predictors of cognitive decline over time. We have developed an interactive instrumented trail-making task (iTMT) platform, which allows quantification of motor planning error (MPE) through a series of ankle reaching tasks. In this study, we examined the accuracy of MPE in identifying cognitive frailty in older adults. Thirty-two older adults (age = 77.3 ± 9.1 years, body-mass-index = 25.3 ± 4.7 kg/m2, female = 38%) were recruited. Using either the Mini-Mental State Examination or Montreal Cognitive Assessment (MoCA), 16 subjects were classified as cognitive-intact and 16 were classified as cognitive-impaired. In addition, 12 young-healthy subjects (age = 26.0 ± 5.2 years, body-mass-index = 25.3 ± 3.9 kg/m2, female = 33%) were recruited to establish a healthy benchmark. Subjects completed the iTMT, using an ankle-worn sensor, which transforms ankle motion into navigation of a computer cursor. The iTMT task included reaching five indexed target circles (including numbers 1-to-3 and letters A&B placed in random order) on the computer-screen by moving the ankle-joint while standing. The ankle-sensor quantifies MPE through analysis of the pattern of ankle velocity. MPE was defined as percentage of time deviation between subject's maximum ankle velocity and the optimal maximum ankle velocity, which is halfway through the reaching pathway. Data from gait tests, including single task and dual task walking, were also collected to determine cognitive–motor performance. The average MPE in young-healthy, elderly cognitive-intact, and elderly cognitive-impaired groups was 11.1 ± 5.7%, 20.3 ± 9.6%, and 34.1 ± 4.2% (*p* < 0.001), respectively. Large effect sizes (Cohen's *d* = 1.17–4.56) were observed for discriminating between groups using MPE. Significant correlations were observed between the MPE and MoCA score (*r* = −0.670, *p* < 0.001) as well as between the MPE and dual task stride velocity (*r* = −0.584, *p* < 0.001). This study demonstrated feasibility and efficacy of estimating MPE from a practical wearable platform with promising results in identifying cognitive–motor impairment and potential application in assessing cognitive frailty. The proposed platform could be also used as an alternative to dual task walking test, where gait assessment may not be practical. Future studies need to confirm these observations in larger samples.

**Keywords:** cognitive frailty; wearable; cognitive–motor impairment; Alzheimer's disease; motor planning error; instrumented trail-making task; ankle reaching task; dual task walking

#### **1. Introduction**

Recent demographic changes have led to emergence of the so-called "dementia epidemic" [1]. Dementia causes great stress to medical, social, and informal care and is currently affecting

approximately 47.5 million persons worldwide [2–4]. This number is projected to increase to 75.6 million by 2030 and 135.5 million by 2050 [3]. This has created an urgent need for a robust and quickly-administered cognitive assessment tool capable of identifying individuals in the earliest stages of cognitive decline and measuring subtle changes in cognitive–motor performance over time [5]. Early diagnosis of mild cognitive impairment (MCI) and detection (or 'measuring') of its subtle progression over time may allow early intervention and prevention of further cognitive and functional decline [6]. In this context, frailty together with cognitive impairment (known as 'cognitive frailty' [7]) has been shown to be a strong and independent predictor of cognitive decline over time [1,8,9].

The concept of 'frailty' is used to identify homeostenotic older adults with low physiological reserves and vulnerability to illness and high risk of disability, institutionalization, and death [10,11]. A recent systematic review identified 67 frailty instruments total; of these, nine were highly-cited (≥200 citations) [12]. Among these, the Physical Frailty Phenotype [10] was the most used frailty instrument in the research literature, followed by the Deficit Accumulation Index [11]. The Physical Frailty Phenotype is based on five criteria (weight loss, exhaustion, low physical activity, weak grip strength, and slow walking speed) [10]. The Deficit Accumulation Index is based on a count of 70 accumulated deficits, including the presence and/or severity of diseases, ability in activities of daily living, and physical signs from clinical and neurological examinations [11]. While both assessments have shown to be accurate for identifying older adults with low resilience and high vulnerability independent of comorbidities and disability [13,14], they all neglect assessing cognitive performance, including working memory and executive function. The cognitive assessment could add complementary information (in addition to frailty) to determine healthy aging trajectory as well as post intervention health outcomes. For example, previous studies have suggested that cognitive performance is an independent risk factor influencing adverse health outcomes after major surgical intervention [15,16].

A recent study by Kelaiditi et al. [9] demonstrated that measures of frailty can be used to identify individuals with Alzheimer's disease (AD) who have a greater risk of cognitive decline and loss of independence. An association between cognitive impairment and specific frailty criteria such as gait speed has been demonstrated [17]. Similarly, several studies have suggested that cognitive and motor declines are independent key predictors of dementia in the geriatric population [8,18,19].

Despite this evidence described above, cognitive impairment and frailty are still assessed and studied separately, mainly because conventional cognitive performance assessment tools do not take into account motor performance (an indicator of frailty), and vice versa. In addition, the current motor or cognitive performance assessment tools often neglect characterization of anticipatory motor planning, an important factor that influences ability to accomplish instrumental activities of daily living (iADL) [20–22].

Studies based on motor learning paradigms highlight the importance of characterizing anticipatory motor planning as an important factor that influences daily motor skills and cognitive performance [23,24]. Most of iADLs rely on the individual's anticipatory planning abilities to appropriately plan movements prior to executing them. For example, to successfully accomplish a behavioral goal such as reaching for a piece of fruit, two problems must be solved: to select which fruit to reach for and to specify the parameters of the reach, such as its direction and extent [24]. Studies on primary motor cortex, premotor cortex, and supplementary motor area have demonstrated that human brains can plan and control physical movement ahead of physical body action [25–27]. It is well established that anticipatory planning abilities improve as a function of age, reaching adult-like levels in late childhood to early adolescence. It then rapidly declines to levels observed in young children, in very old adults, particularly in those with frailty and/or cognitive impairment [28]. It has been postulated that the decrease in anticipatory motor planning abilities in older adults is associated with decline in cognitive skills [28], a theory supported by brain imaging research in which an association between deterioration in frontal and parietal white matter and fine motor dexterity was reported [29,30]. Recently, Stockel and colleagues [28] have demonstrated that up to 64% of

the variance in motor planning performance across age groups can be explained by the cognitive functions of processing speed, response planning, and cognitive flexibility. They have postulated that anticipatory motor planning abilities are strongly influenced by cognitive control processes, which seem to be key mechanisms compensating for age related decline.

To address the practical barriers for clinical use of cognitive–motor assessment tools as well as incorporating a practical method to quantify motor planning, we have developed a novel cognitive frailty measurement tool call instrumented trail-making task (iTMT). The iTMT is based on the combination of a low-cost wearable sensor attached to patient's lower shin and interactive interface technology. The design of the system is inspired by the conventional trail-making test (TMT), which is a useful test to determine cognitive impairment [31]. The second component of the system is an ankle reaching test performed in a virtual environment, which allows for quantification of motor performance and motor planning error. The proposed system provides a unique opportunity to determine cognitive frailty by enabling simultaneous and rapid measurement of both cognitive and motor performance. In our previous study [32], we have demonstrated the effectiveness of the iTMT to determine cognitive impairment by measuring the time duration needed to complete the test. In addition, we have demonstrated that by measuring the ankle velocity during the reaching task, we can also distinguish pre-frail individuals and frail individuals from non-frail individuals. In the present study, we extended our iTMT approach by quantifying anticipatory motor planning error. We extracted the motor planning error (MPE) through measuring ankle velocity patterns during the ankle reaching task. Our hypothesis was that; due to anticipatory motor planning performance decline associated with cognitive impairment, older adults with cognitive impairment would have iTMT MPEs larger than those of elderly without cognitive impairment as well as those of young-healthy subjects.

#### **2. Methods**

#### *2.1. Study Population*

Thirty-two ambulatory older adults (age 65 years or older) were recruited in this study. According to the Mini-Mental State Examination (MMSE) (cutoff score of 23 or less [33]) or Montreal Cognitive Assessment (MoCA) (cutoff score of 25 or less [34]), 16 subjects were classified as cognitive-intact, and 16 were classified as cognitive-impaired. To compare the iTMT results between young and older adults, as well as to establish a healthy benchmark, 12 young-healthy ambulatory subjects with ages ranging from 21 to 34 years were also recruited. Subjects were excluded from the study if they were unable to walk a distance of 20 m with or without walking assistance; had visual problem, limiting their ability to interact with a computer screen; had lower extremity problems, limiting their ability to perform ankle movement needed for the purpose of this study; or had severe balance impairment, limiting their ability to independently stand. Those who could stand behind a chair and perform the iTMT test by holding the chair were not excluded. All subjects signed a consent form for this study. This study was approved by the local institutional review boards.

#### *2.2. Demographic Information*

All subjects' demographics were collected, including age, gender, body mass, height, body-mass-index (BMI), history of falls, and history of depression. Subjects self-reported any fall incidents and number of falls in the past year. The Center for Epidemiologic Studies Depression (CES-D) short-version scale was used to measure self-reported depression symptoms. A cutoff of CES-D score of 16 or greater was used to identify subjects with depression.

#### *2.3. Instrumented Trail-Making Task (iTMT)*

We designed the iTMT platform (Figure 1) based on a wearable sensor (LEGSys™, Biosensics LLC, Watertown, MA, USA), which has a triaxial accelerometer, triaxial gyroscope, and triaxial magnetometer for the purpose of estimating angles and positions. The wearable sensor was attached to the lower shin for tracking three dimensional ankle motion (Figure 1). The real-time data was acquired and transmitted at 100-Hz frequency to an interactive computer interface (developed based on MATLAB® 2013b and Psychophysics Toolbox Version 2.54) installed on a standard computer (Figure 1). The subject could navigate a cursor from a home circle to target circles appearing on the computer screen using ankle rotation.

**Figure 1.** An illustration of the instrumented trail-making task (iTMT) platform with both numbers (1, 2, and 3) and letters (A and B) randomized in the target circles. One inertial sensor—including a triaxial accelerometer, a triaxial gyroscope, and a triaxial magnetometer—was attached to the subject's lower shin using a comfortable elastic band. The sensor allows measurement of three-dimensional motion of the ankle joint in real time. The instantaneous measured joint angle with a sample frequency of 100-Hz was wirelessly transferred to a computer, using low-power Bluetooth, to create an interactive interface for the purpose of the interactive iTMT test. For safety purposes, a research coordinator was in the room supervising the iTMT test at all times. After starting the iTMT test, the research coordinator did not provide any guidance. The interactive interface provided the necessary guidance and instruction to complete the test.

In the iTMT test, six circles appeared on the screen: one home circle in white and five target circles in yellow. The target circles were located in a fanwise position in front of the home circle (Figure 1). The five target circles were randomly marked with numbers (1, 2, and 3) and letters (A and B). At the beginning of the iTMT test, the cursor position was automatically calibrated to the center of the home circle. The iTMT required weight shifting in forward/backward or sideward/diagonal direction, while standing with both feet on the ground in front of the computer screen (Figure 1). To perform the iTMT, the subject would control the movement of ankle joints while maintaining balance. This could be achieved by moving the hip joint in opposite direction of ankle motion (reciprocal compensatory strategy) [35]. Ankle motions on the left foot and right foot are consistent, therefore by attaching just one sensor on one lower shin, three dimensional ankle motion can be tracked. By rotating the ankle joint, the subject could navigate the cursor to the center of target circles with alternating numbers and letters. Specifically, the subject navigated the cursor to the center of the first target circle (with the number '1' inside). The target circle would explode with a reward sound once the cursor stopped in the middle of the target circle. Then the subject navigated the cursor back to the center of the home circle, then to the second target circle (with the letter 'A'), then back to the home circle, then to the third target circle (with the number '2'), and so forth until the end. To efficiently accomplish the test with minimum time to reach all targets, the subject should have optimum motor planning prior to executing the task. For example, to successfully reach a target, two problems should be solved: to select which target to reach for (i.e., working memory task) and to specify the parameters of the reach, such as its direction and extent. In addition, to minimize the time to reach a target, the subject

should determine optimum time to decelerate ankle movement so as stop perfectly at the center of the target without re-adjustment.

#### *2.4. iTMT Motor Planning Error*

We extracted the motor planning error based on ankle velocity pattern measurements during the ankle reaching task. MPE was defined as percentage of time deviation between subject's maximum ankle velocity and the optimal maximum ankle velocity, which is halfway through the reaching pathway (Figure 2). In summary, to minimize the iTMT time, which requires bringing a computer cursor (square shape) from a home circle and stopping at the middle of the target circle (which was defined as the end of a reaching task), the subject should reach maximum ankle velocity at the middle of the reaching pathway and then decelerate to be able to stop at the middle of the target without the need for re-adjustment. Therefore, the ideal ankle velocity curve is a bell shape (indicator of a healthy feedforward performance). Based on our definition, the iTMT MPE can range from 0 to 50%. 0% indicates that the peak ankle velocity appears at the mid-pathway (optimum motor planning). 50% indicates that the peak ankle velocity appears at the beginning or end of target reaching. Figure 2A–C show ankle velocity curves of a typical young-healthy subject, elderly cognitive–motor intact subject, and elderly cognitive–motor impaired subject, respectively. For the typical young-healthy subject with good cognitive perception coordinated with good motor performance, the peak ankle velocity appears very close to the mid-pathway, and the iTMT MPE is very small (Figure 2A). For the typical elderly cognitive–motor intact subject, the peak ankle velocity is lower than the young-healthy subject, and the iTMT MPE is moderate (Figure 2B). For the typical elderly cognitive–motor impaired subject with poor cognition and uncoordinated motion, the peak ankle velocity appears very early or very late, and the iTMT MPE is large. Multiple ankle velocity peaks can be observed in Figure 2C.

**Figure 2.** An illustration of the iTMT MPE. (**A**) Ankle velocity curve of a typical young-healthy subject during the ankle reaching, (**B**) ankle velocity curve of a typical elderly cognitive–motor intact subject during the ankle reaching, (**C**) ankle velocity curve of a typical elderly cognitive–motor impaired subject during the ankle reaching.

#### *2.5. iTMT Time*

The method for measuring the iTMT time has been described in detail in [32]. In summary, at the beginning of the iTMT test, when the cursor position is automatically calibrated to the center of the home circle, the computer program starts recording the time. At the end of the iTMT test, when the cursor position stops at the center of the last target circle (with the number '3'), the computer program stops recording the time. The recorded time period is defined as the iTMT time, which represents the total time a subject uses to complete reaching to all five indexed target circles in the correct order. Our previous study suggests that the iTMT time result is correlated with MoCA test result (*r* = −0.598, *p* = 0.001) and is effective at discriminating between those with and without cognitive impairment with large effect size (Cohen's *d* effect size = 1.02, *p* = 0.015) [32]. We have also demonstrated that the iTMT time represents mainly cognitive performance and corresponds poorly with motor performance

(e.g., gait speed) [32]. To compare this marker with the innovative iTMT MPE proposed in this study, iTMT time was collected for all subjects.

#### *2.6. Walking Test*

For all subjects, wearable sensors were attached to both the left and right lower legs (LegSys™, BioSensics, Watertown, MA, USA) to measure single task and dual task gait performances. Subjects were asked to walk with their habitual gait speed for 20 m with no cognitive task (single task walking). Then, they were asked to repeat the test while loudly counting backward from a random number (dual task walking: walking + working memory test) [2,36]. Gait speed was calculated using validated algorithms [37–39].

#### *2.7. Statistical Analysis*

Age, height, body mass, BMI, single task walking stride velocity, and dual task walking stride velocity for each group were described using both mean ± standard deviation. Gender, history of falls, and history of depression were expressed as count (percentage). The Shapiro–Wilk test was applied to test normality of the data. One-way analysis of variance was employed for pairwise comparison according to the scale of the investigated variable and the distribution of the data. Analysis of covariance was used to compare the iTMT MPE and the iTMT time, with and without adjustment for age and BMI, among groups. Fisher's least significant difference-based post hoc test was performed for pairwise comparison to explore significant main effects and interactions. Cohen's *d* effect size was calculated to assess the magnitude of difference between each group. Values ranging from 0.20 to 0.49 indicate small, and values between 0.50 and 0.79 indicate medium. Values ranging from 0.80 to 1.29 indicate large effects, and values above 1.30 indicate very large effects. Values less than 0.20 are considered as having no noticeable effect. Spearman's correlation coefficients were used to evaluate the degree of association between the iTMT MPEs and conventional dual task walking stride velocity, used as a reference. *p* < 0.050 was considered statistically significant. All statistical analyses were performed using IBM SPSS Statistics 24 (IBM, Chicago, IL, USA).

#### Sample Size Calculation

We estimated prior sample size and power by performing a post hoc analysis based on the observed effect size in the previous iTMT study [32]. Considering the observed effect size of 1.02 for the iTMT time difference between cognitive-intact and cognitive-impaired in [32], to achieve a minimum power of 80%, a minimum sample size of 13 subjects per group are required to observe a statistical significant of 5% or lower using two tails independent sample comparisons.

#### **3. Results**

#### *3.1. Study Population*

Table 1 summarizes demographic and clinical data. The elderly subjects' ages ranged from 60 to 93 years, while the young-healthy subjects' ages ranged from 21 to 34 years. No difference between the elderly cognitive-intact group and elderly cognitive-impaired group was observed for demographic information, including gender, age, height, body mass, and BMI (*p* > 0.050). For the clinical examinations including history of fall and depression, no between-group difference was observed (*p* > 0.050). Motor performance based on single task walking did not show significant difference between the elderly cognitive-intact group and elderly cognitive-impaired group. The only significant difference between the two groups was the dual task walking stride velocity (*p* = 0.031). Compared to the elderly cognitive-intact group, the young-healthy group had significantly increased single task and dual task walking stride velocities (*p* < 0.001 and *p* = 0.001).


**Table 1.** General characteristics of the study population.

BMI: Body-Mass-Index; STW: Single Task Walking; DTW: Dual Task Walking; SV: Stride Velocity. Depression was assessed by Center for Epidemiologic Studies Depression (CES-D) score with a cutoff of 16 or greater. Significant difference between groups were indicated in bold.

#### *3.2. iTMT Motor Planning Error and iTMT Time among Groups*

Table 2 summarizes the iTMT MPE and the iTMT time for different groups, including young-healthy subjects, elderly cognitive-intact subjects, as well as elderly cognitive-impaired subjects. Table 3 and Figure 3 summarize between-group comparisons with and without adjustment for age and BMI. The iTMT MPE generated during the test was 20.3 ± 9.6% in the elderly cognitive-intact group and was significantly increased on average by 68% in the elderly cognitive-impaired group (*d* = 1.86, *p* < 0.001, Figure 3A). The iTMT time spent by the elderly cognitive-intact group to complete the test was 25.2 ± 7.9s. It was significantly increased on average by 100% in the elderly cognitive-impaired group (*d* = 1.21, *p* < 0.001, Figure 3B). When compared with the elderly groups, a much lower iTMT MPE (11.1 ± 5.7%, Figure 3A) and a much shorter iTMT time (18.5 ± 2.1s, Figure 3B) was observed for the young-healthy group. When compared with the iTMT time, the iTMT MPE had much larger effect sizes to discriminate between the three groups (Table 3). After adjusting the results for age and BMI, the between-group difference achieved statistical significance only when comparing between the elderly cognitive-intact and elderly cognitive-impaired groups for both iTMT MPE (*d* = 1.26, *p* < 0.001) and iTMT time (*d* = 0.98, *p* < 0.001) (Table 3).


**Table 2.** iTMT derived parameters for different groups.

iTMT: instrumented trail-making task. Significant difference between groups were indicated in bold.



 trail-making Significant groups 

#### *Sensors* **2018** , *18*, 926

**Figure 3.** The iTMT derived parameters for all groups, including young-healthy group, elderly cognitive-intact group, and elderly cognitive-impaired group. (**A**) The iTMT MPE comparison; (**B**) the iTMT time comparison. Error bar represents the standard error. '\*' denotes when the pairwise group comparison achieved a statistically significant level (*p* < 0.050). *d* denotes Cohen's *d* effect size.

