**Gait Analysis in Parkinson's Disease: An Overview of the Most Accurate Markers for Diagnosis and Symptoms Monitoring**

**Lazzaro di Biase 1,\*, Alessandro Di Santo 1, Maria Letizia Caminiti 1, Alfredo De Liso 1, Syed Ahmar Shah 2, Lorenzo Ricci <sup>1</sup> and Vincenzo Di Lazzaro <sup>1</sup>**


Received: 10 June 2020; Accepted: 17 June 2020; Published: 22 June 2020

**Abstract:** The aim of this review is to summarize that most relevant technologies used to evaluate gait features and the associated algorithms that have shown promise to aid diagnosis and symptom monitoring in Parkinson's disease (PD) patients. We searched PubMed for studies published between 1 January 2005, and 30 August 2019 on gait analysis in PD. We selected studies that have either used technologies to distinguish PD patients from healthy subjects or stratified PD patients according to motor status or disease stages. Only those studies that reported at least 80% sensitivity and specificity were included. Gait analysis algorithms used for diagnosis showed a balanced accuracy range of 83.5–100%, sensitivity of 83.3–100% and specificity of 82–100%. For motor status discrimination the gait analysis algorithms showed a balanced accuracy range of 90.8–100%, sensitivity of 92.5–100% and specificity of 88–100%. Despite a large number of studies on the topic of objective gait analysis in PD, only a limited number of studies reported algorithms that were accurate enough deemed to be useful for diagnosis and symptoms monitoring. In addition, none of the reported algorithms and technologies has been validated in large scale, independent studies.

**Keywords:** Parkinson's disease; gait analysis; diagnosis; symptoms monitoring; wearable; home-monitoring; machine learning

#### **1. Introduction**

Parkinson's' disease (PD) gold standard for diagnosis and symptoms monitoring is based on clinical evaluation, which includes several subjective components. The lack of objective and quantitative biomarkers for diagnosis and symptoms monitoring leads to significant direct and indirect healthcare cost. Based on the current diagnostic criteria [1], the diagnostic error rate is around 20% [2]. In addition, PD is a dynamic disease (i.e., symptoms changes during the disease course) that requires continuous adjustment of therapy.

In the early stages of PD, the most effective treatment to alleviate motor symptoms is oral L-DOPA [3]. However, during moderate and advanced stages, in addition to cardinal motor symptoms, the patient may show motor fluctuations and dyskinesia. During this stage, the brain becomes very sensitive to dopamine level fluctuations, and a continuous stimulation (instead of pulsatile drugs administration) may help in controlling motor fluctuations, dyskinesias and cardinal motor symptoms. This stimulation may be pharmacological with levodopa [4–6] or dopamine agonists [7], or provided

by DBS (deep brain stimulation) [8–12]. During moderate and advanced stages, gait problems, like freezing of gait and reduced balance and postural control, become more evident and unlike cardinal motor symptoms, PD patients respond less to conventional therapy (i.e., oral L-DOPA).

In line with cardinal motor symptoms, to date, gait problems are evaluated with semiquantitative rating scales like the unified Parkinson's disease rating scale (UPDRS) [13] or the movement disorders society unified Parkinson's disease rating scale (MDS-UPDRS) [14]. In an effort to improve PD management and move towards a quantitative and home-oriented assessment and recognition of PD motor symptoms, different technologies have been used to evaluate bradykinesia [15–17], rigidity [17–20], tremor [21–23] and axial symptoms [24–27].

Gait impairment is an evolving condition and different patterns of gait disturbances can be detected throughout the progression of the disease [28]: reduced amplitude of arm swing, reduced smoothness of locomotion, increased interlimb asymmetry [29], low speed, reduced step length [29], shuffling steps, increased double-limb support, increased cadence [28], defragmentation of turns (i.e., turning en block), problems with gait initiation [30], freezing of gait and reduced balance and postural control [28].

Some gait features in PD are specific, and get worse during the disease course. An objective and quantitative gait analysis system could, therefore, potentially improve the current practice (semiquantitative gait evaluation) that may aid in diagnosis, symptom monitoring, therapy management, rehabilitation and fall risk assessment and prevention in Parkinson's disease patients. Among all these promising applications of gait analysis in Parkinson's disease, we have confined the scope of our review to two main unmet needs in this disorder: the diagnostic error, and the lack of objective biomarkers for motor status discrimination. Therefore, the main aim of this overview is to summarize the most important technologies used to evaluate gait features and the associated algorithms that have shown promise in using gait analysis to aid diagnosis and symptom monitoring in Parkinson's disease (PD) patients. The scope of the review was confined to studies that showed any promise of being clinically useful (i.e., are both highly sensitive and specific defined as those with at least 80% sensitivity and 80% specificity) for diagnosis or motor status discrimination.

#### *1.1. Gait Features*

Human gait is a sequence of involuntary movements, cyclically repeated and triggered by voluntary movement. Several components could be used to objectively measure and analyze gait cycle. These components are typically categorized into spatiotemporal, kinematics and kinetics features [31,32].

#### *1.2. Spatiotemporal Features*

Several spatiotemporal features can be used for gait analysis (Table 1). These features are the more commonly used types of features to objectively describe the gait pattern in healthy subjects and patients with several diseases. Spatiotemporal features could refer to the global gait cycle or to the stride cycle.


**Table 1.** Spatiotemporal gait and stride features.

Each gait cycle starts with the initial contact of one foot and ends with a new initial contact of the same foot (Figure 1). One single cycle is composed of a stance and a swing phase: the stance phase is the period during which the foot is in contact with a support surface, and the swing phase is the period during which the same foot is airborne in preparation for the next gait cycle. During the gait cycle, the legs can be individually or simultaneously placed on the ground, so it is possible to identify a single limb support stage, that is the phase during which only one foot is on the surface, and a double limbs support stage during which both legs are on the support surface during a one step cycle (Table 1).

The step and the stride are defined as the length/duration between 2 successive events of same type on opposite limbs, and on the same limb, respectively (Table 1; Figure 2).

**Figure 2.** Stride analysis of a single gait cycle.

The step width represents the horizontal distance measured between the position of the feet on the same event. Finally, foot progression angle can be measured, this represents the angle between the longitudinal axis of the foot and the line of gait progression (Table 1).

#### *1.3. Kinetics Features*

The kinetics analysis (or dynamics of gait) is the study of the forces and their effect on motion. The dynamics forces are the causes of the motion that result in the kinematic movements. Commonly, these forces are represented by the ground reaction force (GRF) on the hip, knee and ankle joints calculated on the sagittal plane. GRF is described only when feet are in contact with the ground (stance phase) and represents the effect of gravity on a body area counterbalanced by the contact with ground and the limb muscular activation [31,32] (Figure 3). GRF refers to a center of pressure (CoP) that is the point of force application.

**Figure 3.** Gait kinetics (dynamics) features.

#### *1.4. Kinematics Features*

Kinematics features describe the movements without taking the forces causing the movements into account [31,32]. Kinematics analysis could describe gait features based on the sagittal, horizontal or frontal plane for several body areas and joints such as the ankle, knee, hip and pelvis. Kinematics features could be extrapolated both from the stance and swing phases. The kinematic analysis of the position, velocity and acceleration of a body part can be determined. Angular kinematics objectively quantify (as degrees) the joint's motion around axes in different gait phases (Figure 4).

**Figure 4.** Gait kinematics features.

#### *1.5. Gait Analysis Technologies*

Several technologies can be used for quantitative gait analysis. Technologies can be divided into two main subtypes: wearables and non-wearables [33]. Wearable sensors used for gait analysis are inertial sensors [34], goniometer [35], pressure and force sensors [36], electromyography (EMG) [37,38], IR-UWB (impulse radio ultra-wideband) [39] and ultrasound [40]. Among non-wearable sensors, the most common types are floor sensors [41,42] and image processing-based technologies (such as a single or multiple cameras [43–45], time of flight [46–48], stereoscopic vision [49,50], structured light [51] and IR thermography [52]).