#### *3.3. Association between iTMT Motor Planning Errors and Conventional Cognitive Assessment*

Figure 4 illustrates the correlation between the iTMT platform and conventional cognitive assessments. In summary, a high correlation was observed between the iTMT MPE and MoCA score (*r* = −0.670, *p* < 0.001, Figure 4A). On the same note, a high correlation was also observed between the iTMT MPE and dual task walking stride velocity (*r* = −0.584, *p* < 0.001, Figure 4B), indicating that the iTMT MPE may be used as a surrogate for the dual task walking test.

**Figure 4.** Correlation between the iTMT MPE and (**A**) MoCA test score and (**B**) dual task walking stride velocity (DTW SV).

#### **4. Discussion**

Cognitive and motor impairments are independent key predictors of further decline in cognitive performance, life independency, and Alzheimer's disease among geriatric population [8,18,19]. In this study, we investigated the feasibility and efficacy of estimating motor planning error from a practical wearable platform with promising results in identifying cognitive–motor impairment in older adults. We were able to confirm our hypothesis that the iTMT MPE is able to discriminate between elderly cognitive-intact and elderly cognitive-impaired groups. The iTMT MPE value is derived from the ankle velocity signals during trail-making and ankle-reaching tasks for examination of both cognitive and motor abilities. In other words, the ankle velocity pattern can provide information on both cognition and motor ability of patients. For example, the velocity pattern of the elderly cognitive-impaired group fluctuated more than that of the elderly cognitive-intact group (Figure 2). This may be due to the fact that both trail-making and ankle-reaching tasks require performing automatic corrections under voluntary control by the brain in order to sharpen target localization [40]. Therefore, the noisy velocity pattern signifies the decrease of motor planning skills in geriatric populations with cognitive impairment.

Cognitive impairment is a symptom that makes it difficult for patients to remember, learn new things, concentrate, or make decisions. While cognitive impairment can be identified with pencil- and paper-based screening tools, such as the MoCA, MMSE, and conventional TMT, these assessments are semi-subjective, time consuming, and insensitive to subtle changes in cognitive frailty. Their accuracy is highly dependent on the examiner's experience and the patient's education level [32,41]. Computerized versions of conventional cognitive screening tools have improved the utility of such measurements. However, they are not capable of monitoring motor performance (an essential component of frailty) and anticipatory planning abilities (an essential component for successfully accomplishment of instrumental activities of daily living) thus are not able to detect cognitive frailty, a known predictor of speed of cognitive decline over time. The dual task paradigm can be used to screen for both cognitive and motor performance and identify both cognitive impairment and frailty [2,36]. The most stablished dual task paradigm is dual task walking. Dual task walking can isolate the cognitive control component of locomotion and expose cognitive deficits through the evaluation of activities that simultaneously demand attentional resources. In addition, dual task walking test is sensitive in determining frailty [36]. However, many older adults have mobility impairment and high risk of falling, making the administration of walking test challenging, in particular in busy clinical settings. In addition, limits on space and time in busy clinics can influence the execution of such tests [42]. Even if gait is assessed over short walking distance to accommodate space limitation, it is debated whether dual-task walking test is reliable over a short walking distance (less than 20 m) [38,43]. The proposed ankle reaching task derived MPE addresses the aforementioned limitations associated with penciland paper-based screening tools as well as dual task gait. The iTMT platform consists of a low-cost wearable sensor combined with an interactive interface installable on any standard computer [32]. It is simple, short, safe, and easy to manage by non-expert users [32]. The iTMT platform does not require walking test. It allows long-term quantification of anticipatory motor planning regardless of setting.

In an early study [32], we demonstrated that the iTMT time is an efficient marker to discriminate among age-matched healthy, amnestic mild cognitive impairment, and Alzheimer's disease groups. However, the iTMT time is a global marker of the trail-making performance. It is the total time that a subject needs to complete the iTMT test and can be influenced by many factors such as the time a subject needs to search and locate targets, the speed a subject can rotate the ankle, the number of mistakes a subject makes during the test, etc. Therefore, the iTMT time is not a specific maker of cognitive–motor impairment. It may not be as sensitive as the iTMT MPE in identifying cognitive–motor impairment and assessing cognitive frailty. The iTMT MPE is derived from ankle velocity measurements during the trail-making test. It is a specific marker which can evaluate both cognitive and motor abilities equally and simultaneously. We collected both iTMT markers in the present study and compared them in Tables 2 and 3. Although Table 2 shows that both iTMT markers possessed significant differences

between the three groups (*p* < 0.001), Table 3 shows that the iTMT MPE had a larger effect size than the iTMT time for discriminating among the three groups, mainly because of relatively lower intra group variation, in particular among the elderly group with cognitive impairment. For instance, when examining the difference between older adults with and without cognitive impairment, the effect size was increased from 1.21 (large effect size) by using the iTMT time to 1.86 (very large effect size) by using the iTMT MPE. In addition, only the iTMT MPE can discriminate between young-healthy and elderly cognitive-intact without adjustment for age. This may be because the major difference between these two groups exists in the motor domain caused by aging. Thus, we may speculate that the iTMT MPE is a cognitive–motor sensitive metric, and may be able to determine motor deficit between groups with comparable cognitive functions. Before adjusting for age and BMI, the iTMT MPE was able to discriminate between all three groups. After adjusting for age and BMI, the difference between young-healthy and older adults was diminished irrespective of cognitive status. However, when comparing between older adults with and without cognitive impairment, the between-group difference remained significant after adjustment. This may suggest that the iTMT MPE is sensitive to age but also able to determine cognitive impairment in the same age categories.

One of the unique opportunities of the iTMT platform is its ability to determine frailty status. In another study, we demonstrated that iTMT parameters were also able to identify presence and absence of frailty phenotypes (Cohen's *d* effect size = 0.81–1.56). Specifically, the iTMT maximum ankle reaching velocity was associated with slowness, changes in ankle velocity from the first reaching task to the last reaching task was an indicator of exhaustion, ankle reaching moment was a representation of weakness, and jerkiness of ankle movement was associated with inactivity. These parameters allow identification of non-frail, pre-frail, and frail older adults, as identified by the Physical Frailty Phenotype [10], with large to very large effect sizes (*d* = 1.04–3.14). Together with results of the current study in which we demonstrated that motor planning error is an indicator of cognitive impairment, it is postulated that the iTMT platform is able to determine cognitive frailty, a strong predictor of elderly dependency in the near future [8]. Recent studies suggest that frailty is also a significant predictor of cognitive decline over a short period of time. For example, Kelaiditi et al. [9] demonstrated that frail patients with AD have almost twice the decline in cognitive performance over one year follow up than AD patients without frailty. The same study concluded that a one-unit (0.033 points) increase in the frailty index (indicator of cumulative deficits) corresponds to significant and clinically relevant cognitive decline, after adjusting for age, sex, and years of education (0.63–4.63 points on the MMSE, *p* = 0.010).

Another advantage of the iTMT platform is that it uses wearable technology, instead of camera or force platform, to evaluate motor-cognitive performance in older adults. Unlike low-cost camera-based motion tracking system (e.g., Microsoft Kinect), wearable sensors do not require a continuous unobstructed sightline; and thus the test could be administrated behind of a chair to provide support if needed. In addition, the examiner could be next to the subject to provide safety, unlike with the camera-based system. This safety feature is especially important during the iTMT in older adults, in particular, those with cognitive impairment and dementia, who have increased fall risk. The force platform is similar to wearable sensor, which is not limited by continuous unobstructed sightline. However it has other limitations making it unsuitable for motor planning error assessment. For instance, commercially available and low-cost force platforms (e.g., Nintendo Wii Fit) restrict the base of support during testing, which is unsuitable for obese patients or those who need large base of support to maintain balance during dynamic tasks [44]. Furthermore, wearable sensors enable measuring kinematics of body joint of interest and thus diversify the tests. Force platforms, however, are limited to measure only center of pressure or ground reaction force without any information from joint angles.

#### *Limitations and Future Directions*

Despite the advantages of the iTMT platform enabling us to prescreen for cognitive–motor impairment and cognitive frailty, there are some limitations in this study. Our subjects were recruited from specialized outpatient clinics (e.g., cancer clinic, vascular clinic, dementia clinic, etc.) instead of older adults dwelling communities. Thus our sample may not represent general elderly population. The instruments (MMSE or MoCA) used to identify cognitive impairment in different clinics were not identical. Future study should be done on different days and be administered by different examiners in order to assess test–retest reliability. In addition, this study analyzed the iTMT MPE while standing, which disqualifies bedbound and non-ambulatory patients. This study invites further analysis of the motor planning error while sitting or lying down in order to identify cognitive–motor impairment in bedbound patients.

#### **5. Conclusions**

In conclusion, we have demonstrated that motor planning error, which could be estimated from a low-cost and simple wearable platform, is able to simultaneously determine cognitive and motor impairments. Based on our observations, the test is feasible and practical even for those with severe cognitive impairment (MoCA < 17) and poor motor performance (single task walking stride velocity < 0.5m/s). The test is fast (taking less than one minute on average, excluding preparation and sensor attachment), it requires a single inertial sensor (i.e., gyroscope) and low-cost standard computer or tablet, and is reliable as demonstrated in our previous study [32]. Despite the simplicity, it can provide important clinical information about cognitive–motor impairment and cognitive frailty in geriatric populations. The proposed iTMT platform could also be used as an alternative test for dual task gait, which is often impractical in busy clinics and homes, where the space for gait assessment is not available, and for patients with limited mobility (e.g., need to walk with walking accessories). Thus, it can easily be used for both community dwelling elderly as well as those in busy hospital settings.

**Acknowledgments:** Partial support was provided by the National Institutes of Health/National Institute on Aging (award number 2R42AG032748), the National Institutes of Health/National Cancer Institute (award number 1R21CA190933-01A1), Baylor College of Medicine, and Michael E. DeBakey Department of Surgery. The content is solely the responsibility of the authors and does not necessarily represent the official views of sponsors. We thank Ana Enriquez and Ivan Marin for assisting with data collection and coordination of this research study between involved key investigators.

**Author Contributions:** H.Z. wrote the manuscript, developed the code for extracting motor planning error, and contributed in data collection and data analysis. H.L., J.L., and M.S. contributed in drafting the manuscript. B.N. contributed in study design and supervising the study. All authors contributed in interpretation of results and critical revision of the study.

**Conflicts of Interest:** The iTMT is protected by a patent pending. H.Z. and B.N. are listed as co-inventors. However, they do not claim any financial conflict of interest relevant to this study.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Article*

### **Nondestructive Estimation of Muscle Contributions to STS Training with Different Loadings Based on Wearable Sensor System**

#### **Kun Liu, Yong Liu, Jianchao Yan \* and Zhenyuan Sun**

School of Mechanical Science and Engineering, Jilin University, Changchun 130025, China; kunliu@jlu.edu.cn (K.L.); liuyong16@mails.jlu.edu.cn (Y.L.); zysun17@mails.jlu.edu.cn (Z.S.) **\*** Correspondence: yanjc16@mails.jlu.edu.cn

Received: 26 January 2018; Accepted: 21 March 2018; Published: 25 March 2018

**Abstract:** Partial body weight support or loading sit-to-stand (STS) rehabilitation can be useful for persons with lower limb dysfunction to achieve movement again based on the internal residual muscle force and external assistance. To explicate how the muscles contribute to the kinetics and kinematics of STS performance by non-invasive in vitro detection and to nondestructively estimate the muscle contributions to STS training with different loadings, a wearable sensor system was developed with ground reaction force (GRF) platforms, motion capture inertial sensors and electromyography (EMG) sensors. To estimate the internal moments of hip, knee and ankle joints and quantify the contributions of individual muscle and gravity to STS movement, the inverse dynamics analysis on a simplified STS biomechanical model with external loading is proposed. The functional roles of the lower limb individual muscles (rectus femoris (RF), gluteus maximus (GM), vastus lateralis (VL), tibialis anterior (TA) and gastrocnemius (GAST)) during STS motion and the mechanism of the muscles' synergies to perform STS-specific subtasks were analyzed. The muscle contributions to the biomechanical STS subtasks of vertical propulsion, anteroposterior (AP) braking and propulsion for body balance in the sagittal plane were quantified by experimental studies with EMG, kinematic and kinetic data.

**Keywords:** nondestructive; joint moment; partial weight loading; muscle contributions; sit-to-stand training

#### **1. Introduction**

Sit-to-stand (STS) movement is a common, but critical action that almost every independent person performs every day, but millions of people have difficulty in STS motion due to neurological pathologies globally [1]. STS movement is also a complex dynamic task that requires the regulation of lower limb muscles to drive the human body and maintain the dynamic balance while performing subtasks, such as leaning forward movement of the HAT (head-arms-trunk segments) in the chair, vertical stretching movement of the whole body and balancing movement of the whole body off the chair [2–5]. Patients with a decrease in leg muscle function experience difficulties in achieving STS movement, but improvement may occur with proper training [6–9]. STS movement requires coordination of several muscle groups to guarantee human balance while achieving the task. However, for the elderly and dependent people who have lost part of the lower limb functionality, the activity becomes tiring and cannot be accomplished without the help of external assistance [10]. To restore muscle strength and improve STS movement coordination, repetitive STS training with appropriate external support is very important for the patients with lower limb motion impairments. Consequently, assistant systems, such as exoskeleton/orthosis or a partial body weight support (PBWS) rehabilitation robot, can be used to provide and control external assistance [11–14].

A greater understanding of the biomechanics of STS and muscle contributions to STS has important implications for rehabilitation training. It is valuable for clinicians to make a targeted, effective and efficient physical therapy plan in individual STS rehabilitation training with motor disability. In common STS movement, the primary mechanism to regulate joint moments is muscle force generation. Muscles accelerate body segments and generate ground reaction forces that alter joint moments about the segments' center-of-mass (COM) to restore and maintain dynamic stability. In addition, gravity contributes to whole-body angular momentum through its contribution to the ground reaction force (GRF) [15]. During STS rehabilitation training, since certain muscles of the lower limb are weak due to neurological pathologies or surgery injury, the patients need external assistant forces (EAF), such as PBWS force on the shoulder or assistant torque on the knee joint. Accordingly the patient is driven to stand up by lower limb muscles and EAF cooperatively. However, because the individual situation varies, it is hard to quantitatively evaluate the muscles' rehabilitation phase and to accurately estimate the muscle contributions to dynamic body support without intrusive observation and measurement. Few previous studies have quantified which muscles primarily contributed to whole-body angular momentum of STS. Identification of muscles' responsibilities and contributions in STS motion has important implications for diagnosis and treatment of balance and movement disorders and the design of effective rehabilitation therapies that target specific muscle groups. Gravity acts throughout the segment, but can be located at the COM for calculation purposes. The gravity contributes to joint reaction forces, and negative joint moments with the moment arms from each segment's COM to the corresponding joint consumed the joint moments that are offered by muscle synergies. There have been many studies on the biomechanics of STS, and many papers have reported findings related to electromyographic (EMG) activity during the action [16–18]. The applications of computational biomechanics for orthopedic treatment and rehabilitation have attracted much attention to provide better quality of life for patients. Especially a greater understanding of the biomechanics and muscle activity of STS movement has important implications to design safe, effective and efficient rehabilitation training guidelines for individuals with motor disability [19–21]. Therefore, non-invasive estimation of joint moments, joint forces and muscle contributions to body support with loading for STS training using a wearable sensor system is very important for STS rehabilitation.

McGowan et al. [22] offered a method to estimate lower limb muscles' contributions to body support and forward propulsion. However, they just analyzed independent effects of weight and mass on plantar flexor activity during walking, not STS movement. Neptune et al. [23] analyzed the muscle contributions to whole-body sagittal plane angular momentum during walking. These results may provide insight into balance and movement disorders and provide a basis for developing locomotor therapies that target specific muscle groups, but there was no practical significance for STS rehabilitation. Wang et al. predicted the joint moments using a neural network model of muscle activations from EMG signals [24], but the method was used to calculate the elbow joint moments, not the lower limb joint moments from the EMG signals of ten flexor and extensor muscles. Thomas Stieglitz et al. presented a noninvasive measurement of torque development in the rat foot, but it is based on electrode implantation in the plantar/dorsiflexion and medial/lateral rotation plane. In a sense, it is not really a noninvasive measurement [25]. Furukawa et al. [26] introduced a newly-developed biosignal-based vertical weight support system, which was composed of pneumatic artificial muscles (PAMs) and an EMG measurement device. It proved that external assistance force can be varied based on measured muscle activities and can be used to instruct STS rehabilitation training. Bonnet et al. [27] investigated the possibility of estimating hip and knee joint angles using a single inertial measurement unit, but they did not pay attention to the rehabilitation of lower limb muscles. Zlatko et al. analyzed and compared kinematics, kinetics and EMG patterns of STS transfer [28] and proved they can be used for STS training. Tatsuya analyzed an EMG-based predictive control model for physical human-robot interaction, but it was not suitable for STS training [29]. Nicole et al. analyzed muscle contribution and coordination during stair ascent using a 26-camera optoelectronic motion capture system. The insightful studies provided a detailed understanding of the biomechanics during stair ascent at the joint level, but it' is not a real-time and wearable method for analyzing joint moments [30].

In this paper, the differences of muscle excitation intensity in different phases during STS are identified. The purpose of the study is to identify the functional roles of the lower limb individual muscles during STS movement and the mechanisms by which muscles work together to perform STS-specific subtasks. The muscle contributions to STS motion at the individual muscle level were analyzed by experimental studies with EMG data, kinematic and kinetic data. The STS kinetics and kinematics were non-invasively estimated and analyzed using developed wearable inertial sensors. The patterns of five groups of the lower limb muscles during standing up were characterized by processed EMG to examine the relationship between the muscle activities and STS kinetics. Furthermore, to identify individual muscle contributions to the biomechanical STS subtasks of vertical propulsion, anteroposterior (AP) braking and propulsion and to analyze how individual muscles, EAF and gravity contribute to STS motion and body balance in the sagittal plane, a partial-muscle-actuated STS movement was generated with different loadings on the HAT of the subject. Based on inverse dynamics analysis of the STS biomechanical model, we estimated hip, knee and ankle joint moments from accelerations, angular velocities and angles measured on the lower limb segments (HAT, thigh and shank) using inertial sensors. The results showed insight into the movement coordination of STS and had implications for the ongoing development and testing of more effective STS training techniques in the clinic.

#### **2. Methods**

Compared to horizontal walking during which the COM of the body is predominantly propelled, STS motion requires an individual to propel the COM significantly more vertically and less horizontally. In addition, it is obvious that dynamic balance is more difficult to maintain after the body left the seat (seat-off). These differences suggest that STS motion likely requires altered muscle contributions relative to different phases before and after seat-off. To estimate the complex relationships between muscle excitation and resultant force, the muscle contributions to accelerate joints and segments in the STS biomechanical task should be analyzed based on the STS dynamic biomechanical model as shown in Figure 1. The subjects rise from sitting in an essentially sagittal symmetric manner; therefore, a two-dimensional three-link segmental model is used in the analysis of the STS task under the assumption of bilateral symmetry in the sagittal plane.

**Figure 1.** Link segment model of dynamic sit-to-stand (STS) biomechanical model.

*Sensors* **2018**, *18*, 971

In most tasks involving kinetic measurements of the STS subject, a direct measurement of internal moments acting on lower limb joints was not feasible for PBWS rehabilitation robot users. A dynamic STS biomechanical model of a trainee consisting of the HAT, thighs, shanks and feet is needed to estimate and analyze joint moments with external GRFs, loading forces and kinetic and kinematic parameters (inertial effects). In addition, initial feet position in the STS biomechanical model directly affects the distance that HAT must move forward in STS motion. It was proven that the amplitude of displacement and velocity at the hip could be significantly optimized when the feet are behind the knee in the AP direction [31,32]. Thus, the STS movement is divided into four phases [33,34]: in the first phase, the subject in the sitting position with feet under the knees bends forward the HAT. In the second phase, the subject lifts off the chair. In the third phase, the subject extends the knee joint and the whole body stretches. Then, in the final phase, the subject is full upright in the stand-up position. Therefore, the human body can be represented by three different parts, shank-foot, thigh and HAT, and can be modeled for the STS movements as a triple inverted pendulum model in the sagittal plane articulated around the ankle, knee and hip joints. In our presented work [35], we had non-invasively estimated the joint moments using the developed wearable sensor system and analyzed the kinematic and kinetic profiles underlying the STS movement. As a further analysis of STS using wearable sensors, we recalculated the joint moments *Mi* (*i* = 1, 2, 3) based on the inertial and static effects of STS motion as follows.

Hip joint moment is shown with Equations (1)–(3):

$$
\stackrel{\rightarrow}{\dot{M}}\_{3-IINERTIA} = \frac{1}{m} \cdot \left(\frac{1}{2} \cdot \stackrel{\rightarrow}{J}\_3 \cdot \stackrel{\rightarrow}{a}\_3 + \stackrel{\rightarrow}{r}\_{33} \times \left(\frac{1}{2} \cdot m\_3 \stackrel{\rightarrow}{a}\_3\right)\right) \tag{1}
$$

$$
\overrightarrow{M}\_{3-STATIC} = \frac{1}{m} \cdot (\overrightarrow{r}\_{33} \times (\frac{1}{2} \cdot m\_3 \cdot \overrightarrow{g} )) \tag{2}
$$

$$
\stackrel{\rightarrow}{M}\_3 = \stackrel{\rightarrow}{M}\_{3-INERTIA} + \stackrel{\rightarrow}{M}\_{3-STATIC} \tag{3}
$$

Knee joint moment is shown with Equations (4)–(6):

$$\stackrel{\rightarrow}{M}\_{2-I\text{INRTIA}} = \frac{1}{m} \cdot (-\frac{1}{2} \cdot l\_3 \cdot \stackrel{\rightarrow}{a}\_3 - l\_2 \cdot \stackrel{\rightarrow}{a}\_2 - \stackrel{\rightarrow}{r}\_{23} \times (\frac{1}{2} \cdot m\_3 \cdot \stackrel{\rightarrow}{a}\_3) - \stackrel{\rightarrow}{r}\_{22} \times (\frac{1}{2} \cdot m\_2 \cdot \stackrel{\rightarrow}{a}\_2)) \tag{4}$$

$$\stackrel{\rightarrow}{\mathcal{M}}\_{2-STATIC} = \frac{1}{m} \cdot \left( -\stackrel{\rightarrow}{r}\_{23} \times \left( \frac{1}{2} \cdot m\_3 \cdot \stackrel{\rightarrow}{\mathcal{g}} \right) - \stackrel{\rightarrow}{r}\_{22} \times \left( m\_2 \cdot \stackrel{\rightarrow}{\mathcal{g}} \right) \right) \tag{5}$$

$$
\stackrel{\rightarrow}{\dot{M}}\_2 = \stackrel{\rightarrow}{\dot{M}}\_{2-INERTIA} + \stackrel{\rightarrow}{\dot{M}}\_{2-STATIC} \tag{6}
$$

Ankle joint moment is shown with Equations (7)–(9):

$$\stackrel{\rightarrow}{M}\_{1-INEXTA} = \frac{1}{\underline{w}} \cdot (\stackrel{1}{2} \cdot I\_3 \cdot \stackrel{\rightarrow}{\cdot a}\_3 + I\_2 \cdot \stackrel{\rightarrow}{\cdot a}\_2 + I\_1 \cdot \stackrel{\rightarrow}{\cdot a}\_1 + \stackrel{\rightarrow}{r}\_{13} \times (\stackrel{1}{2} \cdot m\_3 \cdot \stackrel{\rightarrow}{\cdot a}\_3) + \stackrel{\rightarrow}{r}\_{12} \times (m\_2 \cdot \stackrel{\rightarrow}{a}\_2) + \stackrel{\rightarrow}{r}\_{11} \times (m\_1 \cdot \stackrel{\rightarrow}{a}\_1) \tag{7}$$

$$
\stackrel{\rightarrow}{M}\_{1-STATIC} = \frac{1}{m} \cdot \left( \stackrel{\rightarrow}{r}\_{13} \times \left( \frac{1}{2} \cdot m\_3 \cdot \stackrel{\rightarrow}{\mathcal{g}} \right) + \stackrel{\rightarrow}{r}\_{12} \times \left( m\_2 \cdot \stackrel{\rightarrow}{\mathcal{g}} \right) + \stackrel{\rightarrow}{r}\_{11} \times \left( m\_1 \cdot \stackrel{\rightarrow}{\mathcal{g}} \right) \right) \tag{8}
$$

$$
\stackrel{\rightarrow}{\dot{M}}\_1 = \stackrel{\rightarrow}{\dot{M}}\_{1-INETIA} + \stackrel{\rightarrow}{\dot{M}}\_{1-STATIC} \tag{9}
$$

where <sup>→</sup> *Mi* is the joint moment vector about joint *i*; *Jj* the moment of inertia of segment *j* about the center of mass; <sup>→</sup> *<sup>α</sup> <sup>j</sup>* the angular acceleration vector of segment *j* about the center of mass; <sup>→</sup> *r ij* the position vector from joint *i* to the center of gravity of segment *j*; *mj* the mass of segment *j*; *m* the whole body mass with loading on the HAT; <sup>→</sup> *a <sup>j</sup>* the acceleration vector of the center of gravity of segment *j*.

To quantify the causal relationships between the lower limb muscle excitation inputs and the resulting STS performance, the musculoskeletal modeling with explicit equations for the neuromuscular and musculoskeletal system dynamics should be utilized. Previous studies have used simulations to identify individual muscle contributions to tasks such as walking [36] and wheelchair propulsion [37]. Recently, Lin et al. [38] performed a simulation analysis to determine the contributions of five muscle groups to body support, forward propulsion and balance control during stair ascent by analyzing whole-body COM accelerations. The insightful studies in [30] provided a detailed understanding of the biomechanics of stair ascent. It was obvious that the five lower limb muscles (rectus femoris (RF), gluteus maximus (GM), vastus lateralis (VL), tibialis anterior (TA) and gastrocnemius (GAST)) among the analyzed fifteen muscles made the main contributions to the vertical propulsion. Therefore, to nondestructively identify the functional roles of the lower limb individual muscles during STS training movement, our research focused on the five lower limb muscles' contributions to the biomechanical STS subtasks of vertical propulsion, AP braking and propulsion in the sagittal plane. Meanwhile, to develop a wearable succedaneum of the bulky camera optoelectronic motion capture system for detailed understanding of how individual muscles, EAF and gravity contribute to STS motion and body balance, a wearable sensor system was developed.

The mass and dimension of each segment of the subject were estimated based on the average current Chinese male inertia parameters of body segments according to Chinese national standards, as shown in Table 1. The moment of inertia of each segment was estimated based on the height and the total mass of the subject [39] and shown in Table 1. All segments were assumed to be rigid, and the STS movement was performed only in the sagittal plane.



#### **3. Experiment**

A sensor system composed of two force plates, inertial sensor modules (IMUs) and EMG electrodes was developed. To measure the kinematic parameters of the segments for calculating joint moments according to Equations (1)–(9), three customized IMUs (wearable sensor JY-901B, 1.1 × 1.1 × 0.5 inches with battery and Bluetooth communication) were attached on the position of COM of the shank, thigh and HAT with Velcro, as shown in Figure 1. A microcontroller (Arduino UNO) was used to capture accelerations and angular velocities from the IMUs at a 100-Hz sampling rate, store data and communicate with a PC in real time. The two force plates were developed with pressure sensors (YZC-1B), and the forces were sampled at a 100-Hz rate using the microcontroller (Arduino UNO). Force Plate A was fixed on the chair to measure the vertical chair reaction force (VCRF) before seat-off. Force Plate B was placed under subjects' feet to measure the vertical ground reaction force (VGRF). During the initial calibration, the force plates were positioned horizontally with calibration errors of 0.74% and 0.68% of the measured vertical forces. To relate each IMU's orientation to the corresponding segment's orientation, a sensor-to-segment calibration procedure is performed as presented in [40]. The X-axis of each IMU was repeatedly adjusted to coincide with the axial direction of its corresponding segment in the segment coordinate frames based on the recommendations of the International Society of Biomechanics [41]. The wearable sensor system has been verified in our previous research work [35] to be available for non-invasive estimation of joint moments and analysis of STS kinematics.

EMG signals from five lower limb muscles of the right leg, RF, GM, VL, TA and GAST, were monitored by bipolar surface electrodes (Meditrace 0.11-m self-adhesive Ag/AgCl electrodes). The electrode positions are indicated in Figure 2. The proximal electrode was positioned above either the bulkiest part or the middle of the muscle belly with a 0.01-m standard interelectrode distance according to [42]. At the electrode locations, the skin was shaved, cleaned with alcohol and lightly abrased to further reduce skin resistance. Electrical silence in all five muscles was achieved prior to data collection. The EMG signals were processed by means of two customized self-developed 3-channel amplifiers as shown in Figure 3 with a band-pass filter from 10 Hz–500 Hz. The signal was amplified (1 K) and converted from the analog-to-digital form via an analog-to-digital interface board (PCI 1713U) at a sampling frequency of 1000 Hz. To procure a relative impression of the degree of activity, all EMG signals were full-wave rectified and low-pass filtered.

**Figure 2.** Schematic presentation of the EMG electrodes' placement and muscles in the sagittal plane. GM, gluteus maximus; VL, vastus lateralis; GAST, gastrocnemius; RF, rectus femoris; TA, tibialis anterior.

**Figure 3.** Schematic diagram of the proposed method.

Ten healthy male subjects (age 26 ± 2 years, height 1.74 ± 0.05 m, weight 66 ± 6.5 kg) without any known musculoskeletal or neurological dysfunction participated in this study and received informed

consent. The experimental protocol was approved by the Human Ethical Review Committee of Jilin University. After familiarization and practice, the subjects reported no serious impediment of the developed sensor system. Then, each subject performed three STS trials as a task and five STS tasks with different loadings (0 kg, 5 kg, 10 kg, 15 kg and 20 kg) on the HAT at a self-selected preferred speed and movement pattern without any support. In each initial sitting posture of the STS trials, the feet position were located behind the knee in the anteroposterior direction with self-selected appropriate distance of about half the length of the foot. Therefore, 150 trials (3 trials × 5 groups of tasks per subject × 10 subjects) were achieved. Although a task of three STS trials was performed by one subject, the STS time and the amount of captured data of each trial could not be absolutely the same. Three events provided reference points for kinematic data: movement onset (when the angular velocity of the COM of HAT being greater than 0.01 rad/s), seat-off (zero VCRF) and movement end (when the angular velocity of the COM of thigh was less than 0.01 rad/s or VCRF approximately equaled the subject's weight). The three events enabled the action to be divided into two phases: a pre-stretch phase (movement onset to seat-off) and a stretch phase (seat-off to movement end). Therefore, the time of seat-off (zero VCRF) was designated as the referenced standard point (RSP) within whole STS to synchronize the kinematic, kinetic and EMG data derived from the three STS trials of a task to the same percentage metric, then the ensemble averages of a task were obtained. The kinematic data, VGRF, VCRF and the calculated joint moments were normalized to the subject's height and weight, then were used to analyze the muscle contributions to body support in different-loading STS training compared with the five groups of EMG data captured from RF, GM, VL, TA and GAST.

#### **4. Results**

The wearable sensor system has been verified in our earlier research work [35]. It was available for non-invasive estimation of joint moments and analysis of STS kinematics. All signals captured by the developed sensor system were off-line processed by MATLAB. A typical group of the ensemble averages of the synchronized kinetic and kinematic parameters and EMG signals derived from five different loading STS tasks performed by one healthy subject is shown in Figures 4–8. In each figure, the first column shows the calculated hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly) and VCRF and VGRF (blue and red lines, correspondingly). The second column shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). The third column shows the bar graph of the analysis of the muscle contribution index (MCI) based on the surface EMG signal after further calculation.

Voluntary muscle activity detection from EMG signals can be problematic due to spurious involuntary spikes produced by physiological and extrinsic/accidental origins [43]. Accurate detection of surface EMG signals for the lower limb muscle activities is important to understand how muscles made the contributions to STS motion. However, the EMG signal distortion is unavoidable with the removal of the spike contamination, thus impeding reliable measurement of weak EMG signals. The study in [44] demonstrated the usefulness of the integrated profile of EMG for muscle activity detection using surface EMG signals in the presence of spurious background spike contamination. The integrated profile method was used to determinate the onset of muscle contraction and contribution to STS motion without removing spurious background spikes from raw surface EMG signals. In this paper, as further processing of the integrated profile of EMG, MCI is defined to quantitatively indicate the five lower limb muscles' contributions to the biomechanical STS subtasks of vertical propulsion, AP braking and propulsion in the sagittal plane. The MCI of each muscle was calculated from the ensemble averages of the synchronized integrated profile of EMG of the three STS trials in the same task and was quantified by 50 integers in a whole STS cycle by percentage metric.

**Figure 4.** The analysis result of a typical STS task with no loadings: **the left column** shows the hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly), and vertical chair reaction force (VCRF) and vertical ground reaction force (VGRF) (blue and red lines, correspondingly). **The middle column** shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). **The third column** shows the muscle contribution indexes (MCI) of the five lower limb muscles (TA, GAST, RF, VL and GM).

**Figure 5.** *Cont*.

**Figure 5.** The analysis result of a typical STS task with 5-kg loadings: **the left column** shows the hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly) and VCRF and VGRF (blue and red lines, correspondingly). **The middle column** shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). **The third column** shows the muscle contribution indexes (MCI) of the five lower limb muscles (TA, GAST, RF, VL and GM).

**Figure 6.** *Cont*.

**Figure 6.** The analysis result of a typical STS task with 10-kg loadings: **the left column** shows the hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly) and VCRF and VGRF (blue and red lines, correspondingly). **The middle column** shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). **The third column** shows the muscle contribution indexes (MCI) of the five lower limb muscles (TA, GAST, RF, VL and GM).

**Figure 7.** The analysis result of a typical STS task with 15-kg loadings: **the left column** shows the hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly) and VCRF and VGRF (blue and red lines, correspondingly). **The middle column** shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). **The third column** shows the muscle contribution indexes (MCI) of the five lower limb muscles (TA, GAST, RF, VL and GM).

**Figure 8.** The analysis result of a typical STS task with 20-kg loadings: **the left column** shows the hip, knee and ankle joint angles and moments (green, red and blue lines, correspondingly), angular velocities and angular accelerations of the HAT, thigh and shank (green, red and blue lines, correspondingly) and VCRF and VGRF (blue and red lines, correspondingly). **The middle column** shows the surface EMG signals of the five lower limb muscles (TA, GAST, RF, VL and GM). **The third column** shows the muscle contribution indexes (MCI) of the five lower limb muscles (TA, GAST, RF, VL and GM).

#### **5. Discussion**

As indicated in [35], faster and more fluent STS movements resulted in better accuracy of the kinetic and kinematic analysis of STS using the developed inertial sensor system. Since it was more difficult to firmly fix the IMUs on the soft human body segments than on a rigid body without any slight movement, the systematic error was predictable, but inevitable. Especially in lower speed STS motion, the long duration of skin motion artifact due to impact loading and muscle activation and body-sway motion in non-fluent STS trials would inevitably contaminate the measured original angular velocities, accelerations and the raw EMG signals.

The difference of the five muscles was due to their relative contributions to the horizontal and vertical GRFs and joint moments resulting in different kinematics of STS motions with different loadings. Referring to the joint angles of the hip, knee and ankle in the first column of Figures 4–8, it is suggested that the HAT took action first among the three segments of the subject. When the HAT was still swinging forward before the hip joint angle reached the maximum, the knee joint angle had begun to increase since seat-off; meanwhile, the HAT continued to swing forward into the stretch phase (after seat-off), then the ankle joint began to increase in the terminal stretch phase. Compared with the EMG signals, MCIs and the kinetic and kinematic parameters in Figures 4–8, the highest levels of the five muscles' excitation intensities (indicated by maximum MCI) all occurred before seat-off. Meanwhile, as shown in Figure 9, the percentages of each muscle's accumulated MCI before seat-off were all no less than 58% of the total contributions to whole STS motion, which suggests that the muscles' contributions to the whole STS motion primarily concentrated before seat-off compared to those after seat-off. However the max knee joint moment appeared after seat-off, and the HAT still swung forward after seat-off into the stretch phase. It is, therefore, quite clear that the muscles' co-contractions were recruited ahead of the stretch subtask before seat-off. The superimposition before seat-off could be intended as a synergic action of muscles for controlling weight bearing as static joint moments, then the muscle synergies continued to increase until the inertial joint moments were adequate to promote vertical propulsion, anteroposterior (AP) braking and propulsion for performing a balanced body support of STS motion.

In addition, with regard to the sum of MCI of each muscle before seat-off in different loading tasks in Figure 10, the absolute accumulated MCI of GAST before seat-off was the least among the five muscles in any task (no more than 18 in the 20-kg task, minimum MCI = 10 in the 0-kg task), and as shown in Figures 4–8, the MCI of GAST was even weaker or not detected after seat-off. This suggested that GAST made the least contribution to STS motion among the five muscles in any loading task, but almost all of its contributions were made before seat-off (100% in the 0-kg task in Figure 10). It is, therefore, quite clear that GAST primarily contributed to the preparation of a stable stretch phase by holding the shank before seat-off and mainly contributed to the ankle joint extension and holding after seat-off. Compared the MCIs of TA and GAST in Figures 4–8, TA had the same characteristic as GAST to contribute synergy for holding the shank and ankle joint extension before seat-off, but it was obvious that TA contracted earlier (almost from the beginning) and contributed more synergy in duration and intensity than GAST. Compared with the onset times of the five muscles in each loading task, GAST and VL contracted relatively the latest of the five muscles (almost after 20% of the STS cycle). This once again suggested that the primary contributions of GAST (including VL here) were to vertical STS propulsion preparation and balance holding of the lower limb, but not to the swing of HAT before seat-off. Besides, with regard to the onset time and MCI of each muscle in Figures 4–8, the muscles' activities in the stretch phase were recognized as typical activations of RF muscle as knee extensors, GM muscle as hip extensors and TA muscle as ankle extensors. Comparing the joint angles and moments with the corresponding MCIs of GM, TA and RF in the same task in Figures 4–8, the high correlation of the onset time and amplitude suggested that the GM, TA and RF were the primary contributors to forward propulsion of HAT from the beginning, then all five muscles co-contracted to apply sufficient synergies for vertical propulsion until the subject seated off. Obviously, as shown in Figure 11, the GM contributed the most among the five muscles in duration and intensity to accelerate HAT forwards in the pre-stretch phase and keep stable hip joint extension in the stretch phase in any loading task. For example, in Figure 10, before seat-off in each task, the accumulated MCI of GM was 56, 73, 77, 98 and 113, respectively, any of which was the biggest of the five muscles in the same loading task. It is obvious in Figure 10 that, with the addition of loadings, the muscles' contributions indicated by MCI also increased. This fully illustrates the causal relationship between muscle force and loadings. For example, in the successful STS task with the 20-kg loading, as shown in Figure 8, all of the MCIs of the five muscles indicated the maximum contributions compared with other loading tasks. On the contrary, this also suggested that deficits in the neuromuscular control of the five selected muscles could adversely influence balance recovery, which may be important targets in rehabilitation to improve balance recovery performance. As the MCIs of VL, RF and GM in each task show, their synergic action after seat-off was still obvious. This suggests that the synergies were presented to assist knee and hip extensions and develop muscle tension for weight and external loading acceptance in the stretch phase, meanwhile providing the majority of braking accelerations for the balance recovery in the terminal stretch phase before movement end.