Accelerometers, gyroscopes and magnetometers can be the component of the same inertial measurement unit (IMU) device (Figure 5), one of the most widely used type of sensors in gait analysis especially in PD [33,53]. It can measure velocity, acceleration, orientation and gravitational forces and can be used to study gait initiation [54], assess standing balance [55] and quantify bradykinesia [56].

**Figure 5.** The same inertial measurement unit (IMU) device composed by accelerometers (**A**), a gyroscope (**B**) and a magnetometer (**C**). Legend: (**A**) ax, ay and az = linear acceleration on the three axis x, y and z; (**B**) αx, α<sup>y</sup> and α<sup>z</sup> = angular acceleration on the three axis x, y and z and (**C**) μx, μ<sup>y</sup> and μ<sup>z</sup> = magnetic moment on the three axis x, y and z.

Accelerometers are composed of a mechanical sensing element with a proof mass attached to a mechanical suspension system, with respect to a reference frame, that can be forced to deflect by the inertial force according to the acceleration of gravity. The acceleration can be measured electrically using the physical changes in the displacement of the proof mass with respect to the reference frame [34]. Accelerometer can be attached to the feet, legs or waist [33]. Gyroscope is based on the property that all body that revolves around an axis develop rotational inertia determined by the body's moment of inertia [33]. Basically, a gyroscope is an angular velocity sensor. Magnetometer is based on the magneto resistive effect and can estimate changes in the orientation of a body segment in relation to the magnetic north. It can provide information that cannot be determined by both an accelerometer and gyroscope [34].

Goniometers work with resistance that changes depending on how flexed the sensors is. When flexed, the resistance increases proportionally to the flex angle. Goniometers are easy to set up and use a simple algorithm [33]. Goniometers are commonly used to study the angles for ankles, knees, hips and metatarsals [35].

Pressure and force sensors measure the force applied on the sensor without considering the components of this force on all other axes [33].

Force sensors measure the ground reaction force under the foot and return a current or voltage proportional to the pressure measured. Usually this kind of sensor is easily integrated into instrumented shoes [36]. Pressure and force sensors have been used to study stride length variability in PD patients with freezing of gait (FOG) [57]. Electromyography is a neurophysiologic technique that registers the electrical signals associated with motor unit activity, both voluntary and involuntary. The electrical signal can be recorded with surface electrodes (non-invasively) or needle electrodes (invasively) [58]. Some EMG electrodes are commercialized in combination with wireless technology and play an important role in evaluating walking performance during gait [38]. EMG can be used to study postural disorders in Parkinson's disease patients, like exploring muscular activity in the Pisa syndrome [59].

The impulse radio ultra-wideband (IR-UWB) technique can detect and track movements non-invasively with high resolution and accuracy through emitting impulse radio waves of very short duration, and receiving the reflected waves from the target body [60]. It has been used to quantify activity measurement in movement disorders [61]. This technology also shows a good penetrating power able to detect the motions of internal organs. This technique can also be used for a wearable healthcare system to continuously estimate foot clearance due to its high temporal resolution, low power consumption and multipath immunity [62]. This technology can be used for step and gait phase detection [39].

Ultrasonic sensors can measure the time a sound takes to send and receive the wave produced as it is reflected from an object. Knowing the time and the speed, we can estimate the distance between two points [33]. This kind of technology is useful for step length measurement and gait phase detection [40] to analyze bilateral gait symmetry and coordination [63].

Among non-wearable sensors, the single camera image processing system is composed of single or multiple cameras that can be used to obtain information about gait in selected individuals. This technique allows individual recognition and segment position localization. Image processing has been used to identify people by the way they walk [44] and has several medical applications such as gait recognition considering changes in the subject path [64], and study of the gait kinematic [65].

Time of flight (TOF) systems are based on cameras using signal modulation that measure distances between the camera and the subject based on the phase-shift principle [46]. The TOF system can detect the segment position, gait phase, foot plantar pressure distribution and are useful for individual recognition [47]. TOF systems have been used to assess medication adherence in patients with movement disorders [66].

Stereoscopic vision is used to determine the depth of points in the scene by using a model through the calculation of similar triangles between the optical sensor, the light-emitter and the object in the scene. This could be useful in gait phase detection, segment position and individual recognition [49].

Structured light is the projection of a light pattern under geometric calibration on an object whose shape is to be recovered [33]. This technology is used for segment position study and gait phase detection. Kinetic sensor is one of the most common devices using this technology to create a marker-based real-time biofeedback system for gait retraining [51].

Infrared thermography (IR) creates visual images based on surface temperatures. For studying human gait, its functioning is based on skin emissivity. This method has been applied to recognize the human gait pattern [52].

#### *1.6. Machine Learning Algorithms Application for Gait Analysis*

There has been increasing use of machine learning (ML) in medicine including neurology to aid diagnosis, and patient management using risk stratification [67,68]. ML algorithms learn from data (past experiences) by identifying underlying patterns and relationships. The field of ML can broadly be categorized into supervised, unsupervised and reinforcement learning.

Supervised learning (SL) begins with the aim of predicting a known output or target. Indeed, an SL algorithm takes a known set of input data (the training set) and known responses to the data (output), and trains a model to generate reasonable predictions for the response to new input data. In such algorithms, the artificial intelligence (AI) is approximating what a trained physician is already able to perform with high accuracy. This approach means that the learning algorithm generalized the training data to previously unobserved situations in a "reasonable" way.

All forms of SL algorithms can be classified as either classification or regression. Classification techniques predict discrete responses. Regression techniques, instead, are used to predict continuous responses. They can also be used for modeling the risk, meaning that the computer is doing more than merely reproducing the physician skills. These algorithms are also capable of discovering new associations not apparently evident to human's preliminary interpretation.

Differently from SL, in the unsupervised learning (UL) algorithm, we were no longer interested in predicting outputs. Instead, we aimed to discover naturally occurring patterns or groupings within the data. It is important to emphasize that the examples given to learners were unlabeled; thus, there was no error or reward signal to evaluate a potential solution. Common UL clustering algorithms could broadly divided into three groups: hard clustering, where each data point belongs to only one cluster, and soft clustering, where each data point can belong to more than one cluster; and dimensionality reduction techniques.

Reinforcement learning (RL) is an approach in ML that states what actions an agent should take in an environment to capitalize on the idea of an increasing reward. RL is different form standard SL in that correct input/output pairs are never presented, nor are suboptimal actions explicitly corrected. The primary goal is the direct performance, which involves finding a balance between exploration of unknown datasets and exploitation of current knowledge [69].

The most widely used ML technique in gait analysis is SL, with varying levels of complexity and interpretability. Table 2 describes the most used algorithms in this field.


#### **Table 2.** Machine learning algorithm used for gait analysis.

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

In line with the study of Sánchez-Ferro, et al. [75] we used a similar search string for axial symptoms in Parkinson's disease patients, in PubMed for articles published between 1 January 2005, and 30 August 2019 (Table 3). We identified studies that used technologies to distinguish PD patients from healthy subjects or to differentiate PD motor status or different disease stages. We only selected studies that declared a sensitivity and specificity of at least 80% when using gait analysis for either diagnosis or motor status discrimination. Additionally, further relevant articles based on the author's knowledge of the state of the art in this field were also added.


**Table 3.** Search strategy.

For each selected study, we collected data about the technology, the algorithm used and its performance metrics like accuracy, sensitivity and specificity. In addition for studies, which declares only the regular accuracy ((true positives + true negatives)/(true positives + true negatives + false positives + false negatives)), the balanced accuracy ((sensitivity + specificity)/2) was calculated. This is because for unbalanced test sets, balanced accuracy is a better index for accuracy than regular accuracy [76].