**Figure 9.** Percentage metric of each muscle's accumulated contribution before seat-off in different tasks.

**Figure 10.** The sum of each muscle's MCI before seat-off in different tasks.

**Figure 11.** Percentage metric of each muscle's contraction duration in whole STS cycle of different loading tasks.

Kinetics analysis in the stretch phase: To control STS balance, the antagonistic muscles spanning the hip, knee and ankle joints distributed power from the leg to the remaining body segments. This also suggested that the uniarticular hip and knee extensors (GM and VL) and ankle dorsiflexors (TA) generated backward angular momentum while the ankle plantar flexors (GAS) generated forward momentum; therefore, the whole body began vertical propulsion with enough kinetic energy into the stretch phase, then the muscles' co-contractions obviously reduced with less MCIs after seat-off. That RF and VL overlapped their activities in the early stretch phase just after seat-off (refer to the third column of Figures 4–8) was likely synergizing for modulating rapid knee flexion, although in less than 10% of the whole STS cycle and with low levels of MCI.

In particular, there are still some uncommon and variable activities of the TA, RF and GM indicated by significant MCIs in the middle and terminal stretch phases before movement end (in the third column of Figures 4–8), probably because with the stretching of the body segments and the vertical propulsion of the COM after seat-off, the primary work of lower limb muscles had converted to support body balance. RF and VL supposedly contributed to knee flexion and patella stabilization in the middle and terminal stretch phase; RF and GM participated in hip stabilization; and TA and GAST contributed to ankle flexion and stabilization. According to the consultancies with subjects, as evidence, the deficiency of sufficient stability of the self-made force plate indeed caused an instantaneous stretch balance adjustment in the terminal stretch phase. Especially in the tasks with more loadings (15 kg and 20 kg), TA, RF and GM should instantaneously recruit obvious muscle synergies for stretch balance. Compared with the MCIs of TA, RF and GM in the stretch phases of the five loading tasks in Figures 4–8, the MCIs of the three muscles became higher with more obvious fluctuation as the loadings increased, so this confirmed that more loadings would induce more muscle activities and co-contractions of TA, RF and GM for balance control in the stretch phase.

The results and discussion could provide insight into STS balance and movement disorders and offer a reference for developing locomotor therapies that target specific muscle groups; meanwhile, it may provide further rationale for developing targeted STS rehabilitation robot control strategies to address patient-specific deficits in STS training. Although the selected lower limb muscles' relative contributions to STS motion were compared and analyzed based on the muscle excitation intensities and durations and the kinetics and kinematics of STS motion, outcomes on muscle co-contraction and absolute contributions of more lower limb muscles should be confirmed and supported by further analysis. In addition, differences in designated loadings and self-selected STS speeds across subjects could introduce variability in muscle onset-offset times and co-contraction. Thus, STS speed and loadings could be acknowledged as a limitation of the study.

#### **6. Conclusions**

Quantification of lower limb muscles' co-contraction contributing to STS propulsion movement with different loadings was analyzed based on a self-developed wearable sensor system. How the muscles generate, absorb and/or transfer mechanical power between body segments to accelerate the whole-body COM was identified by quantitative assessment of five lower limb muscles' contributions during able-bodied STS motion, in terms of variability of onset-offset muscular activation, excitation intensity indicated by MCI and the percentage occurrence of EMG signals and kinetic and kinematic information derived from the wearable sensor system. The results would contribute to further understanding of more muscles' contributions and coordination during STS movement and provide additional insight for the development of effective, targeted rehabilitation programs aimed at improving an individual's ability to STS rehabilitation training and useful in the clinical context for designing future STS or gait studies. In the future, partial body weight support or more specified loading would be tested on patients in rehabilitation by more motion models, including assisted standing motion with hands or weight support rehabilitation robots. Finally, the presented method of nondestructive estimation of muscle contributions to STS motion based on the wearable sensor system is expected to be applied to weight support or loading rehabilitation robot control or evaluation of STS rehabilitation therapy.

**Acknowledgments:** The authors wish to acknowledge the support of the volunteer subjects of the Robotics and Dynamics Research Lab in Jilin University. The work was partly supported by the Youth Research Fund Project of the Department of Science and Technology of Jilin province (20160520066JH) and the Graduate Innovation Fund of Jilin University.

**Author Contributions:** Kun Liu, Jianchao Yan, Yong Liu and Zhenyuan Sun conceived of and designed the experiments. Kun Liu and Zhenyuan Sun helped to perform the experiments. Kun Liu, Jianchao Yan and Yong Liu analyzed the data. Kun Liu, Jianchao Yan and Yong Liu wrote the paper. In addition, Kun Liu and Jianchao Yan were responsible for the implementation of the proposed scheme.

**Conflicts of Interest:** The authors declare no conflict of interest.

### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **The Height-Adaptive Parameterized Step Length Measurement Method and Experiment Based on Motion Parameters**

#### **Yanshun Zhang 1, Yingyue Li 1,\*, Chuang Peng 1, Dong Mou 2, Ming Li <sup>1</sup> and Wei Wang <sup>1</sup>**


Received: 8 February 2018; Accepted: 22 March 2018; Published: 30 March 2018

**Abstract:** In order to tackle the inaccurate step length measurement of people with different heights and in different motion states, a height-adaptive method of step length measurement based on motion parameters is proposed in this paper. This method takes people's height, stride frequency, and changing accelerometer output while walking into integrated consideration, and builds a dynamic and parameterized model of their step length. In this study, these parameters were calibrated with thirty sets of experiment data from people with different heights and in different motion states, which were then verified experimentally by motion data of randomly selected subjects, regardless of speed and height. The experiment results indicate that the height-adaptive step length measurement was realized, thus eliminating the influence of height exerted on step length measurement.

**Keywords:** motion parameters; gait analysis; step length; self-adaptation

#### **1. Introduction**

As one of the important parameters reflecting people's motion characteristics, step length can be used in the research of measurements of body motion parameters, disease diagnosis and treatment, health monitoring, rehabilitation training, and pedestrian navigation [1–5]. Motion parameters measured by small Micro Electro Mechanical Systems (MEMS) inertial sensors, at a low cost and with high precision, render step length measurement feasible and effective [3,6].

Step length measurement bas been an important aspect of gait analysis. Many authors employ foot-mounted or leg-mounted IMU which is the abbreviation of "Inertial measurement unit" and can be used as measuring device to get data. However, this can cause a change in angle of the IMU, because of the transformation of the foot or leg when the subjects walk. We employed awaist-mounted IMU to measure the data, so that no matter how the angle of the foot changed, the orientation of the IMU would stay the same; this way, we were able to reduce the parameters of the algorithm [7–10].

Different body characteristics and motion parameters lead to unavoidable errors in step length measurement. This problem can be solved by adopting the corresponding step length measurement under different circumstances. At present, there are three main methods of step length measurement:

The first method calculates the step length based on geometry models. Cavagna et al. (among other scholars) believe that the displacement of the horizontal direction in a single foot motion during walking can be described as an inverted pendulum model, providing the mathematic relationship between the displacement of body's center of gravity in the horizontal direction, and that in the vertical direction [11–13]. A modified pendulum model is proposed by González et al. to estimate real-time step length, with each step divided into the single foot support phase and the double foot support phase [14]. Without the prophase training, Kun-Chan Lan et al. also bring up a step length measurement method based on the Pythagorean theorem [15]. However, because of the integral operation involved, the methods based on geometry models mentioned above can easily result in drift errors.

The second method resorts to the nonlinear empirical formula of the step length. By studying human walking, Weinberg proposes the nonlinear step length measurement method based on the peak values and the valley values of the acceleration in the center of gravity [16]. Due to its simplicity and easy application, this formula is used by a group of scholars researching pedestrian navigation, either directly or indirectly [3,5,17,18]. Parameter recalibration is required when dealing with a variety of pedestrians.

The third method is based on the linear combination: Levi et al. utilize a constant and stride frequency [19], while Ladetto takes both stride frequency and acceleration variance into consideration [2,20,21]. If applied in pedestrians with different heights, these two methods require parameter recalibration, which lacks wide adaptability. Another linear combination method based on height and stride frequency is presented by Renaudin et al., with the measurement precision of step length increased [22]. The measurement precision, however, deteriorates when there are strenuous motions during human walking.

Summarizing advantages and disadvantages of the methods mentioned above, this paper introduces a novel step length measurement method with the integration of pedestrians' height, stride frequency, and acceleration variance during walking, by analyzing the gait characteristics. Meanwhile, the corresponding experimental research is performed.

#### **2. Analysis of Gait Characteristics**

When people walk, there is a motion in virtually each part of body (feet, legs, waist, etc.) [23,24], which means the acceleration speed and angular velocity are constantly changing. The motion of feet and legs are relatively strenuous, with apparent acceleration and angular velocity change, making it easy to extract useful information from them. It is convenient to wear and fix the sensors at the waist, and there is only a little influence exerted on the body motion because of the gentle waist motion. When wearing sensors at the waist, the change of the acceleration and angular velocity in the vertical direction is more obvious than that in other directions, thus facilitating the analysis, extraction, and estimation of the human motion status. The vertical waist acceleration is shown in Figure 1.

**Figure 1.** The relationship between the vertical acceleration and gait.

From the figure above, it is seen that the center of gravity changes periodically up and down with each step. In addition, the vertical acceleration of the center of gravity changes periodically too, which leads to a different step length resulting from different walking habits, body characteristics, and walking status. Here, "step length" is the distance between the footsteps of the left and right foot. As shown in Figure 2, the step length is the distance between the blue foot and the red foot. The relationship between the step length and the changing vertical acceleration needs to be studied. The corresponding parameters have to be identified and extracted from the accelerometer or gyroscope for real-time calculation.

**Figure 2.** The illustration of the step length.

In practical applications, as shown in Figure 3, there are sensor noises, different periodical peak values, and false peak values from the accelerator output resulting from sensor detection errors and step inconformity.

**Figure 3.** The change of vertical acceleration during walking.

The overlay of walking speed, the periodical change of center of gravity, and the variation in the heaviness of step generates the vertical acceleration. All of these factors reflect walking status from which the step length is calculated. It is feasible to study the relationship between these factors and the step length, and then conduct the step length measurement.

#### **3. Step Length Measurement Method**

#### *3.1. The Step Length Measurement Method Based on Stride Frequency and Acceleration Variance*

By analyzing different walking speeds and types of gait per person during walking, Ladetto proposes the step length measurement based on the linear relation of the step length, stride frequency, and acceleration variance [18], which can be expressed as follows:

$$SL\_i = A \cdot f\_i + B \cdot var\_i + \mathbb{C} \tag{1}$$

where *A* is the coefficient of the stride frequency, *B* is the coefficient of the vertical acceleration variance, and *C* is a constant. *A*, *B*, *C* are parameters calculated by the least square method, which is a form of mathematical optimization technology. It finds the best function match of the data by minimizing the sum of the square of the error; *fi* stands for the stride frequency, indicating how fast the pedestrian walks; *var* represents the acceleration variance during walking, describing whether the step is heavy or light: it can be calculated by Equation (4). These two factors indicate the pedestrian's step length indirectly. *SL* is the short name of Step Length. In the positioning of the same walking person, this model is frequently applied in a precise manner [22–24]. However, when the application extends to different people, a lack of consideration about differences between individuals deteriorates the accuracy of the step length measurement.

#### *3.2. The Step Length Measurement Method Based on Height, Stride Frequency and Acceleration Variance*

For pedestrians with different heights, it is found that the step length is also different even when *fi* and *vari* are the same. According to the method mentioned in Section 3.1, when applied to different people, calibration is once again required for each individual to improve measurement accuracy, which limits its application. According to the kinetics of the human body, the step length is proportional to the leg length as well as the body height under normal circumstances.

By analyzing the research and methods described in Section 3.1, a novel step length measurement model based on height, stride frequency, and acceleration variance is proposed:

$$SL\_i = h \cdot (A \cdot f\_i + B \cdot var\_i + C) + D \tag{2}$$

where *i* represents the *i*th step during walking; *h*, *fi* and *vari* stand for height, stride frequency, and vertical acceleration variance during the *i*th step, respectively. *A*, *B*, *C*, *D* are the corresponding model coefficients.

Inputs of this method are height *h*, stride frequency *fi*, and acceleration variance *vari*. *h* is a fixed constant and remains unchanged during different walking processes of the same individual. *fi* can be obtained by Equation (3):

$$f\_i = \frac{1}{t\_i - t\_{i-1}} \tag{3}$$

where *ti* and *ti*−<sup>1</sup> represent the corresponding moments when detecting two adjacent steps [25]. In addition, the vertical acceleration variance of each step during walking can be calculated by Equation (4):

$$var\_i = \frac{1}{N - 1} \sum\_{t=t\_{i-1}}^{t\_i} (a\_t - \overline{a}\_i)^2 \tag{4}$$

where *at*, *ai*, *N* stand for the acceleration at the moment of *t*, the average acceleration, and the number of sampling points within one step, respectively.

According to Equations (2)–(4), the step length measurement is realized. In Figure 3, the flow chart of the step length measurement based on the low-cost MEMS inertial system is presented.

As shown in Figure 4, before calculating step length, errors of inertial sensors and parameters of the step length model are calibrated by the method presented in the reference paper [26]. Once the height of an individual is entered, the height-adaptive step length is calculated according to the program. Once calculations begin, data is read, and then the number of steps is detected. For every single step, the sensors calculate stride frequency *fi*, acceleration variance *vari*, and the step length *SLi*. *SLi* serves as the input to the practical application.

**Figure 4.** The flow chart of the step length measurement.

#### **4. Experiment Research and Analysis**

In order to verify the proposed method in this paper, the following experiments have been conducted. The experiments consist of two parts:


As a consequence of the experimental complexity and the unavailability of a high-speed synchronous camera shooting, the step length was measured, and its accuracy was verified indirectly, by walking along one fixed route several times over.

#### *4.1. Experimental Equipment*

The experimental equipment consisted of a signal acquisition and transmission module, a laptop, and a wearable device. As shown in Figure 5, the acceleration output of MPU6050 is acquired by the microprocessor STM32 in the signal acquisition and transmission module, which is then sent to the laptop through a serial interface. Once the raw data is received, the step length is calculated, and the DR navigation is conducted. DR is the abbreviation of "Dead reckoning". Dead reckoning algorithm uses inertial navigation algorithm to predict motion position MPU-6050 is a MEMS inertial sensor with a 3-axis accelerometer and a 3-axis gyroscope, the advantages of which are its small size, its low cost, and its high precision. MEMS (Micro-Electro-Mechanical System) is also called micro-electromechanical system, micro-system, micro-mechanics, etc. It refers to high-tech devices with a size of a few millimeters or even less. The accelerometer features a measurement range of ±2 g, a zero bias of 50 mg, and random error of 0.4 mg. The random error of the gyroscope is due to the random variation of gyroscope output, which changes with time. It is expressed by the mean square error of the output data during idle state. As for the gyroscope, the measurement range is ±1000◦/s, the zero bias is 0.4◦/s, and the random error is 0.05◦/s. The wearable device is tied up at the waist, as Figure 6 demonstrates.

**Figure 6.** The wearing method.

#### *4.2. Calibration of Experiment Parameters for the Step Length Model*

Before measuring step length using Equation (2), the parameters *A*, *B*, *C*, *D* need to be calibrated. The height was entered and the result was calculated. For more accurate calibrated parameters, the subjects of different heights such as 1.60 m, 1.63 m, 1.71 m, 1.78 m and 1.83 m were selected to walk along one fixed route at different speeds. As shown in Figure 7, each subject was asked to walk a certain distance (24 m) along a flat road at a slow speed, at a preferred speed and at a fast speed. Each trajectory was conducted twice. In total, there are 30 sets of data in Table 1. The walking speed during each set of the experiment was made to be as stable and as consistent as possible. In order to reduce random errors,

the average step length, the stride frequency and the vertical acceleration variance for one step were measured and calculated. Besides the walking data, the subject's height is also added, thus enhancing the experimental data to include different heights, different step frequencies, different variances, and different step lengths. Finally, the model parameters of the step length measurement proposed in this

paper were calibrated by the least square method. The variable was written as *a* = ⎡ ⎢ ⎣ 1 *f*<sup>1</sup> *var*<sup>1</sup> *h*<sup>1</sup> 1 *f*<sup>2</sup> *var*<sup>2</sup> *h*<sup>2</sup> 1 *f*<sup>3</sup> *var*<sup>3</sup> *h*<sup>3</sup> ⎤ ⎥ ⎦,

the coefficient as *b* = ⎡ ⎢ ⎢ ⎢ ⎣ *A B C D* ⎤ ⎥ ⎥ ⎥ ⎦ , and the *SL* as *SL* = ⎡ ⎢ ⎣ *SL*<sup>1</sup> *SL*<sup>2</sup> *SL*<sup>3</sup> ⎤ ⎥ ⎦ . Then the formula was minimized to


The step length model in Section 3.1 has nothing to do with height. For a better comparison and analysis, the experimental data of the subject with the height of 1.71 m was calibrated.


**Table 1.** The data for parameter calibration.

**Figure 7.** The parameter calibration experiment.

#### *4.3. Walking Experiments*

During walking, the pedestrian's step length is not technically consistent for each step, leading to the unobtainability of precise measurements. Consequently, walking experiments of a certain distance are performed to verify the feasibility and accuracy of the step length measurement method proposed in this paper. Three other subjects were chosen to walk along the standard track in the playground of Beihang University three times over, from which the mean value was calculated. The experiment was conducted as follows: (1) the signal acquisition and transmission module was attached to the subject's waist with a belt; (2) the walking experiment was conducted and the motion data was acquired during walking; (3) the step length was measured by both the method explored in this paper and the method based on stride frequency and variance separately; (4) the experiment data was analyzed, and the results of the two different methods were compared.

The step length during walking was summed up to obtain the total walking distance, which was compared with the actual path length. The number of steps, the mean step length, the total walking distance, and the error rate are listed in Table 2.

The actual path length is the length of the track, namely, 400 m, but the length of the track of a few groups ended up being 453 m, because there were students taking physical education class when we were conducting our experiment. Therefore, the subjects were walking on the playground's outer ring, the actual path length of the track thus being 453 m. The walking distance was calculated based on the estimated step length in real time.

Experimental results indicate that the precision of the step length measurement of the method proposed in this paper is superior to that of the method based on stride frequency and variance. By comparing the average error and the standard deviation, we can conclude that the method in this paper can be used for subjects with different heights, the error is more constant than the method based on the frequency and acceleration variance during walking. More specifically, when adopting the step length measurement method based on stride frequency and variance, the user has to be the same person or at least someone with similar physical characteristics. A calibration of parameters had to be conducted again to achieve favorable results for different users, thus restricting its wide application. In this paper, the body height is added to the proposed step length measurement model. Despite the different heights of users, the step length measurement accuracy is relatively high.


**Table 2.** The results of the walking experiments.

*Sensors* **2018** , *18*, 1039

#### **5. Conclusions**

This paper proposes a height-adaptive step length measurement method based on the low-cost MEMS inertial system. Taking the height, the stride frequency, and the vertical acceleration variance into account, the step length was estimated with the motion data measured by the output of the accelerometer worn at the pedestrian's waist. Without any parameter calibration, this method is highly height-adaptive, that is to say, different users just need to input different heights for the step length to be properly measured. In addition, a series of walking experiments were performed, the results of which prove that this method can measure the step length accurately, giving rise to a great application prospect in fields such as auxiliary medical treatment, exercise rehabilitation, and more.

**Acknowledgments:** This work is supported by the National Natural Science Foundation of China under Grant 61473019, the Beijing Natural Science Foundation under Grant 4172036, the Beijing Science and Technology Plan under Grant D171100006217003, the Key Research and Development Project of the Ministry of Science, and Technology of China under Grant 2016YFB051600.