#### **3. Results**

#### *3.1. Discrimination of Parkinson's Disease from Healthy Subjects*

According to the inclusion and exclusion criteria, after the literature search and studies screening, 10 studies were selected that focused on distinguishing Parkinson's disease patients from healthy subjects with gait analysis. To distinguish PD from healthy subjects, several technologies can be used (Tables 4 and 5). One study used the data collected from wireless inertial sensors (Micro-attitude and heading reference system (AHRS) model, MicroStrain, Inc, Williston, VT, USA) placed on the foot in PD patients and healthy subjects to detect peculiar gait features and distinguish PD patients from controls [77]. In particular, authors detected physical kinematic features of pitch, roll and yaw rotations of the foot during walking and used principal component analysis (PCA) to select the best features that were subsequently used for the SVM method to classify PD patients, with and without gait impairment, and healthy subjects. From 67 collected features, they selected 15 kinematic features divided in three categories: pitch, roll and yaw features. The proposed classification has very high sensitivity, specificity and positive predict values (93.3%, 95.8% and 97.7% respectively) to distinguish PD patients from healthy subjects [77].


**Table 4.** Parkinson's disease vs. healthy subjects discrimination.

Abbreviations: CoP: center of pressure; CV: coefficient of variation; kNN: k-nearest neighbor, LDA: linear discriminant analysis; NA: not available; SVM: support vector machine; VGRF: vertical ground reaction force.



**Table 5.** *Cont*.


**Table 5.** *Cont*.


**Table 5.** *Cont*.

Another study compared two machine learning algorithms (decision-tree and neural networks) to differentiate healthy subjects gait patterns in different disease conditions in an elderly population (including patients affected by PD, hemiplegia, leg pain and back pain). Authors used movements data obtained from 12 retroreflective tags placed on the body captured by an infrared camera (Smart IR motion capture system) [78]. They studied 45 healthy controls and 25 PD patients. Predictors were based on velocity and calculated body distances (i.e., difference between average distance between right elbow and right hip and average distance between right wrist and left hip or the angle between two body segments). Global classification accuracy was high for both systems and reached over 95% for decision tree and more than 99% for neural network [78].

Moreover, Barth, et al. [79] demonstrated a good sensitivity and specificity (88% and 86% respectively) to differentiate healthy subjects and early PD patients using only a single mobile inertial sensor (gyroscope and accelerometer, integrated in the SHIMMER Company system) placed over the shoes. The patients performed standardized gait tests and from this data step, signal sequence and frequency features were extrapolated and used as predictors for the linear discriminant analysis (LDA). Moreover, they demonstrated that this system was able to distinguish between mild and severe gait pattern with high sensibility and specificity (100%).

Another approach was used by Yoneyama, et al. [80]: they used a single accelerometer placed on the waist and performed gait analysis to compare a continuous 24-h assessment of 10 PD patients and 17 healthy controls [80]. They used a gait detection algorithm based on the gait cycle (i.e., stride-to-stride time interval) and gait-induced acceleration relationship using 3 features of gait (high intensity, periodicity and biphasicity of gait) that introduced a set of indices in order to quantify subject's walking mode and to assess daily gait characteristics. All the calculated indices were smaller in the PD group, and the proposed method was able to distinguish the PD gait from the normal gait with 100% sensitivity, 94.1% specificity and 96.3% accuracy. These results suggest that the afore-mentioned systems could differentiate normal subjects from those with movement disorders [80].

To easily and objectively assess the difference between PD patients and healthy subjects, Kugler, et al. [81] proposed a classification algorithm based on data collected from surface wireless EMG (Delsys Trigno, Delsys Inc., Boston, MA, USA) positioned on two inferior limbs muscles during the performance of standardized gait tests in five PD patients and five healthy subjects [81]. Furthermore, data from accelerometers (Trigno sensor) placed on heels were collected and used for step segmentation. Statistical and frequency features from EMG signals were used to train an SVM algorithm for step detection. The proposed step detection method reached 98.9% sensitivity and 99.3% specificity and the classification accuracy to distinguish between PD and healthy subjects reached 90% sensitivity and 90% specificity (average value).

Arora, et al. [82] investigated the feasibility and the accuracy of smartphones' built-in tri-axial accelerometer (in LG Optimus S) developing an app to objectively assess PD patients and distinguish them from healthy subjects [82]. They studied 10 PD patients and 10 controls for 1 month and during execution of the gait tests to extract 23 features of frequency and time domain from accelerometer and subsequently used a random forest method to distinguish between PD and controls. This system reached a 98% balanced accuracy, 98.5% sensitivity and 97.6% specificity.

Another study proposed the use of a Bayesian gait recognition method based on data acquisition by video infrared camera system (Microsoft Kinect depth sensors) [83]. This system consisted of an infrared projector and two infrared cameras (on left and right) that follow the structured light principle. The collected data are converted in the depth frame matrix, depth frame contour, image frame matrix and skeleton numbering. Then the acquired data were further analyzed through MATLAB software. Eighteen PD patients, eighteen healthy subjects and fifteen healthy students were assessed and probabilistically classified according to their detected gait featured (stride length and gait speed) through skeletal tracking. This Bayesian system used the stride length, gait speed and age as features and was able to distinguish between PD patients and controls with 92.2% accuracy combining stride length and gait speed, and 94.1% accuracy combining stride length and patient/healthy subjects' age.

Djuri´c-Joviˇci´c, et al. [84] used an electronic walkaway to distinguish between the PD patient and controls. Authors compared 40 de novo PD patients and 40 controls while walking selecting three different tasks: normal pace walking, dual motor task (as walking and carrying a glass of water) and walking during mental task execution. The most relevant predictor variables were selected (19 features) including the stride length, stride length coefficient of variation (CV), swing time, step time asymmetry and heel-to-heel base support CV. These features were selected with the random forests algorithm and the classification accuracy of these selected features was tested with the support vector machine. The overall accuracy combining the three conditions was 85% to identify de novo PD patients from healthy subjects, with a sensitivity of 85% and a specificity of 82%. Their study also found that step time asymmetry and the support base CV are the most relevant factors that contribute to global system accuracy.

Alam, et al. [85] proposed a novel mathematical method to assess the gait in PD patients and controls using data from eight sensors below each foot and extrapolated features from vertical ground reaction force (VGRF) data recorded during subjects walking [85]. Twenty-nine PD patients and eighteen healthy subjects were enrolled in this study. Three different algorithms (sequential forward selection, minimum redundancy maximum relevancy feature selection (MRMR) and the mutual information-based feature ranking method) were applied to select the best features extracted from VGRF data and 13 features (like CV swing time, CV stride time and centre of pression data) were chosen. Finally, four different machine learning classifiers (SVM, k-nearest neighbor-kNN, random forest and decision trees) were compared to distinguish the gait pattern between healthy subjects and PD patients. The accuracy of the machine learning methods ranged from 85.21% (kNN) to 95.7% (SVM cubic kernel). SVM with cubic kernel showed a sensitivity of 94.4% and a specificity of 96.6%.

Finally, Pham and Yan [86] used a tensor decomposition algorithm called canonical polyadic decomposition (CPD) also known as Parallel Factor Analysis (PARAFAC), a generalization of PCA, to differentiate the multisensors time series of the gait between PD and controls in a previous published dataset of 93 PD patients and 72 controls [86]. Data were collected from load sensors (Ultraflex Computer Dyno Graphy, Infotronic Inc., Vriezenveen, NL, USA) placed on each shoe that recorded force in a function of time. Tensor-decomposition factors of control and PD patients showed a distinct relationship. This system used the full length of the VGRF time without considering the minimum number of strides required for effective tensor decomposition analysis of the gait dynamics. This system can be applied for very short time duration signals and can resolve the problem of obtaining several trials for stable and trustable results. Authors showed 100% of accuracy, sensitivity and specificity to distinguish PD and controls.

#### *3.2. Parkinson's Disease Motor Status Discrimination*

According to the inclusion and exclusion criteria, after a literature search and studies screening, only three studies were selected and these papers aimed at identifying different motor statuses in Parkinson's disease (Tables 6 and 7). Sensors can be used to discriminate the different motor status in individual PD patients, in a single disease's stage, monitoring the motor fluctuations, or can be used in a longitudinal way to monitor the motor status changes during the disease evolution.