**Author Contributions:** Yanshun Zhang conceived and designed the experiments and contributed materials. Yingyue Li performed the experiments; Chuang Peng wrote the paper. Yingyue Li, Dong Mou, Ming Li and Wei Wang analyzed the data.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Motor Subtypes of Parkinson's Disease Can Be Identified by Frequency Component of Postural Stability**

### **Saba Rezvanian 1, Thurmon Lockhart 1,\*, Christopher Frames 1,2, Rahul Soangra <sup>3</sup> and Abraham Lieberman <sup>2</sup>**


Received: 26 February 2018; Accepted: 4 April 2018; Published: 5 April 2018

**Abstract:** Parkinson's disease (PD) can be divided into two subtypes based on clinical features—namely tremor dominant (TD) and postural instability and gait difficulty (PIGD). This categorization is important at the early stage of PD, since identifying the subtypes can help to predict the clinical progression of the disease. Accordingly, correctly diagnosing subtypes is critical in initiating appropriate early interventions and tracking the progression of the disease. However, as the disease progresses, it becomes increasingly difficult to further distinguish those attributes that are relevant to the subtypes. In this study, we investigated whether a method using the standing center of pressure (COP) time series data can separate two subtypes of PD by looking at the frequency component of COP (i.e., COP position and speed). Thirty-six participants diagnosed with PD were evaluated, with their bare feet on the force platform, and were instructed to stand upright with their arms by their sides for 20 s (with their eyes open and closed), which is consistent with the traditional COP measures. Fast Fourier transform (FFT) and wavelet transform (WT) were performed to distinguish between the motor subtypes using the COP measures. The TD group exhibited larger amplitudes at the frequency range of 3–7 Hz when compared to the PIGD group. Both the FFT and WT methods were able to differentiate the subtypes. COP time series information can be used to differentiate between the two motor subtypes of PD, using the frequency component of postural stability.

**Keywords:** Parkinson's disease (PD); tremor dominant (TD); postural instability and gait difficulty (PIGD); center of pressure (COP); fast Fourier transform (FFT); wavelet transform (WT)

#### **1. Introduction**

In 2010, approximately 630,000 people in the U.S. were diagnosed with Parkinson's disease (PD)—a number that is estimated to double by 2040 [1]. PD is a progressive neurodegenerative disorder that includes motor and non-motor features [2]. PD can be further divided into two subtypes based on clinical features—namely tremor dominant (TD) and postural instability and gait difficulty (PIGD) [2–5]. This categorization is important at the early stage of PD, since identifying the PD subtypes can help to predict the clinical progression of the disease. Several studies have confirmed that the PIGD subtype has a faster disease progression and greater motor function impairment [6], and is less responsive to levodopa and deep brain stimulation when compared to the TD subtype [3,5,7,8]. It has also been reported that there is a correlation between the freezing of gait (FOG) score and the

PIGD score [7]. Additionally, the PIGD subtype can place PD patients at a higher risk of falling when compared to TD patients [9]. It has been shown that PIGD patients have worse postural control when compared to TD patients [9,10]. Accordingly, correctly diagnosing subtypes can help caregivers to initiate early amenable interventions and track the progression of the disease. It should be noted that the diagnosis would not lead to a different medical treatment. However, another treatment needs to be taken alongside the medical treatment for PIGD patients in order to reduce the loss of balance and falling, since dopaminergic medications may result in limited improvement in postural instability and gait [11,12]. Thus, the diagnosis leads to the specific path that should be taken for the patient to manage the symptoms.

The differentiation of TD from PIGD is currently based on sub-scores of the Unified Parkinson's Disease Rating Scale (UPDRS) [3,11]. The UPDRS is scored by clinicians, and thus is subjective and prone to error [12]. Subtype-specific biomarkers may improve the accuracy of the diagnoses that are relevant to the PD subtypes and progression.

The center of pressure (COP) measure is widely employed in assessing postural control, and has been utilized for analyzing the disease-related features in PD patients [13–16]. The results of different studies have indicated that COP was more variable for PD patients, relative to the control participants [14,15], and that COP-derived velocities were abnormally large for PD patients with FOG when compared to the patients without FOG [13]. Thus, COP is considered as a good measure for representing PD disease-related postural characteristics.

PD tremor is present while resting, and is typically dampened with kinetic movement. Therefore, in order to distinguish between the two subtypes, proposing a static test appears to be more appropriate than a dynamic task [17]. Several studies have reported a frequency range of 3–7 Hz for PD tremor [17–19]. It has also been demonstrated that the whole-body COP signal has a frequency lower than 2 Hz [20–22]. Subtype-specific postural instability in PD may be better identified by the frequencies that make up the COP signal. We hypothesized that the whole-body COP frequency may be a better and more objective means of identifying the PD subtypes. The most common method to investigate the tremor in PD is fast Fourier transformation (FFT) [5,6]. FFT transfers a signal from the time domain to the frequency domain. In this method, the time information is lost after the transformation. Therefore, a method such as wavelet transformation (WT)—which includes both the time and frequency information of the signal [7]—might help to diagnose the subtypes better than FFT. Based on the importance of correct PD subtype diagnosis and the lack of an objective method among the current diagnosis techniques, this study aims to develop an objective method to diagnose PD motor subtypes by employing COP data and using the FFT and WT methods.

#### **2. Materials and Methods**

#### *2.1. Participants*

Thirty-six participants that were diagnosed with PD by specialists at the Muhammad Ali Parkinson Center at the Barrow Neurological Institute (Phoenix, AZ, USA) were recruited for this study. The participants' demographic information is presented in Table 1. The Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) was employed to identify the TD and PIGD groups [23]. The designated items for TD (kinetic and postural tremor in both the right and left hand; tremor—while at rest—of either the face and lips or the chain, arms, and legs) and PIGD (freezing, walking, posture, gait, and postural stability) were used to calculate the mean TD and PIGD scores. The ratio of the mean TD score to the mean PIGD score was used to identify the TD group. The patients with a ratio greater than or equal to 1.5 were classified as TD, while those with a ratio less than or equal to 1.0 were classified as PIGD. The patients with ratios ranging from 1.0 to 1.5 were classified as mixed-type, and were considered as an exclusionary criterion for this study [5,11,24].


**Table 1.** Demographics of the tremor dominant (TD) and postural instability and gait difficulty (PIGD) groups (mean ± standard deviation—SD). MDS-UPDRS: Movement Disorder Society Unified Parkinson's Disease Rating Scale.

The study was approved by the Institutional Review Board at the Barrow Neurological Institute and Arizona State University, Tempe, AZ, USA. The participants provided informed consent prior to their inclusion in the study. All of the assessments were performed while subjects were in the "on" medication status—approximately 1 to 1.5 h after taking the PD medication.

#### *2.2. Experimental Procedure*

The participants were placed with their bare feet on the force platform and were instructed to stand upright with their feet shoulder width apart and their arms by their sides for 20 s, and look straight ahead during the experiment. They were instructed not to talk or bend their knees throughout the experimental trials. Harnesses were fitted onto the participants to avoid falls. This task was performed under two conditions—namely eyes open and eyes closed. For the eyes closed condition, the subjects were asked to close their eyes during the experiment. Each participant performed the experiment under both conditions. Each condition had three trials.

#### *2.3. Data Analysis*

COP data were derived using force plate data sampled at 100 Hz. Both anterior–posterior (AP) and medial–lateral (ML) COP data were low-pass-filtered using a fourth-order, zero lag Butterworth filter with a cut-off frequency of 10 Hz. Five traditional COP measures were calculated to assess whether or not the two subtypes of PD can be distinguished by using the time domain information. The measures included the following: COP range (the range of COP displacement), resultant COP path length (the total COP trajectory length), resultant mean velocity (the resultant path length divided by the total duration), and a 95% confidence ellipse area (the smallest ellipse that will cover 95% of the points of the COP diagram). Based on previous studies, these traditional parameters are good indicators of postural instability [14,25,26] and were considered as variables that might help us to distinguish PIGD from TD. All of the analyses were performed in MATLAB version 2015a.

#### *2.4. TD vs. PIGD Detection Method*

In order to distinguish between the TD and PIGD subtypes, the following two methods were utilized: fast Fourier transform (FFT) and wavelet transform (WT). In the FFT method, the PD subtypes were identified by the frequency spectra of COP signals. Two frequency bands were introduced [27–29]: the COP band and the tremor band. The COP and tremor bands were defined as the frequency components from 0–3 Hz to 3–7 Hz, respectively. The detection method was defined as the ratio of the area under the power spectra of the tremor band to the summation of the areas under the power spectra of the COP band and the tremor band.

COP data were transformed into the wavelet domain using daubechies mother wavelet (db6). It was chosen because it has been widely employed in different human posture and movement studies [27,30]. Various mother wavelets were also applied to ensure that the optimal selection was made appropriately. The results supported the notion that daubechies mother wavelet was the best choice. In the WT method, the COP and tremor bands were defined as the scales that corresponded to the frequency ranges of 0–3 Hz and 3–7 Hz, respectively. The detection method was defined in

a similar manner to the way it was defined in the FFT method: the ratio of the averaged WT coefficients of the tremor band to the summation of the averaged WT coefficients of the COP band and the tremor band. This ratio was unitless because it was a ratio of values with the same unit. In both methods, the defined ratio was multiplied by 100 in order to obtain a value between 0 and 100. Values that were closer to 100 indicated a higher possibility of the TD subtype, while the possibility of the PIGD subtype increased as the values approached 0. The first time derivative of COP time series was defined as COP velocity (V-COP). The ratio that was defined above was applied to COP (RCOP) and COP velocity (RVCOP) in both the AP and ML directions.

#### *2.5. Statistical Analysis*

Analysis of variance (ANOVA) with repeated measures on the traditional COP measures and the proposed detection ratio (using both the FFT and WT methods) were performed. Different factors—such as condition (two levels: eyes open (EO) and eyes closed (EC)) and group (or subtype) of PD (two levels: TD and PIGD)—were considered as within-subject and between-subject factors, respectively. Comparisons of interest exhibiting statistically significant differences (*p* < 0.05) were further analyzed using post hoc tests with Bonferroni corrections. In all analyses, sphericity assumptions were tested (Greenhouse–Geisser analysis). The diagnostic performance of the proposed method—or the accuracy of a test to discriminate between the subtypes—was further evaluated using receiver operating characteristic (ROC) curve analysis [31] for the directions and factors of both methods. In a ROC curve, the true positive rate (sensitivity) is plotted as a function of the false positive rate (100—specificity) at different cut-off points. Therefore, each point on the ROC curve corresponds to a sensitivity/specificity pair for a particular decision threshold. Therefore, the upper-left corner denotes a test with perfect discrimination (no overlap in the two distributions) in a ROC curve analysis. Accordingly, the closer the ROC curve is to the upper-left corner, the higher the overall accuracy of the test [31]. In this study, PD subtypes were diagnosed by utilizing UPDRS and were considered as a correct diagnosis. All of the statistical analyses were performed based on this assumption. In all tests, *p* < 0.05 was considered as a significant level. Statistical analyses were performed using IBM SPSS Statistics 22.

#### **3. Results**

The results of the traditional COP measures—under both the eyes open and eyes closed conditions—are provided in Table 2. All of the variables had larger values in the eyes closed condition compared to the eyes open condition. Because these parameters did not have a normal distribution, a Box-Cox transformation was applied and parametric methods were performed. There was no significant difference between the two groups for all the variables. However, there was a significant difference between the conditions for all of the parameters (range AP: F(1,34) = 4.252, *p* = 0.047; range ML: F(1,34) = 60.34, *p* = 0.001; path length: F(1,34) = 29.797, *p* = 0.001; mean velocity: F(1,34) = 29.795, *p* = 0.001; area: F(1,34) = 11.847, *p* = 0.002).

**Table 2.** Selected postural stability parameters. Range anterior–posterior (AP): center of pressure (COP) range in the AP direction, range medial–lateral (ML): COP range in the ML direction, path length: resultant COP path length, mean velocity: resultant COP mean velocity, and area: 95% ellipse area. The symbols \* or \*\* denote which of the two variables were significantly different at each parameter (*p* < 0.05).


A power spectral analysis of the COP and COP velocity of a TD patient and a PIGD patient are plotted in Figure 1, revealing that both patients had frequency components ranging from 0 to 2 Hz in their COP and COP velocity signals. However, only the TD patient had an increase in power spectrum in the frequency band of 3–7 Hz. This increase was larger in the ML direction.

**Figure 1.** Power spectrum of COP and COP velocity of a tremor dominant (TD) patient and a postural instability and gait difficulty (PIGD) patient for both the medial–lateral (ML) and anterior–posterior (AP) directions. The graphs on the left and right sides of the page present the power spectrum signal of a TD patient and a PIGD patient, respectively. COPML: COP in the ML direction, COPAP: COP in the AP direction, V-COPML: COP velocity in the ML direction, and V-COPAP: COP velocity in the AP direction.

The WT of COP and COP velocity of a TD patient and a PIGD patient in both the ML and AP directions are plotted in Figure 2. The horizontal white lines in each figure indicate the PD tremor scale range corresponding to the frequency range of 3–7 Hz. The WT coefficients in Figure 2 display relatively larger values in the PD tremor scale range (i.e., lighter blue values appeared in between two horizontal white lines) for the TD patient when compared to the PIGD patient. Similar to the power spectral analysis (Figure 1), these increases were larger in the ML direction.

The results of the proposed detection ratio for COP and its velocity in both directions using FFT are presented in Figure 3. Neither the ratio of COP (RCOP\_ML) nor its velocity (RVCOP\_ML) in the ML direction were significantly different across the different conditions (RCOP\_ML: F(1,34) = 2.006, *p* = 0.112; RVCOP\_ML: F(1,34) = 2.67, *p* = 0.112). However, a statistically significant difference in RVCOP\_ML across the groups (F(1,34) = 7.978, *p* = 0.008) was observed, although no significant difference was found in RCOP\_ML (F(1,34) = 3.449, *p* = 0.072). In both RCOP\_ML and RVCOP\_ML, there was no significant interaction between the condition and the group (RCOP\_ML: F(1,34) = 1.181, *p* = 0.285; RVCOP\_ML: F(1,34) = 2.037, *p* = 0.163). RVCOP\_ML was larger for the TD group than for the PIGD group (Figure 3A,B). This indicated that there were larger amplitudes in the frequency range of 3–7 Hz in this group. In the AP direction, there was no significant difference across the groups (RCOP\_AP: F(1,34) = 0.498, *p* = 0.485; RVCOP\_AP: F(1,34) = 0.628, *p* = 0.433) and the conditions (RCOP\_AP: F(1,34) = 1.306, *p* = 0.201; RVCOP\_AP: F(1,34) = 3.45, *p* = 0.08) in both RCOP\_AP and RVCOP\_AP.

The explained WT method was applied to COP and its velocity in both directions. The results are presented in Figure 4. We found a significant difference between the groups for RCOP\_ML and RVCOP\_ML (RCOP\_ML: F(1,34) = 7.589, *p* = 0.009; RVCOP\_ML: F(1,34) = 10.066, *p* = 0.003), but no significant difference between the conditions (RCOP\_ML: F(1,34) = 0.373, *p* = 0.814; RVCOP\_ML: F(1,34) = 2.5, *p* = 0.123). There was no significant interaction between the conditions and the groups (RCOP\_ML: F(1,34) = 3.044, *p* = 0.09; RVCOP\_ML: F(1,34) = 2.828, *p* = 0.102). Both RCOP\_ML and RVCOP\_ML had larger values for the TD group than for the PIGD group (Figure 3A,B). These increases occurred because of the larger amplitude values in the scales corresponding to the frequency range of 3–7 Hz. In the AP direction, there were no significant differences across the groups (RCOP\_AP: F(1,34) = 0.004, *p* = 0.952; RVCOP\_AP: F(1,34) = 0.854, *p* = 0.362) or conditions (RCOP\_AP: F(1,34) = 0.011, *p* = 0.916; RVCOP\_AP: F(1,34) = 3.047, *p* = 0.091) in both RCOP\_AP and RVCOP\_AP.

**Figure 2.** Wavelet transform (WT) of COP and COP velocity of a TD patient and a PIGD patient for both the ML and AP directions. The horizontal white lines in each plot indicate the PD tremor scale range corresponding to the frequency range of 3–7 Hz. The frequencies of 3 Hz and 7 Hz correspond to the scales of 24 and 10, respectively. COPML: COP in the ML direction, COPAP: COP in the AP direction, V-COPML: COP velocity in the ML direction, and V-COPAP: COP velocity in the AP direction.

**Figure 3.** Fast Fourier transform (FFT) results of the proposed detection ratio for COP and its velocity in both the ML and AP directions. (**A**) RCOP\_ML: the detection ratio using COP data in the ML direction, (**B**) RVCOP\_ML: the detection ratio using COP velocity data in the ML direction, (**C**) RCOP\_AP: the detection ratio using COP data in the AP direction, and (**D**) RVCOP\_AP: the detection ratio using COP velocity data in the AP direction. The asterisks (\*) placed over the vertical bars denote a significant difference (*p* < 0.05). EC: eyes closed condition; EO: eyes open condition.

**Figure 4.** WT results of the proposed detection ratio for COP and its velocity in both the ML and AP directions. (**A**) RCOP\_ML: the detection ratio using COP data in the ML direction, (**B**) RVCOP\_ML: the detection ratio using COP velocity data in the ML direction, (**C**) RCOP\_AP: the detection ratio using COP data in the AP direction, and (**D**) RVCOP\_AP: the detection ratio using COP velocity data in the AP direction. The asterisks (\*) placed over the vertical bars denote a significant difference (*p* < 0.05).

The ROC curves of the proposed detection ratio for COP and its velocity in both directions and under both conditions are plotted in Figures 5 and 6 for the FFT and WT methods, respectively. In both methods, the ROC curves were closer to the upper-left corner in the ML direction than they were in the AP direction, which indicated a higher overall accuracy of the test in the ML direction [31].

**Figure 5.** Receiver operating characteristic (ROC) curves of the proposed detection ratio using the FFT method for COP and its velocity: (**A**) ML direction and (**B**) AP direction. EO-RCOP\_ML: the detection ratio using COP data in the ML direction under the eyes open condition, EC-RCOP\_ML: the detection ratio using COP data in the ML direction under the eyes closed condition, EO-RVCOP\_ML: the detection ratio using COP velocity data in the ML direction under the eyes open condition, EC-RVCOP\_ML: the detection ratio using COP velocity data in the ML direction under the eyes closed condition, EO-RCOP\_AP: the detection ratio using COP data in the AP direction under the eyes open condition, EC-RCOP\_AP: the detection ratio using COP data in the AP direction under the eyes closed condition, EO-RVCOP\_AP: the detection ratio using COP velocity data in the AP direction under the eyes open condition, and EC-RVCOP\_AP: the detection ratio using COP velocity data in the AP direction under the eyes closed condition.

**Figure 6.** ROC curves of the proposed detection ratio using the WT method for COP and its velocity: (**A**) ML direction and (**B**) AP direction. EO-RCOP\_ML: the detection ratio using COP data in the ML direction under the eyes open condition, EC-RCOP\_ML: the detection ratio using COP data in the ML direction under the eyes closed condition, EO-RVCOP\_ML: the detection ratio using COP velocity data in the ML direction under the eyes open condition, EC-RVCOP\_ML: the detection ratio using COP velocity data in the ML direction under the eyes closed condition, EO-RCOP\_AP: the detection ratio using COP data in the AP direction under the eyes open condition, EC-RCOP\_AP: the detection ratio using COP data in the AP direction under the eyes closed condition, EO-RVCOP\_AP: the detection ratio using COP velocity data in the AP direction under the eyes open condition, and EC-RVCOP\_AP: the detection ratio using COP velocity data in the AP direction under the eyes closed condition.

The ROC curves were further analyzed by calculating the areas under each curve. The results are presented in Table 3. Only COP velocity data in the ML direction could significantly distinguish between the two subtypes using the FFT method. The results of the area under the ROC curves also revealed that the WT method could significantly distinguish between the two subtypes by using either COP or COP velocity data in the ML direction, regardless of the conditions.