**Table 6.** Parkinson's disease motor status discrimination.

Abbreviations: LDA: linear discriminant analysis; NA: not available; SVM: support vector machine.


**Table 7.** Parkinson's disease motor status discrimination features selected.

Samà, et al. [87] focused their research on automatic detection of bradykinesia, the cardinal symptom of PD [87]. They proposed a mathematical algorithm to automatically identify bradykinesia in PD patients at home using an SVM classifier (that detects strides) based on data collected from a single accelerometer placed on the waist combined with a video-recording of the examination, they correlated the stride frequency with the UPDRS bradykinesia score. Their methods showed a high accuracy (>90%) to identify bradykinesia, a high sensitivity and specificity (92.52% and 89.07%, respectively) and a good correlation with UPDRS specific items.

To objectively detect the motor on–off fluctuations, an integrated system like REMPARK (personal health device for the remote and autonomous management of Parkinson's disease, FP7 project REMPARK ICT-287677) was used [88]. The REMPARK system consists of an algorithm added in an app inside a smartphone that used the data from an accelerometer placed on the iliac crest. This system was developed for longitudinal evaluation. In this study, 41 PD patients were enrolled for a 3-day monitoring. For the on/off state discrimination, authors developed an algorithm, which analyzed gait [89] and dyskinesias [90]. The algorithm responses were compared to a self-reported on–off diary. The REMPARK system showed 97% sensitivity in detecting off states and 88% specificity in detecting on phases compared to diaries.

Barth, et al. [79] demonstrated, in a cohort of 14 early stage PD patients and 13 intermediate PD stage patients, that a mobile and light inertial sensor (gyroscope and accelerometer, integrated in the SHIMMER system) placed over the shoes allows differentiation between the two groups with 100% sensitivity and specificity with using a linear discriminant analysis (LDA) classifier that combines step, signal sequence and frequency features as predictors [79].

#### **4. Discussion**

Several studies aimed at detecting specific patterns of gait alterations in Parkinson's disease by using a quantitative technology-based assessment [28,29,91,92]. Gait impairment is an evolving condition throughout the progression of the disease and different patterns of gait disturbances can be detected in early, mild to moderate and advanced stages [28].

Early specific alterations include reduced amplitude of arm swing, smoothness of locomotion and increased interlimb asymmetry [29]. Impaired muscle contraction, rigidity and postural instability contribute to reduced forward limb propulsion, which, in turn, can negatively affect spatiotemporal gait parameters, such as speed and step length [29]. In particular, reduced step length seems to be a specific feature of Parkinson's disease gait [92]. Sensor-based observations showed that the increased variability in gait reflects increased gait instability that can be detected early in the disease and can be a useful marker of disease progression [91,93]. To carry out two tasks at the same time, a paradigm known as dual-task interference, is particularly complex in PD patients because of two independent effects influencing gait: the first is an age-associated reduction in gait performance unrelated to pathology, and the second one is a PD-specific effect due to a dual-task coordination deficit interfering with postural control. The latter suggests reduced stability and ability to adapt to PD patients under dual-task conditions [93]. Arm swing outcomes provide a sensitive measure of decline in gait function in PD under dual-task conditions [91]. On the other hand, one of the most representative early feature of Parkinsonian gait, reduced speed, is not disease specific [28].

In the mild-to-moderate stage, symptoms spread bilaterally so that asymmetry might decrease [94]. Gait problems worsen and shuffling steps, increased double-limb support and increased cadence become common [28]. Motor automaticity becomes further impaired, resulting in fragmented motor function, such as defragmentation of turns (i.e., turning en block) and problems with gait initiation [30].

Further worsening in gait characterizes the advanced stage of the disease, with more frequent freezing of gait (FOG) and motor blocks, reduced balance and postural control, motor fluctuations and dyskinesia [28].

The use of several sensors technologies to objectively assess Parkinson's disease (PD) symptoms has exponentially increased in the last twenty years. Despite different features such as analysis of the face [95,96], speech [97,98], bradykinesia [15–17], rigidity [17–20] and tremor [21–23] being explored for objective PD evaluation, gait analysis has received widespread attention as part of an objective assessment for PD patient examination. Gait analysis is based on capturing movement with a motion capture device that can be wearable or non-wearable. The analysis of the acquired signals used different statistical or machine learning algorithms like the support vector machine (SVM), dynamic neural network (DNN), naïve Bayes, random forests or decision tree. For machine learning algorithms, the more complex the algorithm, the more likely it is able to determine an optimal decision boundary and hence improved accuracy. However, this improvement in accuracy comes at the cost of reduced interpretability. Among the algorithms included in this survey, decision trees are the most interpretable. All the algorithms used data derived from several features of the gait pattern, which can be grouped into three parameters groups: spatiotemporal, kinematic and kinetic [85]. Spatial parameters measure the physical distance between two steps (like strength length); temporal parameters evaluate the time spent to complete a gait cycle (like the cadence, duration of swing and stance phase) and kinematic parameters evaluate the movement of an object without consideration of its cause while kinetic parameters measure the force that cause the movement (like the ground reaction force during walking) [85].

The majority of these studies assessed if the motion capture device and associated algorithms can distinguish between PD patients and healthy subjects. Other studies investigated if these tools could help to classify different motor status of PD or identify various disease stages.

Among studies focused on discriminating PD vs. healthy subjects, various studies showed high accuracy (more than 90%; Table 4). In addition, regarding studies focused on motor status discrimination, Bayes, et al. [88] were able to discriminate the on/off state by merging an algorithm, which analyzed gait and dyskinesia; instead Barth, et al. [79] were able to differentiate the early vs. intermediate PD stage, with 100% sensitivity and specificity (Table 6). However, it should be considered that all these algorithms need to be validated on larger and representative populations in order to avoid overfitting the problem, which makes the algorithm valid only for the analyzed sample. In addition, the simplest algorithms that provide acceptable accuracy are preferable (i.e., logistic regression and decision trees) over more complex algorithms (e.g., SVM and neural networks) that may provide slightly higher accuracy but are less interpretable.

#### **5. Conclusions**

The present overview showed that among the high volume of literature, published on the topic of objective gait analysis in PD, only few studies showed accurate algorithms that can potentially be clinically useful for diagnosis and symptoms monitoring. However, none of those studies have been independently validated or tested on a large scale.

**Author Contributions:** Writing—original draft preparation, L.d.B.; A.D.S.; M.L.C.; A.D.L.; S.A.S. and L.R.; writing—review and editing: L.d.B., S.A.S. and V.D.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

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

#### **References**


© 2020 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* **Fifteen Years of Wireless Sensors for Balance Assessment in Neurological Disorders**

**Alessandro Zampogna 1,**†**, Ilaria Mileti 2,**†**, Eduardo Palermo 2, Claudia Celletti 3, Marco Paoloni 3, AlessandroManoni 4, IvanMazzetta 4, Gloria Dalla Costa 5, Carlos Pérez-López 6,7, Filippo Camerota 3, Letizia Leocani 5, Joan Cabestany 6,7, Fernanda Irrera <sup>4</sup> and Antonio Suppa 1,8,\***


Received: 3 April 2020; Accepted: 3 June 2020; Published: 7 June 2020

**Abstract:** Balance impairment is a major mechanism behind falling along with environmental hazards. Under physiological conditions, ageing leads to a progressive decline in balance control per se. Moreover, various neurological disorders further increase the risk of falls by deteriorating specific nervous system functions contributing to balance. Over the last 15 years, significant advancements in technology have provided wearable solutions for balance evaluation and the management of postural instability in patients with neurological disorders. This narrative review aims to address the topic of balance and wireless sensors in several neurological disorders, including Alzheimer's disease, Parkinson's disease, multiple sclerosis, stroke, and other neurodegenerative and acute clinical syndromes. The review discusses the physiological and pathophysiological bases of balance in neurological disorders as well as the traditional and innovative instruments currently available for balance assessment. The technical and clinical perspectives of wearable technologies, as well as current challenges in the field of teleneurology, are also examined.

**Keywords:** wireless sensors; wearables; balance; posturography; Alzheimer's disease; Parkinson's disease; multiple sclerosis; cerebellar ataxia; stroke; vestibular syndrome

#### **1. Introduction**

Countries are globally experiencing a demographic shift in the distribution of the population towards older ages [1] and every year up to 35% of people aged 65 and over fall, often requiring hospital admission after mild to severe injuries [2]. Falls account for 40% of all injury-related deaths [2], and even when non-fatal, commonly cause a "post-fall syndrome", a psychomotor regression condition responsible for psychological, postural and gait dysfunction in elderly [3]. In terms of the economic burden of falls, in 2015 the estimated medical costs attributable to fatal and non-fatal falls increased to 50 billion dollars in the United States [4]. Falls represent a major public health concern and have an enormous economic impact on society, thus requiring the development of effective strategies to prevent underlying causes. Among these, balance impairment is one of the leading determinants of falls along with ecological factors, such as environmental hazards [5]. Ageing significantly impacts on postural ability due to age-related changes in the sensorimotor and cognitive function [6]. Moreover, balance impairment frequently affects patients with neurological disorders who are twice as likely to fall compared to an age-matched healthy population [7].

To date, a history of falls is the strongest predictor of future falls [8,9], thus underscoring the need for predictive measures to determine early preventive interventions. However, clinical assessment is subjective and is not sensitive enough to identify early balance control dysfunction [10]. Conversely, traditional laboratory evaluation, including posturography through force platforms and optoelectronic systems, is objective and sensitive enough to identify subtle abnormalities but does not always reflect real-life situations. Over the last 15 years, advancements in healthcare technology have allowed analysing physiological measures of motor and non-motor behaviour objectively and unobtrusively [11]. Indeed, the availability of wearable devices has opened to the instrumental evaluation of clinical phenomena in free-living conditions. Accordingly, several authors have made a great effort to use wireless sensors in the study of balance impairment in patients with neurological disorders, thus offering new solutions for diagnosis and rehabilitation [12].

Despite several previous reviews discussing specific technical or clinical aspects of balance assessment through wearables, this narrative review aims to discuss the whole topic of balance evaluation, through wireless sensors, in patients with neurological disorders. Accordingly, in this review, we first introduce the physiology and pathophysiology of balance, including the main mechanisms underlying postural dysfunction in several neurological disorders, and report clinical tools commonly used for balance assessment. We then summarise the instrumental assessment of balance, including static and dynamic posturography. Moreover, we analyse wearable technologies available for balance assessment in neurological disorders. Finally, we speculate about prospects and challenges of wireless sensors for balance assessment in teleneurology and telerehabilitation.

#### **2. Physiology and Pathophysiology of Balance**

Balance is the ability to maintain body orientation in space under static and dynamic conditions [13], respectivelyintended as postural stability at rest andin response to activemovement or external perturbations. Over the course of evolution, the complexity of this function greatly increased with the acquisition of vertical posture and bipedalism in humans, representing the main transformation in primates [14]. A composite sensorimotor-control system based on a closed-loop circuit dynamically coordinates body segments according to environmental hazards through feedback and feed-forward strategies [15].

The central nervous system oversees balance maintenance by integrating sensory inputs from the peripheral nervous system (e.g., receptors and nerves) and motor outputs to the musculoskeletal system [15,16] (Figure 1). Brainstem nuclei, along with basal ganglia, the cerebellum, and other subcortical structures (e.g., thalamus) play crucial roles in the integration of sensory cues from the somatosensory, vestibular, and visual systems, which continuously provide an overall representation of body movement, acceleration, and position in space [15,17] (Figure 1A). By encoding an internal postural model based on reciprocal connections with the parietal cortex, the cerebellum contributes to dynamic balance control through postural responses that serve as an error-correction mechanism [16] (Figure 1B). Finally, the cerebral cortex oversees attentional and visuospatial balance requirements and manages anticipatory postural adjustments (APAs) before and during voluntary movements [18]. Cognitive-motor processes are responsible for postural optimisation based on prior experience, current context, and learning through long-latency components of postural responses [19].

**Figure 1.** Physiology of balance. (**A**) The visual system provides information on the surrounding environment; the vestibular system, consisting of the two inner-ear balance organs and several nervous structures (nerves and central nuclei), encodes angular and linear accelerations of the head to support the clear vision and balance control via rapid eye movements (vestibulo-ocular reflexes) and postural reflexes (vestibulo-spinal reflexes); the somatosensory system senses self-movement and body position through specialised sensory receptors located in the muscles (muscle spindles), joints (Ruffini endings, Pacinian corpuscles, and Golgi-like receptors), tendons (Golgi tendon organs), and skin (Merkel cells, Ruffini endings, Meissner corpuscles, and Pacinian corpuscles) [20,21]. (**B**) Multisensory signals from visual, vestibular and somatosensory receptors are integrated in the central nervous system to provide an internal postural model and in turn, descending motor commands to muscles. (**C**) Reactive postural strategies and anticipatory postural adjustments allow balance control under environmental circumstances (e.g., external postural perturbations) and motor initiative (e.g., voluntary movement), respectively.

The main goal of physiological mechanisms underlying balance control is the maintenance of postural stability by managing the spatio-temporal relationship between the body's centre of mass (COM) and base of support (BOS) [22]. While reactive postural responses compensate for unexpected external perturbations, proactive postural responses allow balance control under expected external perturbations or self-produced balance disturbances through a motor prediction strategy [22]. When an external balance perturbation occurs, different postural strategies are adopted to maintain the COM projection within the BOS. Indeed, minor postural perturbations are usually counteracted by corrective strategies involving body rotations around the ankle (ankle strategy) or hip (hip strategy) that move the COM projection. Conversely, major postural disturbances require a broadening or displacement of the BOS in order to maintain the COM projection within the BOS (protective strategy) [22] (Figure 1C).

Three main pathophysiological mechanisms are responsible for balance dysfunction: (i) abnormal acquisition, transmission, or perception of sensory signals (Figure 1A); (ii) abnormal sensorimotor integration and motor planning (Figure 1B); (iii) impaired transmission of motor output or musculoskeletal system damage [23] (Figure 1B,C). In patients with impaired afferent sensory information (e.g., somatosensory, vestibular or visual inputs), balance control requires compensatory strategies including attentional resources [24] and sensory reweighting [25].

Ageing is commonly associated with a progressive loss of sensorimotor function, including structural and functional changes in the somatosensory, visual, and vestibular systems, along with a decline in central neural processing and muscle strength [6]. Accordingly, ageing leads to slower reaction times and reduced limits of stability, thus worsening balance control mainly under cognitive loads and unexpected postural perturbations [6,26].

Patients with neurological disorders may manifest balance dysfunction as a result of impairment of at least one physiological component responsible for balance control significantly increasing the risk of falls compared to age-matched healthy subjects [7]. Pathophysiological mechanisms leading to balance impairment in various neurological disorders are summarized in Table 1 along with the main nervous system structures underpinning postural dysfunction. Understanding the physiological mechanisms underlying balance control in humans is the necessary background to measure balance objectively, through conventional as well as wearable technologies, in patients with neurological disorders.


 **1.** Balance impairment in neurological disorders.

**Table**



#### **3. Clinical Assessment of Balance**

The clinical assessment aims at recognizing balance impairment and identifying possible underlying causes [63]. Neurological examination routinely involves several clinical manoeuvres, including the Romberg's test [64], the pull test [65], and the tandem gait test [66], designed to examine individual balance performance qualitatively (Table 2). In addition to these clinical manoeuvres, several standardized scales and tests provide a semiquantitative evaluation of balance (Table 2). A secondary task during motor performance (i.e., dual task) is commonly used to assess the involvement of cognitive function in balance control.