**Table 3.** The area under the receiver operating characteristic (ROC) curves of the proposed detection ratio, using both the FFT and WT methods, for COP and its velocity in the ML and AP directions under the two conditions (eyes open (EO) and eyes closed (EC)). The *p*-values of each parameter are presented in parentheses. The asterisks (\*) indicate that the area under the ROC curve was significantly different from 0.5 (*p* < 0.05).


#### **4. Discussion**

This study addressed subtype-specific biomarkers in order to classify the inherent heterogeneity of PD. This categorization can help to predict the clinical progression of the disease. Thus, the correct diagnosis of the subtypes can assist caregivers in initiating early amenable interventions and managing symptoms. The COP time series of PD patients were analyzed to distinguish between the two subtypes of PD. To the best of our knowledge, this study is the first to attempt to objectively diagnose the TD and PIGD subtypes of PD. Postural stability is maintained through neuromuscular feedback loops and open loop control processes that constantly adapt to internal and external perturbations [32,33]. Utilizing specific statistical and numerical tools, these control mechanisms can be quantified to identify neuromuscular changes that occur with pathology. Thus, traditional linear postural measures and Fourier transformation were applied to the COP time series and the increment of the COP time series in both the AP and ML directions. Furthermore, in order to quantify the changes in COP dynamics that occur at multiple timescales, a wavelet transform was employed to infer the underlying nature and control mechanisms involved in balance maintenance and the disease state.

In the traditional measures of postural sway, the parameters that denoted the magnitude of the postural movements were unable to discriminate between the TD and PIGD subtypes (Table 2). However, when visual information was occluded, a coincident decrease in postural stability was reflected in both subtypes for the linear postural measures (i.e., COP range, mean velocity, path length, and a 95% confidence ellipse area). These results were consistent with previous investigations regarding postural stability in PD patients [34].

Both the power spectral density and the WT of the COP time series and its velocity (Figures 1 and 2) revealed an increase in the 3–7 Hz frequency range of the TD group, a frequency spectra that is reportedly symptomatic of parkinsonian tremor [17–19]. In fact, the ML COP data exhibited a greater frequency content than the AP COP data, which was consistent with previous investigations, which reported that PD patients exhibited increased ML sway amplitude, decreased AP sway amplitude, and possibly postural inflexibility in the AP direction [15,35–37]. In this context, the preponderance of the ML frequency in the ML direction, coupled with the impaired movement in the AP direction, suggested an underlying postural inflexibility in PD patients, where the tremor reflected in the ML time domain might be a consequence of the AP direction's inability to contain movements in a higher frequency range [35,38,39]. Our proposed ratio was not able to show a statistically significant difference

between the TD and PIGD patients in the AP direction using either of the methods—even accounting for both COP and COP velocity. The reason was that the tremor frequency had a larger amplitude in the ML direction than it did in the AP direction (as shown in Figures 1 and 2). However, both the FFT and WT methods were able to distinguish the TD patients from the PIGD patients using the ML-COP velocity signal, while only the WT method was able to specify the subtype with the COP position time series. This could be explained by the fact that the FFT method used only the frequency information from the signals, while the WT method employed both the frequency and time components. The information from the signals that was utilized by WT enabled us to specify the subtypes of PD using both COP and COP velocity. Additionally, FFT displayed significant results when it employed COP velocity—as opposed to COP in itself—because the velocity of the signal was a first time derivative of the signal, which captured more variation of the signal. Hence, FFT could assess more information about the signals when it employed COP velocity. The results of the proposed method were consistent across both conditions (EO and EC) in both methods (FFT and WT). This consistency indicated the strength of the proposed diagnostic method using the proposed ratio. Although the proposed method can distinguish the TD from the PIGD subtypes, further studies are required to define the threshold value ranges that can classify the patients.

**Acknowledgments:** This research was supported by the NSF-Information and Intelligent Systems (IIS) and Smart and Connected Health (1065442, and 1547466, and secondary 1065262).

**Author Contributions:** Saba Rezvanian and Thurmon Lockhart conceived and designed the experiments; Saba Rezvanian and Christopher Frames performed the experiment; Saba Rezvanian analyzed the data and wrote the manuscript with support from Thurmon Lockhart, Abraham Lieberman and Rahul Soangra. All authors discussed the results and contributed to the final manuscript.

**Conflicts of Interest:** The authors declare no conflicts of interest.

**Ethical Statements:** All subjects gave their informed consent for inclusion before they participated in the study. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the St. Joseph's Hospital and Medical Center (PHX-16-0227-71-04).

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Analysis of a Smartphone-Based Architecture with Multiple Mobility Sensors for Fall Detection with Supervised Learning**

#### **José Antonio Santoyo-Ramón, Eduardo Casilari \* and José Manuel Cano-García**

Departamento de Tecnología Electrónica, Universidad de Málaga, ETSI Telecomunicación, 29071 Málaga, Spain; jasantoyo@uma.es (J.A.S.-R.); jcgarcia@uma.es (J.M.C.-G.)

**\*** Correspondence: ecasilari@uma.es

Received: 15 February 2018; Accepted: 4 April 2018; Published: 10 April 2018

**Abstract:** This paper describes a wearable Fall Detection System (FDS) based on a body-area network consisting of four nodes provided with inertial sensors and Bluetooth wireless interfaces. The signals captured by the nodes are sent to a smartphone which simultaneously acts as another sensing point. In contrast to many FDSs proposed by the literature (which only consider a single sensor), the multisensory nature of the prototype is utilized to investigate the impact of the number and the positions of the sensors on the effectiveness of the production of the fall detection decision. In particular, the study assesses the capability of four popular machine learning algorithms to discriminate the dynamics of the Activities of Daily Living (ADLs) and falls generated by a set of experimental subjects, when the combined use of the sensors located on different parts of the body is considered. Prior to this, the election of the statistics that optimize the characterization of the acceleration signals and the efficacy of the FDS is also investigated. As another important methodological novelty in this field, the statistical significance of all the results (an aspect which is usually neglected by other works) is validated by an analysis of variance (ANOVA).

**Keywords:** fall detection system; inertial sensors; smartphones; accelerometers; machine learning algorithms; supervised learning; ANOVA analysis

#### **1. Introduction**

According to some forecasts [1], it is expected that between 2015 and 2050 the over-60 years of age world population will grow from 900 to 2000 million. This dramatic demographic change will undoubtedly give rise to a series of challenges in the health systems that must be faced in order to sustain and increase the quality of life of the citizens. The present study focuses on one of the biggest public health problems faced by the world society: falls.

A fall can result in minor consequences such as bruises, lacerations or abrasions, but it can also lead to more serious and dangerous situations such as bone fractures or psychological disorders (such as the so-called Fear of Falling or FoF syndrome). The World Health Organization has reported that falls are the second worldwide cause of mortality provoked by accidental or unintentional injuries. 37.3 million falls requiring medical attention occur annually, while it has been estimated that around 646,000 people die every year due to falls [2] A remarkable fact is that people over 65 years old experience the greatest number of falls. Approximately 28–35% of people over 65 suffer a fall each year, whereas this percentage climbs with age, reaching 32–42% for people over 70 [3]. Risk groups are not limited to the older people. Other population groups are exposed to endure severe falls during their work or leisure time (cyclists, mountaineers, firemen, antenna installers, cable layers, etc.).

The probability of survival after a serious fall is strongly related to the rapidness of the medical assistance. As mentioned in [4], 50% of people affected by a bad fall who remain without assistance for more than an hour die before the six months following the accident. Consequently, the research on reliable and cost-effective systems for the automatic detection of falls has gained attention during the last decade.

#### **2. State of the Art on Wearable Fall Detection Systems and Multisensory Architectures**

Fall Detection Systems (FDS) are intended to permanently monitor the mobility of a target user or patient in order to automatically alert his/her relatives or the medical staff whenever a fall occurrence is identified. The goal of a FDS is to maximize the possibility of detecting actual falls while minimizing the number of false alarms, i.e., the conventional movements or Activities of Daily Living (ADLs) that are mistaken as falls.

FDSs can be generically classified into two large groups. On the one hand, in context-aware systems, the detection decision is based on the signals captured by a network of environmental sensors (microphones, cameras, vibration sensors, etc.) located in a specific area around the user to be monitored. These FDSs have clear drawbacks provoked by their high installation and maintenance costs, their limitation in terms of the area where the user can be monitored or their high probability of spurious interference caused by external factors such as changes in the light, noises, presence of another individual, displacement of furniture, domestic pets or other falling objects, etc. In addition, under a context-aware tracking solution, the user can feel their privacy invaded because of the continuous use of audiovisual equipment.

In contrast, wearable systems use sensors that are integrated into the patient's clothing. Hence, these sensors only track magnitudes unambiguously linked to the patient mobility, such as the acceleration or angular velocity of the body. Wearable systems offer some clear advantages over the context-aware solutions since the restrictions about the monitoring area are eliminated, as long as the sensors always accompany the patient. Besides, they are also less intrusive, more economical and less vulnerable to the effects of external factors. As disadvantages, wearable FDSs can adversely affect the patients' comfort during their daily routines (especially if they are too bulky). They can also be useless if the user forgets to wear it or to recharge the corresponding battery or if the FDS is unintentionally misplaced or not fit properly on the patient's body. In addition, in certain environments such as toilets or bathrooms or during certain activities (such as showers) the monitoring activity may be inviable.

Due to the fact that most present smartphones integrate the inertial sensors that are required by a wearable FDS (mainly an accelerometer but also a gyroscope and a magnetometer), smartphone technology has been proposed as a basis to deploy cost-effective FDSs [5,6]. The use of smartphones can benefit from the plummeting costs, increasing hardware capabilities and the widespread popularity of these personal devices. Although the access to a mobile phone is lower for older age groups, smartphone usage among people over 60 years old in the United States was about 81% in 2005, while 60% of the elderly above 60 declare to own a smartphone [7]. Moreover, these percentages tend to increase. As predicted in [8], almost the entire population over 60 years old in Western countries is expected to use cell phones in the next 10 years.

As a result of the easiness of developing a FDS on a smartphone, most works of the recent related literature on wearable FDS have focused on the study of 'smartphone-only' based detectors, i.e., architectures that employ the smartphone as the unique element in the system, simultaneously acting as a sensing unit, data processor node and communication gateway [9]. As smartphones are natively provided with cellular (3G/4G) communications, the patient can be tracked almost ubiquitously (at least in urban scenarios) and fall detection alarms can be seamlessly integrated with emergency response systems by means of a SMS, an automatic phone call to a predefined set of phone numbers, a notification through a remote server on the Internet, etc. The use of smartphones as mobility monitoring units have been even proposed to characterize the gait speed as a predictor of post-operative morbidity and mortality in elderly cardiovascular disease [10].

However, it has been have shown [11] that the position of the sensor that monitors the movements of the subject is crucial for the effectiveness of the FDS. In this regard, recommended locations such as the chest or waist are not comfortable positions to place a smartphone (which is conventionally transported in a hand-bag or in a loose pocket, where the phone may exhibit a certain freedom of movements that can affect the representativeness of the mobility measurements provided by its sensors). Furthermore, the range of the sensors integrated in the smartphone were not originally conceived to quantify the intensity of the movements that a fall can produce. Thus, the use of smaller and specific sensors that can be easily incorporated into the patient's clothing seems advisable for the sake of ergonomics. Nowadays there is a great variety of low-cost sensing motes that can be used for this purpose. These devices embed inertial sensors accelerometers as well as wireless communication interfaces (Bluetooth, 802.15.4 ...) that can be utilized to send the monitored signals to devices with a greater processing capacity such as a smartphone.

Multisensory Body Area Networks (BANs) have been proposed in works such as [12–14] to recognize and differentiate diverse activities of daily life. These networks are composed of a set of sensing nodes located on different parts of the body of the subjects under test. Nonetheless, in these works the network was never employed to identify falls and in most of them the importance of the position of the sensors to identify the activities was not specifically addressed as a research issue.

The work by Hyunh et al. in [15] (whose results are commented by the same authors in [16]) investigated the optimal position of a fall detection sensor, concluding that the sensor should be attached to the center of the chest.

An architecture with multiple sensors for fall detection has been recently investigated by Turan in the interesting work presented in [17]. This author analyzed the performance of a FDS built on a BAN consisting of six wireless motes placed at six positions (head, chest, waist, wrist, thigh and ankle). After recording a dataset of falls and movements of daily life executed by a group of experimental subjects, the effectiveness of different detection algorithms to discriminate falls is assessed taking into account the positions of the motes that are considered to produce the detection decision. In the deployed testbed, the network did not employ any smartphone (the signals are sent via ZigBee to a computer) while the ADLs emulated by the volunteers did not include activities (such as hopping or climbing stairs) that can be easily mistaken as falls. The statistical significance of the results was not evaluated either. In fact, except for a very reduced set of works (such as [18], in which a *p*-value analysis is carried out to examine if the election of the samples influences the efficacy of the detection method), the literature on FDSs disregards the analysis of the statistical significance of the metrics that are obtained to compare the performance of the detection algorithms.

The present study will extend a previous work [11] to systematically evaluate the application of different learning machine strategies when they are applied to the traces captured by a hybrid multisensory FDS architecture (consisting of a set of sensing motes and a smartphone). The study tries to optimize the election of the acceleration statistics that characterize the body mobility as well as to identify the sensor placement combinations that produce the best detection performance. In all the cases, the validity of the conclusions will be contrasted through an analysis of variance (ANOVA) of the results.

#### **3. Description of the Experimental Testbed**

In order to characterize the movements of a set of experimental subjects, we developed a monitoring system, sketched in Figure 1, based on a smartphone which is wirelessly connected to a set of four sensing nodes or 'motes', which are located on different parts of the body of the subject (chest, waist, wrist and ankle). The motes measure the acceleration (expressed in *g* –or g-force-), angular velocity (in ◦/s) and flux density of the magnetic field (in μT).

The sensing nodes were deployed on CC2650 SimpleLink™ Bluetooth low energy/Multi-standard SensorTag modules [19], manufactured by Texas Instruments. Every SensorTag incorporates a CC2650 ARM microcontroller and a set of MEMS sensors, including an InveSense MPU-9250 Inertial Measurement Unit (IMU) which integrates three triaxial sensors: an accelerometer, a gyroscope and a magnetometer. The SensorTag modules, which are powered through a CR2032-type battery, support different wireless communications, which avoids wiring and provides the user with a complete freedom of movement. In particular, the sensing motes embed a multi-standard 2.4 GHz ultra-low power wireless MCU compatible with ZigBee, 6LowPAN or Bluetooth Low Energy (BLE) communications. In our prototype, benefiting from the fact that most current smartphones support BLE, a smartphone was selected as the central node of a BLE piconet. The piconet follows a typical star topology. Thus, the smartphone, which plays the role of the piconet master, receives all the packets containing the measurements captured by four SensorTags, which act as slaves in the BLE piconet.

The best election of the sensor sampling rate is a still open issue in the area of fall detection system (see the works by Medrano in [20] and Fudickar in [21] for a further analysis of the importance of the sampling rate in the efficacy of a FDS). As a matter of fact, sampling frequencies ranging from 5 Hz to 256 Hz have been used by the related literature to generate benchmarking datasets to compare fall detection algorithms. Aiming at avoiding saturation problems in the Bluetooth network, the sampling rate of the sensors was set to 20 Hz. For that purpose, the original firmware of the SensorTags was modified, so that the 9 values periodically collected by the three IMU triaxial sensors could be transmitted every 50 ms through the BLE connection. In addition, as the smartphone in turn also integrates an IMU, we used the mobile phone, which was located in a trouser pocket for all the experiments, as a fifth sensor to describe the mobility registered on the subject's thigh. The employed sampling frequency of the smartphone measurement was 200 Hz.

**Figure 1.** Basic architecture of the testbed.

Besides, a specific Android application (app) was programmed to receive and save the information captured by all the sensors, so that a repository with the generated mobility traces could be progressively generated. The app is in charge of storing the measurements transmitted from the four SensorTags as well as those captured by the smartphone in a CSV (Comma Separated Value)-format log file. Every sample received from any of the mobility sensors and any source (the four SensorTags and the smartphone) is recorded in an independent line of the log file, together with a timestamp generated by the app. No further synchronization between the sensing nodes is needed as long as a default value of 30 ms is selected by Android OS for the polling BLE Connection Interval (the time between two consecutive transmissions from the same slave to the Bluetooth master). So, the communication delay that may be introduced by BLE technology can be considered negligible when compared to the observation window that is required to detect a fall event.

For each experiment the program creates a new log file that stores the signals obtained from the subject during the execution of a certain movement (ADL or fall) for a fixed period of time (15 s). The log file also includes the personal characteristics of the experimental subject (weight, age, gender, height) as well as the typology of the experiment performed movement. In the log files, whenever

a sample is received from a certain sensor, the app adds a timestamp and the MAC address of the mote (so that the sensor that has sent the information can be identified). Finally, in every log file, the app includes the employed sampling rate and additional information of the sensing device (manufacturer, range, etc.).

The final resulting dataset (called UMAFall) containing all the log files has been made publicly available in Internet [22] (together with some video clips that illustrate the activity of the subjects) as a benchmarking tool for the research on fall detection systems.

With regard to energy consumption, the batteries of the sensors allowed generating the dataset without being replaced. Similarly, the battery of the smartphone supported experimental one-day sessions of continuous use without having to be recharged.

Battery-life is obviously a key issue for the feasibility of any wearable monitoring system. However, the detailed study of the consumption of the implemented architecture is out of the analysis of this work (which focuses on evaluating the performance of the detection algorithms). Refer to [23–25] for a further study of the challenges caused by mobile sensing to battery-life in wearables and smart personal devices.

As it respects to the employed methodology to validate the detection strategies, due to the obvious difficulties of utilizing measurements from real-world falls, we followed the procedure that is commonly observed by the related literature. Thus, a group of 19 experimental subjects, whose basic characteristics are presented in Table 1, systematically performed a series of predefined movements, including 11 types of Activities of Daily Life (ADL) and 3 types of emulated falls, which were executed on a mattress to avoid injuries. Table 2 in Section 5 includes a description of each type of movement as well as the number of samples collected from the subjects. For the selection of the typology of ADLs and falls in the testbed we considered the following criteria:



**Table 1.** Personal characteristics of the participants in the testbed.


**Table 2.** Typology and number of the movements executed by the experimental subjects during the testbed. The table also indicates the number of samples utilized for both training and testing phases during the validation of the machine learning algorithms.

The divergence in the mobility patterns of actual falls suffered by older people and falls mimicked by young and healthy adults on a cushioning surface is still under discussion (see, for example, the studies in [27–29]) but that issue is beyond the scope of this paper. In any case, as it can be noticed from Table 1, two experimental subjects older than 65 participated in our testbed although they did not perform falling movements for safety reasons.

In all the trials, the starting position of the subject performing the movement (ADL or fall) was standing with hands on hips. All the experiments took place in a home environment. As it can be contemplated in Figure 2, the motes were tightly attached to the subject's body by means of elastic bands. For all the tests, the accelerometer axes of the SensorTag modules and the Smartphone were oriented and aligned as depicted in Figure 3. Thus, the components of the measurements of the tri-axial sensors are always associated to the same actual direction of the body position.

The log files obtained (one for each experiment) were formatted and stored in the internal memory of the smartphone. After the experiments were finished, the data were extracted and moved to a PC where they are processed in an offline way with Matlab scripts that implement the different machine learning algorithms to be tested.

**Figure 2.** Example of the distribution of the sensors worn by an experimental subject. The green arrow indicates the position of the smartphone. Red arrows correspond to the SensorTag modules, which are attached to the user's body by means of elastic bands.

**Figure 3.** Representation of the spatial reference system of the employed sensing devices (devices are firmly attached to the subjects' body to guarantee that the reference system does not change during the experiments). (**a**) SensorTag (**b**) Smartphone.

#### **4. Machine Learning Algorithms and Selection of the Input Features**

The data collected through the developed system are processed and subsequently utilized to study the behavior of four supervised learning classification algorithms. Such algorithms (whose operational procedure is sketched in Figure 4) need to be trained before being applied to the test data. This training is achieved by providing the algorithm with a series of features (statistics that characterize the movements and which are computed from the mobility traces) that are extracted from a set of training samples as well as with the decisions that the detector should make (ADL or fall) for every sample. From the training phase, the algorithm builds a mathematical model that is later employed with the test data to decide the data 'class', that is to say, to discriminate falls from ADLs [30].

**Figure 4.** Typical Flow chart of a supervised learning classification algorithm.

#### *4.1. Feature Extraction: Selection of the Input Statistics of the Machine Learning Algorithms*

In the development of any machine learning algorithm, a proper selection of the input characteristics is a key aspect.

As in most works in the literature, the statistical characterization of the movements will be based on the measurements (*AXi* , *AYi* , *AZi* ) of the acceleration captured by the tri-axial accelerometers embedded in the five sensing points: chest, wrist, waist, ankle and thigh (pocket). Future studies should contemplate the information provided by the gyroscope and the magnetometer.

The statistics to be studied as the input of the supervised algorithms are derived from the samples of 15 s generated through the experimental testbed. Falls are typically associated to a brusque decay (which tends to 0 g) of the acceleration components (during an initial free-fall period) followed by one or several acceleration peaks caused by the impact (or impact) against the floor [31]. Thus, we focus our analysis of the signal on the period of the signal where the difference between the 'peaks' and the

valleys' in the acceleration is maximized. In particular, to analyze the acceleration measurements in every axis (*x*, *y* or *z*), we utilize a sliding window of duration *tW* = 0.5 s or *NW* samples, where:

$$N\_W = t\_W \cdot f\_s \tag{1}$$

being *fs* the sampling rate of the sensors (20 Hz for the SensorTags and 200 Hz for the smartphone). This value of 0.5 s is coherent with other related studies on context aware and wearable systems for fall detection, which also employ a sliding window to detect the falls: 0.4 s [32], 0.6 s [33], 0.75 s [34], 1 s in [35,36] or 2 s in [17,37,38]. In [39] authors claim that an observation interval of 0.5 is the best compromise between effectiveness, complexity and low power consumption to track the acceleration measurements in a fall detection system.

In order to locate the interval where the acceleration components suffer the highest variation, we compute for each window the module of the maximum variation of the acceleration components in the three axes. For the *j*-th window, this parameter (*Awdi f f*(*j*)) is calculated as:

$$A\_{w\_{diff}}(j) = \sqrt{\left(A\_{X\_{\text{max}\_j}} - A\_{X\_{\text{min}\_j}}\right)^2 + \left(A\_{Y\_{\text{max}\_j}} - A\_{Y\_{\text{min}\_j}}\right)^2 + \left(A\_{Z\_{\text{max}\_j}} - A\_{Z\_{\text{min}\_j}}\right)^2} \tag{2}$$

where *AX*max*<sup>j</sup>* , *AY*max*<sup>j</sup>* and *AZ*max*<sup>j</sup>* indicate the maximum values of the acceleration components in the *x*, *<sup>y</sup>* and *z-* axes during the *<sup>j</sup>*-th sliding window (e.g., *AX*max*<sup>j</sup>* = max(*AXi* ) ∀*i* ∈ [*j*, *j* + *NW* − 1]).

Thus, the analysis interval will be defined as the subset of consecutive samples [*ko*, *ko* + *NW* − 1] where the maximum value of *Awdi f f*(*j*) is found to be:

$$A\_{\mathcal{W}diff(\text{max})} = A\_{\mathcal{W}diff}(k\_0) = \max \left( A\_{\mathcal{W}diff}(j) \right) \forall j \in [1, N - j] \tag{3}$$

where *ko* is the first sample of the analysis interval while *N* indicates the total number of samples of the trace (for each axis).

The rest of the input features for the detection algorithms are computed just taking into account the values of the acceleration components during this analysis interval. We consider as candidate features to feed the machine learning algorithms the following statistics:


$$
\mu\_{SMV} = \frac{1}{N\_{\mathcal{W}}} \cdot \sum\_{i=k\_{\mathcal{o}}}^{k\_{\mathcal{o}} + N\_{\mathcal{W}} - 1} SMV\_i \tag{4}
$$

where *SMVi* represents the Signal Magnitude Vector for the *i*-th measurement of the accelerometer:

$$SMV\_i = \sqrt{A\_{X\_i}^2 + A\_{Y\_i}^2 + A\_{Z\_i}^2} \tag{5}$$


$$
\sigma\_{SMV} = \sqrt{\frac{1}{N\_W} \cdot \sum\_{i=k\_o}^{k\_o + N\_W - 1} \left( SMV\_i - \mu\_{SMV} \right)^2} \tag{6}
$$

This parameter may be clearly affected by the presence of 'valleys' and 'peaks' in the evolution of the acceleration.


$$
\mu\_{SMV\_{diff}} = \frac{1}{N\_W} \cdot \sum\_{i=k\_\sigma}^{k\_\sigma + N\_W - 1} |SMV\_{i+1} - SMV\_i| \tag{7}
$$


$$\mu\_{\theta} = \frac{1}{N\_W} \cdot \sum\_{i=k\_{\theta}}^{k\_{\theta} + N\_W - 1} \left( \cos^{-1} \left[ \frac{A\_{Xi} \cdot A\_{Xi+1} + A\_{Yi} \cdot A\_{Yi+1} + A\_{Zi} \cdot A\_{Zi+1}}{SMV\_i \cdot SMV\_{i+1}} \right] \right) \tag{8}$$


$$
\mu\_{Ap} = \frac{1}{N\_W} \cdot \sum\_{i=k\_o}^{k\_o + N\_W - 1} \sqrt{A\_{Y\_i}^2 + A\_{Z\_i}^{-2}} \text{ for the SensorTags} \tag{9}
$$

$$\mu\_{Ap} = \frac{1}{N\_W} \cdot \sum\_{i=k\_v}^{k\_o + N\_W - 1} \sqrt{A\_{X\_i}^2 + A\_{Z\_i}^2} \text{ for the smartphone} \tag{10}$$

#### *4.2. Employed Supervised Learning Classification Algorithm*

As the detection techniques, we compare four supervised learning classification algorithms that are commonly employed by the related literature (refer to [5,8,41–45] for a state-of-the-art on the detection techniques used in FDS): Support Vector Machine, *k*-Nearest Neighbors, Naives Bayes and Decision Tree, which are briefly presented in the following sub-sections (see [30,46] to gain a much deeper insight into this set of techniques).

#### 4.2.1. Support Vector Machine (SVM)

Support Vector Machine (SVM) is perhaps the most popular learning algorithm employed in the fall detection system literature [31,47–51]. According to the SVM algorithm, the input space defined trough the features of the different training samples is converted into a multi-dimensional space by means of a non-linear mapping.

Thus, as reflected in Figure 5, from the training dataset the algorithm is capable of building a 'maximum margin hyperplane' that acts as a decision boundary to categorize and discriminate the samples. The hyperplane, which separates two multi-dimensional regions, is defined so that the distances to the nearest points (from both sides) are maximized [30]. Once that the system is trained, the classification of the samples is directly based on the region where the sample is included.

**Figure 5.** Example of the performance of the SVM algorithm for a two-dimensional space (samples characterized by two input features): (**a**) distribution of the training data on the two-dimensional space (**b**) creation of the hyperplane and classification decision for a certain test data.

#### 4.2.2. *k*-Nearest Neighbors (*k*-NN)

This instance-based classifier has been utilized as a decision algorithm in works such as [17,50–54]. The typical operation of *k*-NN is represented in Figure 6, utilizes the training dataset in a very simple way: whenever a new activity has to be classified, *k*-NN searches for the *k* already classified samples that are closest to this new uncategorized data. After this, the classification decision is based on the most popular class that is found among these neighbors. Although other techniques (such as Manhattan and city-block distances) have been suggested [42], we compute the Euclidean distance between the numerical features to estimate the distance between the training samples and the sample to be classified.

**Figure 6.** Operation of *k*-NN classifier: (**a**) the data under study is located among the training patterns aiming at finding the *k* nearest neighbor (in the example *k* = 5); (**b**) after detecting the *k* nearest samples, the new sample is classified considering the 'majority vote' (most common class) of its neighbors.

#### 4.2.3. Naive Bayes

This probabilistic classifier (used for FDS in papers such as [50,55]) calculates the probabilities of a particular sample to belong to the existing classes as a function of its associated multi-variate features. By applying the Bayes Theorem, these probabilities are estimated from the prior probability of the classes and the features, as well as the so-called *likelihood* (conditional probability of the features given a class). For this purpose, the statistical characterization of these variables must be previously computed from the training data, 'naively' assuming that the features are mutually independent. From these probabilities, the class of an unclassified pattern is decided taking into account the most likely hypothesis.

#### 4.2.4. Decision Tree

The use of Decision Trees has also been considered by the related literature [50,54,56–58]. The basic operation of the algorithm is exemplified in Figure 7, makes use of the training data to create a decision tree that will allow assigning a class to the unclassified mobility patterns. For this purpose, the algorithm selects a certain feature as a decision 'root' of the tree while different 'branches' are progressively created depending on the values of the other features. The tree is completed when all the features have been considered or when all the training samples in the same branch belong to the same class. The selection criteria of this flowchart-like structure are established to maximize the probability of assigning the correct class to the training data when the different classification rules of the branches are applied.

**Figure 7.** Example of an operation of a decision-tree algorithm: (**a**) after the training phase, the branches and the decision rules the tree are configured; (**b**) during the test phase, the features of an unclassified sample are utilized to apply the decision rules and determine the sample class.

#### **5. Result and Discussion**

In this section we analyze the performance of the four afore-described algorithms when they are employed to detect falls in the datasets obtained from the testbed. Apart from comparing the results of the four machine learning techniques, the goal is to explore two elements that impact on the precision of the fall detection process: (1) the selection of the input characteristics for the machine learning techniques, (2) the number and positions of the sensors that are more relevant for the fall identification.

As performance metrics, we utilize the recall or sensitivity (*Se*) and the specificity (*Sp*), which are commonly employed to evaluate the effectiveness of binary classification systems. In contrast with other metrics used in pattern recognition (such as precision or accuracy), *Se* and *Sp* are not affected by the unbalance between the number of existing samples of both types (in this case, falls and ADLs) which are employed to test the detection algorithms.

The sensitivity describes the capacity of the classificatory to correctly identify an event of the 'positive' class (here: falls) when this event actually occurs. So, it can be computed as the true positive rate:

$$Sc = \frac{TP}{TP + FN} \tag{11}$$

where *TP* indicates the number of 'True Positives', i.e., the number of falls correctly labeled by the classifier, while *FN* is the number of 'False Negatives' or the number of falls that are misidentified as ADLs. The sum *TP* + *TN* corresponds to the total number of falls in the testing dataset.

Similarly, the specificity informs about the efficacy of the detection system to avoid false alarms (or 'False Positives') by properly classifying the actual ADLs. This true negative rate is defined as it follows:

$$Sp = \frac{TN}{TN + FP} \tag{12}$$

where *TN* and *FP* represent the number of 'True Negatives' (ADLs which are well identified) and 'False Positives' (ADLs mistaken as falls), respectively.

In all the detection schemes, a higher specificity is attained at the cost of decreasing the sensitivity of the system. Therefore, as a trade-off between specificity and sensitivity must be achieved in a FDS, we consider the geometric mean of these two parameters (\**Sp*·*Se*) as a global metric to characterize the goodness of the detection process.

Once that the performance metric is defined, for the study of these elements we utilize a 2*<sup>k</sup>* factorial design, where *k* designs the number of possible factors (selected input statistic or sensor position) that can affect the detection process.

Every comparison is repeated for the four considered machine-learning based detection techniques. Thus the effectiveness of the algorithms to detect the falls is also compared in a systematic way under a wide range of circumstances.

To investigate the statistical representativeness of the comparison, the differences in the different results that the alternative use of the factors entails are assessed by means of an ANOVA test. The ANOVA test allows to decide if the means of two or more different populations are equal or not. The test is aimed at determining if different treatments in an experiment cause significant differences in the final results, and consequently, it permits evaluating the importance of the different factors that may alter the operation of the fall detector.

In particular, the factorial ANOVA analyzes the series of residuals resulting from subtracting the values of the obtained observations and the mean of the observations. ANOVA assumes that residuals are independent and follow a normal distribution with a common finite variance (homoscedasticity) [59].

Table 2 details the number of experiments employed for every type of movement (ADLs and falls) as well as the distribution in groups employed for both training and testing with an ANOVA analysis.

To proceed with the supervised learning of the algorithms, the 746 samples obtained from the testbed were divided in a training and a test dataset of 183 and 563 samples, respectively. The number of samples selected to define the training datasets for each type of fall or ADL are described in Table 2. Within the same type, the samples for each phase (training and testing) were randomly chosen.

To apply the ANOVA test, we also divided at random the test dataset into six different 'blocks' or 'subsets' of approximately 93 samples. We decided to divide the samples into six subsets as long as this number would allow representing the Gaussian nature of the residuals with six points while keeping a population of almost 100 samples in each subset for an adequate estimation of the performance metric.

The number of samples of the same movement type is almost homogeneous for the six sub-sets (as also indicated in Table 2). The values of the specificity (*Sp*) and sensitivity (*Se*) were computed after separately applying the four detection algorithms to the six testing sub-sets for all the possible combinations of the impacting factors (input characteristics and sensor positions). Thus, by replicating the test with the 6 sub-sets, the variance of the series of the performance metric (\**Sp*·*Se*) measured for the subsets can be investigated.

#### *5.1. Analysis of the Impact of the Selection of the Acceleration-Based Features on the Fall Detection Performance*

We firstly investigate which characteristic (or set of characteristics) should be selected as input statistics of the machine learning algorithms to maximize the accuracy of the detection decision. For that purpose, we compare the algorithms by analyzing the performance metrics that are obtained when different sets of the input features (described in Section 4.1) are considered. For each set of

characteristics, the global specificity and sensitivity are computed by averaging the metrics that are obtained when a specific combination of body sensors are employed. Thus, the specificity (*Spi*) and sensitivity (*Sei*) for the *i*-th combination of characteristics are calculated as:

$$Sp\_i = \frac{1}{N\_{cs}} \sum\_{j=1}^{N\_{cs}} Sp\_{ij} \tag{13}$$

$$\text{Se}\_{i} = \frac{1}{N\_{cs}} \sum\_{j=1}^{N\_{cs}} \text{Se}\_{ij} \tag{14}$$

where *Spij* and *Seij* indicate the specificity and the sensitivity estimated for the *i*-th combination of input characteristics and the *j*-th combination of employed sensors, respectively. In the equations *Ncs* represents the total number of the possible 31 combinations of the five body sensors, as the algorithms may consider the signals from just one single device to the five sensors (*Ncs* = 25 − <sup>1</sup> = 31).

To proceed with the ANOVA analysis of the results, we focus the comparison on the series of the geometric means of the specificity and the sensitivity (\**Spi* · *Sei*) which are computed for the different sets of test samples (defined in Table 2) and for every combination of input characteristic and considered sensors. The goal is to identify the combination of input features that lead to the best detection performance in a statistically significant way. In case that the best options do not differ significantly, the combination with a lower dimension (i.e., the lower number of features) should be preferred in order to reduce the computational complexity of the system (which is normally intended to be implemented on a wearable device, such as a smartwatch or a smartphone, which can exhibit memory and computation restrictions). This procedure is repeated for the four machine-learning algorithms, whose results are discussed in the following sub-sections.

#### 5.1.1. Results for Support Vector Machine (SVM) Algorithm

We focus on the analysis of the performance of the SVM algorithm when it is alternatively fed with the different possible combinations (2<sup>6</sup> − 1 = 63) of the six 'candidate' input features.

Firstly, to prove the validity of the ANOVA analysis of the results, we check the assumptions of normality and homogeneity of variance of the residuals (differences between the obtained metrics for every sample sub-test and the global mean value of the metric for each combination of input features).

Figure 8a shows that the Cumulative Distributed Function (CDF) of the residuals can be reasonably approximated by a normal distribution. Similarly, Figure 8b depicts the scatterplot of residuals versus the predicted values (estimated means) of the series of performance metrics obtained for every combination of inputs. The figure illustrates that the variability of the series is only clearly different for those input combinations for which the algorithm presents a very poor behavior. In these cases, where the input features do not feed the algorithm with a convenient characterization of the movements, a very low value of the performance metric (with values of the specificity or the sensitivity close to zero) is achieved, and, consequently residuals also tend to zero. Conversely, this problem does not appear for the rest of combinations, for which the variability of the residuals is homogeneous. This lack of homoscedasticity is not critical for the ANOVA test, provided that the experiments are balanced (all series have the same size) and that the ratio of the maximum to the minimum variances of the series does not exceed a proportion of 4 to 1 [59].

**Figure 8.** Results for the SVM algorithm. Analysis of the residuals: (**a**) Normal probability plot (crosses: empirical data, dashed line: theoretical normal fit); (**b**) Values of the residuals versus the predicted (or fitted) values.

In fact, different data transformations that are recommended by the literature [59,60] when the criterion of a common variance is not completely met were also applied to the data series. As the conclusions drawn with these transformations were basically the same for all the cases, for the sake of clarity, we only show the results obtained with the untransformed series and in the original units.

Figure 9 shows the post hoc multiple comparison of the computed means based on the results generated by the one-way ANOVA analysis. The central marks in the bars of the graph represents the estimated mean while the lines expand to the corresponding comparison interval for a 95% confidence level.

In the figure the blue mark and line refers to the optimal combination, i.e., that which minimizes the number of required input features while achieving an estimated mean of the performance metric which is not significantly lower than those obtained with the rest of the combinations. The gray color of the marks and lines indicates those combinations that offer a similar performance to the optimal case but requiring a higher number of input features. The red color is in turn utilized to illustrate the results of employing those feature combinations that underperform (when compared to the optimal case) in a statistically significant way. From the figure, we can observe that the optimal results are generated by the combination that employs *μSMV*, *Awdi f f*(*max*) , *σSMV*, and *μAp* (A = B = C = F = 1, D = E = 0), without considering *μθ* and *μSMVdi f f* . For the combination of these four inputs, the achieved value of the geometric mean of the specificity and the sensitivity lies in the interval [0.735–0.785], which is significantly different from the result attained with the use of 59 out of the other 62 possible combinations of input features.

**Figure 9.** Comparison of the means of the performance metric (\**Sp*·*Se* or geometric mean of the sensitivity and the specificity) obtained with the SVM algorithm for all the possible combinations of input features: A = *μSMV*,B= *Awdi f f*(*max*) ,C= *σSMV*,D= *μθ* ,E= *μSMVdi f f* ,F= *μAp*. For each combination, the *y*-axis of the figure indicates with '1' or '0' whether the corresponding feature is considered ('1') or not ('0'). The combination with the best results is indicated with a blue arrow.

According to the ANOVA analysis, Table 3 shows the relative variation (expressed as a percentage) that every single input feature produces in the results, with respect to the global mean of the metric. The Table also includes the same values for the combinations of two inputs (*μSMV* and *Awdi f f*(*max*) , and *Awdi f f*(*max*) and *μAp*) that also produce the highest variation. As it could be expected, the best global set of inputs is that resulting from combining the four inputs that individually have a higher impact on the result, especially *Awdi f f*(*max*) (42.312% of variation) *μAp* (25.807%) and *σSMV* (14.971%). The interaction of the inputs *Awdi f f*(*max*) and *μAp* adds an increase of 3.032%. To characterize the robustness of the results attained by the algorithm, Table 3 also includes the error reported by the ANOVA analysis, which is 5.218%. This error, which should be minimized, may be explained by other factors that have

not been considered as input features, such as the small differences in the composition of the test subsets (type and number of executed movements, election and particularities of the experimental subjects for each subset, etc.).