When considering the clinical assessment of balance, several issues should be taken into account. First, the clinical assessment unlikely detects early postural abnormalities since it identifies balance impairment when significant pathological changes in the nervous system have already occurred. Second, the clinical assessment provides qualitative rather than quantitative evaluations of postural

ability, thus representing a subjective tool. Third, standardised clinical scales or indices, such as the Berg balance scale [75] or the dynamic gait index [76], are semiquantitative evaluations of balance, but are time-consuming and suffer from floor and ceiling effects. Lastly, the clinical setting usually involves rather predictable environments with poor ecological value. As a result, evaluation through instrumental tools, such as wearable sensors, would contribute to providing more sensitive, objective, multidimensional, long-term and ecological measures.

#### **4. Static and Dynamic Posturography**

Posturography refers to the instrumental assessment of balance [77–79] under static or dynamic conditions [80,81]. Static posturography examines body postural sway while subjects maintain a static stance on a non-movable surface [79,81]. During the upright stance, the human body can be considered an unstable system in which force gravity and body inertia generate torques to be balanced [82]. Indeed, the vertical projection of the whole body mass constantly varies over time, deviating from the ankle joint centre of rotation [83]. Human standing balance can be represented by a reduced number of joints resembling an unstable single-link inverted pendulum [84].

Unlike static evaluation, dynamic posturography includes several postural tests and ad-hoc instruments designed to assess balance under experimentally-induced external perturbations [85]. External disturbances are often designed to simulate environmental hazards occurring in daily activities including a set of visual and motor challenges [85,86]. Postural responses to external perturbations can be assessed by a non-motorised movable platform, such as the Biomechanical Ankle Platform System [87], or more complex commercial robotic systems, such as the Equitest system (Neurocom International, Clackmas, OR, USA) [85], the Balance Master (Micromedical Technologies, Chatham, IL, USA) [88], or Caren (Motek, Amsterdam, the Netherlands) [89].

Several non-commercial robotic platforms have been recently designed to provide various patterns of mechanical perturbation [90–94]. Common approaches include unidirectional [95] or multidirectional [85,96,97] disturbances, such as rotational [93,98–101] and translational perturbations [96,102–104], or forces applied to specific body segments [105,106]. Abrupt perturbations allow the examination of reactive postural responses, whereas continuous and oscillatory perturbations are used for the assessment of anticipatory postural strategies [101,102,104,107,108]. Postural perturbations can be also defined as predictable or unpredictable according to the subject's awareness. The predictability/unpredictability of a specific perturbation allows the experimental investigation of reactive or anticipatory postural strategies [105,106,109]. Mechanical perturbations are often merged with visual, vestibular, and proprioceptive disturbances such as visual scene movements, imposed head accelerations, galvanic vestibular stimulation, and tendon vibration [81,85,89,110,111]. The most common tests used are the Sensory Organization Test (SOT) [112], the Motor Control Test (MCT) [113], and the Adaptation Test (AT) [81]. In the SOT, subjects are elicited through visual, vestibular, and proprioceptive modifications of the support surface and visual surroundings to create sensory conflict conditions. The MCT consists of antero-posterior perturbations at different intensity levels, while in the AT subjects experience toes-up and toes-down rotations.

Several biomechanical parameters quantify balance dysfunction [114,115] by referring to two main variables: the centre of pressure (COP) and COM [116]. The COP is the application point of the total ground reaction force vector, whereas the COM refers to the average position in 3D space of all body segment positions according to their specific masses [116]. COM can be considered representative of the movements of the entire human body [116]. Several indices considering acceleration, velocity, displacement of single or multiple body segments, joint angles, and muscle activity can be measured using both traditional and wearable instrumentation (Table 3).


*Sensors* **2020** , *20*, 3247


**Table 3.** *Cont.*

*Sensors* **2020**, *20*, 3247

Overall, classical laboratory posturography through force plates and optoelectronic systems provides reliable, accurate, and comprehensive measurements for balance assessment. However, these techniques are generally expensive, encumbering, and also require supervised settings as well as technical expertise, thus precluding their use for long-term monitoring in daily life situations. Accordingly, current research on posturography has recently moved on wearable technologies [132–137] possibly providing objective, long-term and free-living monitoring of postural ability at a negligible cost.

#### **5. Wearable Technologies**

Recent advances in microelectronics have led to the production of small flexible sensors, even integrated into clothing ("e-textile") [138], thus making wearable devices suitable for free-living applications [139]. To date, the main wearable technologies available for balance assessment include mechanical devices, such as inertial and pressure sensors, and physiological devices, such as surface electromyography sensors (sEMG) (Figure 2). Wireless inertial sensors are the most used solution in wearable systems and have been widely adopted for balance and gait assessment [115,140–142]. Half of the previous studies used commercial inertial measurement unit (IMU) sensors including triaxial accelerometers and gyroscopes, and half adopted stand-alone accelerometers [143] or gyroscopes [144]. The combination of triaxial accelerometers, triaxial gyroscopes and magnetometers compose magnetic and inertial measurement units. Sensor placement depends on the specific postural task under investigation [115]. For instance, wearable sensors can be placed over the waist or trunk in order to measure postural sway and trunk acceleration. Other possible body locations include the lower limbs, sternum, upper limbs and forehead. Triaxial sensors can capture spatio-temporal and 3D kinematic data including joint and segment angles [145–147]. Overall, the combination of accelerometers, gyroscopes, and magnetometers provides accurate information on body spatial orientation and motion (Figure 2A). Besides inertial devices, wearable sEMG sensors evaluate specific patterns of muscle activation during static and dynamic postural perturbations. sEMG, therefore, allows a better understanding of physiological mechanisms responsible for balance control [148,149] (Figure 2B). Lastly, wearable pressure sensors are instrumented insoles placed or integrated into the shoe to measure pressure changes between the foot and ground [150]. The accuracy of this discrete sensor system is comparable to non-wearable technologies such as the laboratory force platform (Figure 2C). In addition to mechanical and physiological devices, there are wearable sensors able to continuously monitor the concentration of specific biochemical markers in biofluids, through miniaturized and flexible devices [151]. These innovative sensors would open to interesting prospects also referring to the assessment of balance. For instance, monitoring L-Dopa or dopamine concentration by microneedle patches would be a helpful tool to correlate postural ability with dopaminergic treatments in patients with Parkinson's disease [152,153]. Currently, several wearable sensors, mostly including inertial devices, are available on the market for approved clinical use in balance assessment [154], also including self-adhesive biosensors (for further details see www.clinicaltrials.gov).

The large volume of data produced by wearable sensors requires the development of specialised algorithms and machine learning algorithms to select clinically-valuable measures [138]. Owing to the considerable processing capacity of wearable devices, embedded algorithmic sets can be used for the online and remote execution, but at the expense of the battery charge duration. To optimise the performance of these algorithms in recognising clinical phenomena, a common approach leverages the so-called "sensor fusion", which consists of the combination of sensory data and signals derived from distinct sources so that the resulting information is more accurate (e.g., integration of inertial and electromyography signals) [155]. Accordingly, the emerging trends in wearables are moving towards the design of integrated sensors, including devices composed of IMUs and sEMG [148], to be user-friendly, waterproof and unobtrusive. Table 4 summarises the strengths, limitations and challenges of each type of wireless sensors currently used for balance assessment. Moreover, Table S1 reports all the previously published reviews on balance assessment through wearable devices in healthy subjects and patients affected by various medical conditions.

**Table 4.** Strengths, limitations and challenges of wireless sensors currently available for balance assessment.


**IMU**: Inertial Measurement Unit; **sEMG**: surface electromyography

#### **6. Literature Research Strategy and Criteria**

Literature research of studies investigating balance impairment through the use of wireless sensors in neurological disorders was performed using the following databases: MEDLINE, Scopus, PubMed, Web of Science, EMBASE and the Cochrane Library. Literature criteria included the following terms: "wireless sensors" OR "wearables" OR "inertial measurement unit" OR "surface electromyography" OR "pressure sensors" AND "neurological disorders" OR "Alzheimer's disease" OR "stroke" OR "Parkinson's disease" OR "multiple sclerosis" OR "vestibular disorders" OR "cerebellar ataxia" OR "traumatic brain injury" OR "Huntington's disease" OR "neuropathy" AND "balance" OR "posturography" OR "postural control." Eligible studies were experimental studies published from January 2005 to March 2020, examining balance through wireless sensors in patients suffering from the above reported neurological disorders. The reference lists of retrieved articles were also manually searched for additional studies. Reviews, reports, conference proceedings, and articles in languages other than English were not considered in the evaluation of eligible studies.