**Table 3.** Error reported by the ANOVA analysis and relative influence of the election of the input features on the global result when the Support Vector Machine (SVM) algorithm is applied. Inputs: A = *μSMV*,B= *Awdi f f*(*max*) ,C= *σSMV*,D= *μAp<sup>θ</sup>* ,E= *μSMVdi f f* ,F= *μAp*.


5.1.2. Results for the *k*-Nearest Neighbors (*k*-NN) Algorithm

The previous analysis is repeated for the case in which the detection decision is based on the *k*-Nearest Neighbors (*k*-NN) algorithm. For the application of the algorithm, we set the value of *k* to 9 (nine neighbors). This value, which is selected after different tests, improves the behavior of the algorithm for the employed datasets, as it is close to the number of samples of the different movement types which are employed during the initial training phase of the algorithm.

The results (not depicted here for reasons of space) illustrate again that the CDF of the residuals can be reasonably approximated by a normal distribution while no relevant variation the variability of the residuals is detected for the whole range of predicted values. Consequently, the ANOVA analysis can be considered valid. The post hoc multiple comparison of the performance metric obtained for the different combination of input features indicate that the best performance for the *k*-NN algorithm is achieved when the chosen input characteristics are *μSMV*, *σSMV*, *μθ* and *μAp*. The use of this particular set of characteristics outperforms the results of 59 combinations of the six possible input statistics in a statistically significant way. The particular contribution of the six statistics in the global performance of the detection process is described in Table 4. As it can be observed, the best combination includes the two factors (C = *σSMV*,F= *μAp*) with the highest impact in the results (21.329% and 19.393%, respectively). The high value (6.064%) of the percentage computed for *Awdi f f*(*max*) , which is not included in the best combination of input characteristics, can be actually explained by the negative impact (reduction of the performance metric) that is the achieved if this factor is considered. The table with the conclusions of the ANOVA analysis also informs that the error due to other factors (not considered as inputs) amounts to 21.25%.

**Table 4.** Error reported by the ANOVA analysis and relative influence of the election of the input features on the global result when the *k*-Nearest Neighbor (*k*-NN) algorithm is applied. Inputs: A = *μSMV*,B= *Awdi f f*(*max*) ,C= *σSMV*,D= *μθ* ,E= *μSMVdi f f* ,F= *μAp*.


#### 5.1.3. Results for the Naïve Bayes Algorithm

As in the previous cases, the visual inspection of the CDF and the variability of the residuals show that the conditions of the normal distribution of the population and the homogeneity of the variances are reasonably met. Again, we assume that the small divergences between the actual CDF and the Gaussian fit can be neglected if we take into account that the model is balanced as the number of samples in each test group is basically the same. Thus, we assume the ANOVA analysis can be reasonably applied as a valid tool. Besides, the post hoc comparison of the method when the different combinations of input features are considered indicate that, for the Naïve Bayes algorithm, the best results are achieved when the systems utilizes as inputs the following characteristics: B = *Awdi f f*(*max*) , C = *σSMV*,D= *μθ* and F = *μAp*. The performance of the algorithm for this combination of inputs (which yields a global metric in the interval [0.671–0.745]) significantly differs from those obtained with other 48 combinations. Table 5 in turn specifies the individual influence of each possible input characteristics as well as the four combinations of two inputs that presents the highest impact on the results. The table reveals the importance of selecting *μAp* and *Awdi f f*(*max*) (and their corresponding combination) on the global performance. On the other hand, the error (due to the presence of other external components that affect the evaluation) computed by the ANOVA analysis is 13.484%.

**Table 5.** Error reported by the ANOVA analysis and relative influence of the election of the input features on the global result when the Naïve Bayes algorithm is applied. Inputs: A = *μSMV*, B = *Awdi f f*(*max*) ,C= *σSMV*,D= *μθ* ,E= *μSMVdi f f* ,F= *μAp*.


#### 5.1.4. Results for the Decision Tree Algorithm

The same analysis procedure of the Sections 5.1.1–5.1.3 is applied to the results achieved by the Decision Tree learning method. The normality and the homoscedasticity of the residuals are contrasted and again, the graphs (not depicted here) seem to hold the conditions of the ANOVA test.

From the comparison of the mean performance metrics achieved for the different combinations of input features, we conclude that the best performance of the Decision Tree algorithm takes place when *μSMV*, *Awdi f f*(*max*) *σSMV*, and *μAp* are used as input variables to characterize the mobility of the individuals. For this combination, the performance metric (the geometric mean of the specificity and sensitivity) is estimated to be in the range [0.801–0.849], which is significantly higher than the intervals computed for most other combinations (53 out of 62).

Table 6 describes the contribution of the different statistics (and the most impacting combinations of two factors) to the global performance of the algorithm. The table highlights the importance of employing, as input parameters, the variance of the standard deviation of the acceleration module (*σSMV*) and, in the second place, *μAp* to discriminate falls from ADLs.

The error due to non-considered factors is computed to be 12.110%.

**Table 6.** Error reported by the ANOVA analysis and relative influence of the election of the input features on the global result when the Decision Tree algorithm is applied. Inputs: A = *μSMV*, B = *Awdi f f*(*max*) ,C= *σSMV*,D= *μθ* ,E= *μSMVdi f f* ,F= *μAp*.


5.1.5. Impact of the Election of the Input Characteristics: Summary of the Results

From the study performed in the four previous subsections we can draw the following conclusions:


**Table 7.** Error reported by the ANOVA analysis and relative influence of the election of the six considered input features on the global result for the four learning machine strategies. The symbol ( ) or (-) indicates if the corresponding feature is included (or not) in the combination that achieves the (statistically significant) best performance.


#### *5.2. Study of the Importance of the Sensor Position for the Decision of Machine-Learning Fall Detection Algorithms*

Taking advantage from the multisensory Body Area Network developed for the testbed, this section investigates the importance of the election of the sensor positions for an adequate detection decision. In particular, we try to assess if the simultaneous use of several wearable sensors may introduce any improvement. This analysis, which is aimed at minimizing the number of sensors that should be transported in a real application scenario, will be carried out for each of the four considered supervised learning algorithms. The input statistics that will be taken into account for each algorithm in order to classify the samples will be those (summarized in Table 7) that were proved to induce a better discrimination between ADLs and falls.

In all the tests, we evaluate the performance of the algorithms (expressed again in terms of the geometric mean of the specificity and the sensitivity) for all the 31 possible combinations of the five sensors of the body area network (from the simplest case where only the signals captured by just one single sensor are considered, to the case where the measurements of the five sensors are utilized as the input for the machine learning technique).

In the comparison, the algorithms are always individually applied for each considered position. Thus, when more than one sensor is employed, an 'AND' policy is applied, that is to say, a fall is assumed to have occurred only if it is simultaneously detected by the algorithm in all the considered positions. From some tentative experiments, we checked that the detection based on an 'OR' operation (i.e., a fall is presumed if just one sensor detects it, with independence of the results at the other sensing points) is proved to dramatically increase the number of false positives.

The same procedure of the previous section is again utilized to assess the influence of the position and number of sensors. Firstly, it is confirmed that the series of performance metrics that have been obtained meet the assumptions required by the ANOVA analysis. Then, the results of the analysis itself are presented.

#### 5.2.1. Results for the SVM Algorithm

Figure 10 again illustrates the two graphical tests that can be used to check if the data meet the assumptions of normality and equality of variance. At first glance it is verified that the series do not strictly comply with these conditions. This fact is perceptible in Figure 10b, where the assumption homoscedasticity is shown not to be met completely.

**Figure 10.** Results for the SVM algorithm. Analysis of the residuals: (**a**) Normal probability plot (crosses: empirical data, dashed line: theoretical normal fit); (**b**) Values of the residuals versus the predicted (or fitted) values.

As discussed in the previous sections, there are some possible solutions to reduce this problem. In our case, as recommended in [59,60], we have proved to transform the data (*y*) by applying an exponential function of the type *yα*, where α is a constant. This transformation is intended to improve the normality and equality of variance of the numerical series. However, as this experiment is unbalanced, the variations obtained after this transformation do not meaningfully vary. Thus, so we will continue with the ANOVA analysis with the original data to simplify their understanding.

Figure 11 displays the post hoc multiple comparison of the computed means of the performance metric that resulted from the ANOVA analysis (for which an error of 14.162% was computed), when all the 31 possible combinations of the five BAN sensors are considered. As in the previous section, the figure indicates the estimated mean by depicting a circle in the corresponding bar, which in turn expands to a comparison interval for a 95% confidence level. The blue color is again employed to underline the combination which maximizes the performance, while the red lines correspond to those combinations whose behavior is significantly worse than the optimal case. Gray lines are utilized to identify those combinations whose results present a confidence interval which partially overlap the optimal case (and, consequently, cannot be considered significantly different).

**Figure 11.** Comparison of the means of the performance metric (\**Sp*·*Se* or geometric mean of the sensitivity and the specificity) obtained with the SVM algorithm for all the possible combinations and positions of the sensors. The positions are indicated as: P (Pocket), C (Chest), W (Waist), Wr (Wrist) or A (Ankle). For each combination, the y-axis of the figure indicates with '1' or '0' whether the corresponding sensor is considered ('1') or not ('0') for the detection decision. The combination with the best results is indicated with a blue arrow.

From the results we can infer that the best performance is achieved by using the sensors attached to the waist and chest. Thus, the utilization of any of these two devices as the unique sensor of the system or the combination of both sensors clearly improves the efficiency of the algorithm, as their results significantly differ from the rest. For these three configurations (only-waist, only-chest or the combined use of the sensors on the waist and chest) the computed geometric mean of the sensitivity and the specificity is around 0.95 (which implies false negative and false positive rates lower than 5%). This outperformance could be attributed to the fact that the chest and waist are the sensing points which are closest to the gravity center of the body so that they are less affected by spurious and sudden movements of the limbs. Consequently, the sensors on these positions probably characterize better the mobility of the whole body. Contrariwise, the utilization of the signals captured by the other three devices (the phone in the pocket and the SensorTags on the wrist and on the ankle) do not contribute to the system efficacy and even degrade the algorithm performance. This is particularly true for the case of the sensor on the ankle, as the performance metric plummets below 0.75 for any combination that employs the measurements of the accelerometer located on this point.

#### 5.2.2. Results for the *k*-NN Algorithm

In the case of studying the residuals plots computed when the ANOVA test is applied to the performance metric obtained with the *k*-NN algorithm, we can conclude that the assumptions of normality and homoscedasticity are met. The small variations of the variance as a function of the range of the predicted values can be neglected due to the balanced design of the experiment.

The ANOVA analysis results in an error of 22.857%, which can be justified by the presence of factors different from the position of the sensor, which cause a deviation from the overall average of the experiment, such as the physiognomics and mobility differences of the subjects

From the multiple comparison of the mean values of the performance metric and their corresponding 95% confidence intervals for the different combinations of sensors, we obtain that the best results are again achieved by using the acceleration signals measured at the chest and waist. The differences of the results for these three configurations (only-waist, only-chest and waist-chest) combinations are not statistically significant although they are slightly higher (close to 0.975) than those obtained with the SVM algorithm. Again, the consideration of a more complex BAN (by combining the detection decision on other locations) does not introduce any benefit in the system. Similarly, the use of the signals at the ankle particularly deteriorates the system behavior.

#### 5.2.3. Results for the Naive Bayes Algorithm

The same conclusions as for the previous algorithms can be drawn if we employ a Naïve Bayes strategy to detect the falls. Firstly, the conditions of normality and homoscedasticity of the residuals present a similar behavior to those registered for SVM and *k*-NN algorithms.

The study of the performance metric also reveals that the best option is to consider the detection carried out by the algorithm at the chest, which yields a value of the geometric mean of *Sp* and *Se* in the interval [0.922–0.981]. Similarly, the 95% confidence intervals of the results obtained with the sensor at the waist and with a combination of the sensors on the waist and chest overlaps with that achieved on the chest. Thus, they cannot be considered as poorer. Conversely, the worst performance is again attained whenever the detection process takes into account the application of the algorithm to the signals captured at the ankle. The error of the ANOVA analysis for this algorithm was estimated to be 22.857% (higher than for the precedent algorithms).

#### 5.2.4. Results for the Decision Tree Algorithm

We repeat the testing procedure when the detection algorithm is based on a Decision Tree learning mechanism. We observe again that the criteria of normality and equality of variance of the residuals are not completely met. However, as in the previous experiments, although the data do not strictly comply with these assumptions, the balanced nature of the experiment allows assuming that their influence on the validity of the final results will be limited.

The multiple comparison in turn shows that the optimal operation of the algorithm (with a performance metric in the interval [0.938–0.991]) is also achieved by the use of the signals captured on the chest followed by those of the waist, the combination of waist and chest, the wrist and the combination of wrist and chest. On the other hand, the detection based on the sensors located on the pocket and/or the ankle remarkably deteriorates the effectiveness of the algorithm. Besides, the computed error of the ANOVA analysis was 40.620% (higher than for the precedent algorithms). This is a sign that there are a number of external factors—not considered in the analysis as system variables—that could notably impact on the results.

#### 5.2.5. Summary and Discussion of the Results

In the previous subsections we have assessed the behavior of the four considered machine-learning strategies to detect falls when they are individually applied to the signals produced by the 31 possible combinations of sensors employed in our testbed.

Table 8 summarizes the results obtained by the best combinations that lead to the best results of the geometric mean of the specificity and the sensitivity, as well as the results that are obtained when just a single sensor is employed. The table indicates the sensors that are considered for each combination and highlights in bold those cases in which the performance metric is above 0.9 (which implies that the specificity and sensitivity simultaneously reach values higher than 94.8%). From these results we can draw the following conclusions:


the execution of certain types of ADLs. This bad behavior of this position forces to reconsider the role of the smartphone as a wearable sensor in fall detection systems. Although many studies have proposed its use in FDSs (taking advantage of its popularity and ease of programming), its effectiveness would be very limited unless it is transported in an uncomfortable (e.g., on a belt) or unnatural place (some works have even tested its effectiveness in FDSs when it is firmly attached to the chest). In the case of the smartphones employed in this study, the sampling frequency of its acceleration sensor was higher than that of the sensor motes (200 Hz vs. 20 Hz), so it can also be deduced that a higher sampling rate does not correlate with a better performance.



**Table 8.** Summary of the results of the study on the impact of the sensor position. For every combination, the tick symbol ( ) identifies those sensors that are employed by the algorithm.

#### **6. Conclusions**

This article has presented a systematic assessment of the performance of different supervised learning algorithms when they are utilized for the automatic discrimination of falls and ADLs in wearable fall detection system. The study is based on the analysis of a dataset (UMAFall) obtained through a testbed with experimental subjects. In order to generate the mobility data traces, the subjects were requested to executed a series of predefined movements (ADLs and mimicked falls) while wearing a multisensory wireless Body Area Network, consisting of a Bluetooth network with five inertial mobility units (embedded in four commercial sensing motes and a smartphone) located on five different parts of the body. In our framework, the quality metric selected to weigh up the efficiency of the algorithms to detect falls was the geometric mean of the specificity and variance, which can be considered a good trade-off between the needs of avoiding false negatives (falls identified as ADLs) and false positives (ADLs mistaken as falls).

The purpose of the analysis was twofold. Firstly, we intended to assess the impact of the election of the statistics employed (as input variables for the algorithms) to describe the evolution of the subjects' mobility during the movements. Secondly, we evaluated the importance of the location of the sensors for the detection effectiveness of the machine learning strategies.

The final results enable quantifying the contribution of the different factors involved in the detection of a fall and discerning if they lead to significant changes in the final performance of the system.

As it refers to the characterization of the accelerometer signals, it has been shown that each detection mechanism needs a different combination of input characteristics to achieve its best performance. Nevertheless, there are statistics that augment the effectiveness of all the algorithms whenever they are considered: the standard deviation of the acceleration module (measured during a time window around the acceleration peak), which describes the variability of the human mobility, and the mean module of the non-vertical acceleration components, which characterizes the alterations of the perpendicularity of the subjects' body with respect to the floor plane.

In any case, results suggest that a greater number of input features does not necessarily lead to a better operation of the machine learning algorithms. In fact, in many cases, this increases in the complexity of the system and even produces a decrease in the measured performance metric.

Besides, it has been evidenced that the location of the employed sensors strongly influences the efficacy of the FDS. In this regard, for all the studied algorithms, the best results correspond to the detection based on the signals captured on the chest and/or waist (the two considered sensing points which are closest to the center of mass and gravity of the human body). In addition, it has also been inferred that the combined use of several sensors (aimed at decreasing the number of false alarms) does not improve the performance, since it heavily decreases the sensitivity. The study has also shown that a sensor in a trouser pocket (the position that is conventionally utilized by for many users to transport a smartphone) is not the best option for an adequate discrimination rate. Consequently, a fall detection system merely based on a smartphone (as it is massively proposed by many papers in the literature) is not feasible as long as the phone should be constantly fixed to a belt or 'worn' in a quite unnatural way (firmly attached to the chest).

Another important contribution of the paper is the employed methodology as the validity of all the comparisons was evaluated through an ANOVA analysis. Thus, it has been possible to determine the statistical significance of the differences obtained when the algorithms were tested under different configurations of the detection process (input variables, selection of the sensors). The study of the statistical significance of the results (an aspect that is normally neglected by the existing bibliography about FDSs) is completed by the information reported by the error computed by the ANOVA analysis. This error can be used as a metric to evaluate the robustness of the algorithms with respect to the presence of other impacting factors that have not been taken into consideration for the analysis (such as the personal characteristics of the subjects).

Many mid- and low-end smartphones are still manufactured without gyroscopes. Thus, as the smartphone is a key element in our architecture, this study has only focused on the discrimination of falls based on signals captured by the accelerometers. However, the trunk angle velocity and trunk orientation have been proved to be two key parameters to detect anomalies in the mobility of the body during the moments immediately preceding the impact [71]. Our future research plans to extend this study by incorporating metrics derived from the signals measured by the gyroscope embedded in the sensors. So, the contribution of the information on the changes of the body orientation to the fall detection process could also be evaluated.

The analysis of the performance of the fall detection schemes of this work was developed in an offline mode (by processing the samples on a computer using Matlab scripts), as the general goal was to assess the 'abstract' effectiveness of the machine learning strategies under different circumstances. Therefore, the problems related to the implementation of this type of architectures in a wearable system (with inherent hardware and software limitations) were outside the scope of our study. However, these aspects (which are normally ignored by the literature) should also be systematically investigated.

Future studies should analyze in detail the implementability of systems that combine wireless communications with external sensors and complex artificial intelligence techniques that must make decisions in real time.

Battery-powered wearable units acting as the core of a Body Area Network for a FDS may have serious shortcomings to provide the computational resources (computing speed, memory, etc.) required by the communication dynamics and the algorithm of the fall detection service. In the case of using a conventional smartphone as the center of the sensor network, the practical coexistence of the detection application and the other typical activities executed by the device (making calls, messaging, browsing, etc.) should be also studied in detail. Otherwise, the use of a smartphone as a specific tool exclusively intended for fall detections (as it is conceived in those architectures where the phone is firmly attached to the user's chest) is pointless. In this context, battery drain due to FDS applications should be also studied in detail.

**Acknowledgments:** This work was supported by Universidad de Málaga, Campus de Excelencia Internacional Andalucia Tech, Málaga, Spain.

**Author Contributions:** J.A.S.-R. deployed the testbed, executed the tests to collect the dataset, proposed the ANOVA analysis methodology, programmed and performed the off-line study of the algorithms, analyzed the results and co-wrote the paper. E.C.-P. designed the experimental setup, defined the comparison procedure and the machine-learning algorithms to be compared, co-analyzed the results, elaborated the critical review, wrote and revised the paper. J.M.C.-G. helped in the collection of datasets and gave technical advice regarding the implementation of the testbed.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Article*