#### **7. Wearable Technologies in Neurological Disorders**

Previous studies using wearable sensors have investigated balance impairment in Parkinson's disease [114,122,124,125,156–168], multiple sclerosis [118,146,169–177], stroke [52,178–184], traumatic brain injuries [123,126,185–189], cerebellar ataxia [130,190–195], vestibular syndromes [196–199], neuropathies [199–201], Alzheimer's disease [32,202,203], and Huntington's disease [46,204]. Most of these studies have compared patients affected by neurological disorders with healthy subjects. However, a minority of authors [52,167,176,178,180,187] have analysed postural ability only in a group of patients with neurological disorders without including a control group.

Concerning the type of sensors used for balance assessment, most of the existing studies have applied inertial devices, primarily accelerometers and gyroscopes. Several authors [46,162,163,183,184] have even used inertial sensors installed in common tablet computers and smartphones. Conversely, no authors have used pressure sensors, while only a few have adopted wireless sEMG sensors [166–168] to analyse balance impairment in patients with Parkinson's disease. Strengths and limitations of each type of sensor are shown in Table 4. Each type of sensor technology would be implemented by addressing some challenges, including the elaboration of new algorithms, the development of implantable EMG tools and, finally, the use of unobtrusive "e-textile" devices (see Table 4). Also, future studies would benefit from the integration of various sensor technologies (i.e., sensor fusion) to optimize the measure of balance dysfunction in patients with neurological disorders.

Regarding the number and body location of sensors, authors have used 1 to 8 inertial devices and multiple body segments, including the upper (10 studies) and lower limbs (21 studies), head (1 study), trunk (18 studies), and waist (48 studies), depending on the static or dynamic postural task chosen for balance assessment. Indeed, some authors who investigated postural evaluation during gait (e.g., [122,146,161,169,172,175]) and instrumented versions of clinical tests, such as the push and release test [171] and the Fukuda Stepping Test [182], have usually applied more sensors than those evaluating static balance during upright stance (e.g., [52,114,125,157,158,163,177,183–185,188,191–194,199,203,204]. However, despite one study [204], all authors have included the lumbo-sacral region as the main location of inertial sensors for the analysis of postural sway, according to the COM position. Conversely, multiple sEMG sensors have been placed mainly on lower limbs to monitor muscle activity during postural perturbations [166–168]. The number of sensors and their placement on the body is a relevant issue for balance assessment, also requiring to consider a proper cost and energy-benefit analysis, as well as the efforts for patients and caregivers. The number of sensors to be used depends on the specific clinical phenomenon under investigation (e.g., postural sway for balance control) and the need for maintaining high-quality measurements, through appropriate sampling rate and estimated energy consumption. Indeed, though more informative, a high number of devices would be computationally demanding and expensive, as well as uncomfortable to be applied in a domestic environment.

Considering the accuracy of sensors in balance assessment, some authors [52,114,156,157,159, 162,164,171,173,174,186,191,193,194,200] have compared wearable device measurements with those of standardised laboratory measurement systems, such as force plates and 3D motion-capture systems. These authors have agreed on the moderate or strong correlation between specific inertial indices (e.g., root mean square of acceleration time series [114], acceleration peaks of anticipatory postural adjustments [156,159], time to reach stability [171]) and COP or optical measures, thus suggesting an accurate performance of inertial wearable devices compared to standardised instrumentations in the laboratory. However, validation studies in unsupervised settings are warranted to further support the reliability of wireless sensors for balance assessment in domestic environments.

Most authors [32,52,114,118,123–126,130,146,157,158,163,165,169,170,173,174,176–178,183,185–195, 197–204] have performed a static balance evaluation by analysing maintenance of the upright stance with different amplitudes of the BOS (e.g., side-by-side, tandem, single-leg stance). These protocols have also included the assessment of sensory and cognitive contribution to balance control by removing visual and/or proprioceptive cues (e.g., closed eyes, foam surface) and by increasing cognitive load (e.g., dual-task). Moreover, a large number of authors [46,118,122,146,156,159–162,164,166–169,171– 173,175,179–182,190,196] have investigated dynamic postural control, mostly through the use of walking tasks, instrumented versions of clinical tests (e.g., Timed-Up and Go, stand and walk, and push and release tests), and external or self-triggered postural perturbations. Although several authors [46,123,125,157,161,166–168,171–173,178–181,190,196,202] have assessed balance during tasks possibly reflecting daily postural challenges, all research protocols have been conducted in a laboratory setting. However, since supervised laboratory settings only partially reflect challenging "real-life" situations, these studies do not provide firm conclusions about the application of wireless sensors in a domestic environment.

Concerning biomechanical measures, previous studies have used filtered acceleration signals by inertial sensors to measure body sway in all the neurological disorders here considered, but have evaluated APAs during gait initiation only in patients with Parkinson's disease. Overall, these measures have shown increased postural sway in patients with neurological disorders and decreased APAs during gait initiation in patients with Parkinson's disease, as compared to age-matched healthy subjects. These parameters have also identified subclinical postural abnormalities (e.g., in vestibular syndromes) correlating with the amount of clinical disability [114,118,124,146,163,165,170,171,175,176,184,190,195]. A few authors [166–168] have measured muscle postural synergies with sEMG sensors in patients with Parkinson's disease. Given that no studies have directly compared biomechanical indices in patients with different neurological disorders, it is unclear whether any of the measures may discriminate the various conditions. These findings overall have shown that wireless sensors can accurately quantify several kinematic measures, including the time and frequency COM dynamics [114,174,200], the 3-D trajectory of body sway angles [191], the joint range of motion [205], the stepping latency [171], and the APAs [159]. Conversely, the evaluation of kinetic measures, including the analysis of internal forces and moments acting on human joints, by wearable systems remains quite challenging [206]. Although the novel approach by wearables would help to partially overcome this issue with inertial and pressure sensors, inverse dynamics techniques, through motion capture systems and force platforms, are currently more suitable to achieve these measures. Moreover, to date, other dynamic variables, including the joint power and the energy cost of a movement, have not yet been evaluated by wearable sensors. Specifically concerning APAs, in addition to inertial measurements, wearable technologies would also allow long-term APAs recordings, through wearable sEMG, in more ecological environments. However, APAs recordings through wearable sEMG would require advanced algorithms for pattern recognition to achieve consistent observation. A further consideration concerns the generalizability to more ecological environments of behavioural measures observed in the laboratory setting. Unlike motor performance under "real-world" postural perturbations, experimental measures under a supervised laboratory setting would improve per se patients' motor behaviour owing to unspecific and disease-unrelated factors, such as attentional and emotional aspects. The appropriate selection of a standardised measure for balance assessment would promote more consistent evaluation among the various neurological disorders. Table 5 provides an overall overview of the methodological approaches and findings from studies here examined. Also, a more detailed description of these studies is shown in Table S2. Finally, Figure 3 shows the positive trend of published studies on wireless sensors for balance assessment in the various neurological disorders.


*Sensors* **2020**, *20*, 3247


**Table 5.** *Cont.*



sEMG: surface test; UPDRS-III: Unified Parkinson's Disease Rating Scale—part III; VAF: variance accounted for; VS: patients with vestibular syndrome.

electromyographic

 sensors; SOT: Sensory Organisation

 Test; ST: patients with previous stroke; TBI: patients with previous traumatic brain injury; TUG: Timed-Up and Go

**Figure 3.** Trend of published studies on wireless sensors for balance assessment in neurological disorders. The figure includes studies reported in Table 5 and Table S2, based on the literature research through MEDLINE, Scopus, PubMed, Web of Science, EMBASE and the Cochrane Library (accessed on 31 March 2020). Note that the figure does not include a study published in 2005 [190] for graphical reasons.

#### **8. Teleneurology and Telerehabilitation for Balance: Prospects and Challenges**

Along with the ageing of the population, the prevalence of neurological disorders will also significantly increase in the next decades [207]. Accordingly, public health challenges will burden society and healthcare systems, which will face a heavy demand for the neurologic care of acute and chronic conditions. By allowing long-term monitoring for preventive and recovery strategies, wireless sensors will promote teleneurology and telerehabilitation and take some of the burden off of healthcare facilities.

Concerning the role of teleneurology for balance assessment through wireless sensors, so far, a few studies have addressed this topic in patients with neurological disorders. Nevertheless, several advantageous clinical prospects related to this issue should be considered. First, access to care for patients with balance impairment is quite challenging due to transportation difficulties and dependence on caregivers. Wireless sensors would be a sensitive and objective tool for the domestic measurement of balance control during the performance of validated instrumented tasks, such as maintenance of an upright stance. Moreover, other symptoms commonly associated with postural dysfunction, such as gait disorders [208], would also be measured, thus providing more detailed clinical information. Current evidence suggests that teleneurology promotes a reduction of patient and caregiver burden [209]. Second, medical visits in a hospital setting do not always reflect real-life situations, which commonly present insidious postural challenges. Therefore, the long-term monitoring of postural ability during common daily activities could provide ecological data on patient balance control in free-living conditions. This approach would help to identify early subclinical changes of balance, allow the objective assessment of fall risk and design individualised strategies for fall prevention (e.g., use of mobility aids and changes of environmental hazards). Third, the real-time identification of situations at high risk of falling would also allow patients to benefit from temporary preventive or rescue interventions. For instance, the detection of near-falls could be used for the automatic activation of protective tools, such as inflatable hip pads aimed to prevent fall-related injuries [210]. A further strategy would include the improvement of balance control by wearable-based sensory biofeedback, able to enhance patients' awareness and in turn, prevent falls [211,212].

Along with fall prevention strategies, rehabilitation is the main therapeutic approach for improving balance in patients with neurological disorders. The main goal of rehabilitation is to enhance individual postural skills, supporting patient independence in ecological settings. To this aim, by using information and communication technologies, telerehabilitation would provide rehabilitative services directly at home [213] with similar effectiveness to conventional therapy [214]. Wireless sensors would allow monitoring of individual postural ability in a domestic environment, increasing adherence to the rehabilitative programme, and thus promoting tailored therapeutic approaches [215]. Moreover, wireless sensors would also support home-based interactive rehabilitation programmes by providing real-time feedback during unsupervised training. Nowadays, the increasing use of mobile phones and other technological tools in multiple aspects of daily life is promoting a widespread technological education in the general population, including the elderly. Accordingly, in the next decades, user-friendly wearables will be increasingly used to increase adherence to telerehabilitation strategies. Owing to remote and continuous evaluation by physicians and physical therapists, telerehabilitation would reduce the number of periodic hospital admissions. However, some initial education to patients and caregivers concerning wearables applications for therapeutic purposes is likely required. So far, several clinical trials have already adopted sensor-based measurements to objectively evaluate balance and its response to pharmacological as well as non-pharmacological interventions [216] (for further details see www.clinicaltrials.gov). However, only a few authors [216–220] have examined the effectiveness of sensor-based balance training in patients with neurological disorders. Furthermore, most of these studies involved a laboratory or clinical setting supervised by experienced staff [216]. Hence, to reach some firm conclusion, new randomised controlled trials should assess large samples of patients in ecological settings, including the domestic environment [216].

The main current challenge is the technological migration of wireless sensors from the laboratory setting to a domestic and unsupervised environment. The technological feasibility of sensor systems primarily depends on the variables to be measured as well as on the computing-capacity integrated into the wearables. Unlike conventional laboratory systems, the domestic use of wearable sensors would imply some limitations such as autonomy and interface capabilities (e.g., interaction with the user, communication with external devices and servers for information sharing). Concerning IMUs, challenges include the calculation capacity, which mainly depends on the running algorithms thus influencing the selection of a specific device, processing characteristics, memory capacity and communication protocol. Overall, the technological migration of wireless sensors from the laboratory setting to a domestic environment would benefit from the identification of standardized and accurate measures. To this aim, understanding the physiological and pathophysiological mechanisms underlying balance is the background for selecting, measuring and interpreting the specific postural variables to be assessed. Also, the improvement of communication between wearable sensors and external devices, as well as the implementation of standardized and low energy-consuming algorithms are additional limitations to overcome. To support this migration process, current commercialization efforts are reducing sensor dimensions to ensure the unobtrusiveness of the devices, though maintaining safety and accuracy standards. "Real-world" evidence aimed at monitoring balance disorders through wireless sensors in ecological settings (e.g., patients' home or nursing home) will further clarify strengths and limitations in the telemedicine and telerehabilitation approaches.

Several open questions remain when considering teleneurology and telerehabilitation approaches. To date, only a few randomised controlled trials have addressed this topic in patients with neurological disorders, thus pointing to the weak internal validity of the current clinical evidence. Future studies should propose easier solutions to be applied in unsupervised settings without requiring technical expertise (e.g., issues related to data storage, access platforms and software/app usage). As a possible solution, machine-learning algorithms, including those using artificial neural networks (deep learning algorithms) [221], would be suitable tools for the automatic storage, interpretation and management of healthcare data [222–224]. Indeed, by learning from massive amounts of longitudinal data, machine-learning systems could lighten the burden of technical expertise and improve clinical decision making through a tailored approach. Another relevant point concerns some ethical issues, such as the security of the overwhelming amount of healthcare sensitive data derived from the use of wireless sensors, possibly leading to the generation of discriminatory profiles, manipulative marketing or data breaches [225]. Accordingly, limiting the wireless transmission to a small number of selected data (e.g., fall episodes) would help to preserve the confidentiality of a large amount of recorded information in case of privacy violation. Using proper encryption technology and increasing the users' awareness of privacy rights would help to address ethical issues. Nonetheless, strict regulations for data management should also be adopted to guarantee users' confidentiality and integrity [226]. The use of inertial sensors included in smartphones would address the issue of the cost and availability of wearable sensors [227].

#### **9. Conclusions**

Over the last 15 years, wearable devices have been largely used for the assessment of balance in patients affected by neurological disorders, providing valuable data compared with standard laboratory instrumentation. Indeed, a great experience in the use of wireless sensors for balance evaluation has been achieved in the laboratory setting. Conversely, much still needs to be done for the technological migration of wearable devices from the laboratory to the domestic unsupervised environment. This migration would open several valuable prospects, including teleneurology and telerehabilitation approaches.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/1424-8220/20/11/3247/s1, Table S1: previous reviews on balance and fall risk assessment through wireless sensors and Table S2: sensor-based balance evaluation in neurological disorders. References [32,46,52,114,115,118,122–126,130,146,147,156–204,211, 216] are cited in the supplementary materials.

**Author Contributions:** Conceptualization, A.S., A.Z. and I.M. (Ilaria Mileti); Methodology, A.S., F.I., J.C., A.Z. and I.M. (Ilaria Mileti); Electronic Database Search, A.Z., I.M. (Ilaria Mileti), E.P., C.C., M.P., A.M., I.M. (Ivan Mazzetta), G.D.C., L.L., C.P.-L. and J.C.; Quality Assessment, A.S., F.I., J.C., E.P., F.C. and L.L.; Writing – Original Draft Preparation, A.Z., I.M. (Ilaria Mileti), C.C., M.P., A.M., I.M. (Ivan Mazzetta), G.D.C., L.L., J.C. and C.P.-L.; Writing – Review & Editing, A.S., F.I., J.C., A.Z., I.M. (Ilaria Mileti) and A.M.; Supervision, A.S., F.I., J.C., E.P., F.C. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

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

#### **Abbreviations**


#### **References**


© 2020 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*
