**Reactive Postural Responses to Continuous Yaw Perturbations in Healthy Humans: The E**ff**ect of Aging**

#### **Ilaria Mileti 1,\*, Juri Taborri 2, Stefano Rossi 2, Zaccaria Del Prete 1, Marco Paoloni 3, Antonio Suppa 4,5 and Eduardo Palermo <sup>1</sup>**


Received: 31 October 2019; Accepted: 18 December 2019; Published: 20 December 2019

**Abstract:** Maintaining balance stability while turning in a quasi-static stance and/or in dynamic motion requires proper recovery mechanisms to manage sudden center-of-mass displacement. Furthermore, falls during turning are among the main concerns of community-dwelling elderly population. This study investigates the effect of aging on reactive postural responses to continuous yaw perturbations on a cohort of 10 young adults (mean age 28 ± 3 years old) and 10 older adults (mean age 61 ± 4 years old). Subjects underwent external continuous yaw perturbations provided by the RotoBit1D platform. Different conditions of visual feedback (eyes opened and eyes closed) and perturbation intensity, i.e., sinusoidal rotations on the horizontal plane at different frequencies (0.2 Hz and 0.3 Hz), were applied. Kinematics of axial body segments was gathered using three inertial measurement units. In order to measure reactive postural responses, we measured body-absolute and joint absolute rotations, center-of-mass displacement, body sway, and inter-joint coordination. Older adults showed significant reduction in horizontal rotations of body segments and joints, as well as in center-of-mass displacement. Furthermore, older adults manifested a greater variability in reactive postural responses than younger adults. The abnormal reactive postural responses observed in older adults might contribute to the well-known age-related difficulty in dealing with balance control during turning.

**Keywords:** aging; reactive postural responses; yaw perturbation; kinematics; postural stability; dynamic posturography

#### **1. Introduction**

Falls are among the most common leading causes of accidental death, hospitalization, or injuries, such as broken bones, head, and spinal cord injuries [1]. Since falling still represents a challenging social, medical, and economical matter [2], several research efforts have been spent in understanding the mechanisms leading to falls in the elderly [3]. A possible theory for explaining the increased risk of falling in the elderly population is that age compromises balance capability. However, understanding the effects of aging on balance is still an unanswered question.

Exploring dynamic balance control of upright stance through imposed external perturbations has the main objective to identify deficit in reactive postural responses. According to the amplitude of COM dislocation and the intensity of perturbation, different reactive strategies, such as hip and ankle strategies and/or stepping responses, prevent imbalance [4,5].

Various imposed posture manipulations, such as mechanical, visual, and vestibular perturbations, have been designed to reproduce free-living imbalance conditions and to investigate risk of fall [6–9], especially in elderly [10–14]. In a mechanical perspective, popular approaches focused on rotational [8,13,15–17] and/or translational support surface movements [9,18,19] or on body release paradigms [20,21] to simulate common postural disturbances (e.g., standing on a bus, slipping on a slippery surface, and falling due to a sudden boost). In this context, the majority of the literature studies focused on the postural responses under forward/backward perturbations to assess risk of fall in the elderly. Across this range of perturbation paradigms, older adults exhibited greater difficulty than young adults in recovering loss of balance through protective postural strategies, such as stepping, especially in case of a forward fall. Accordingly, a lower maximum body lean angle [20], smaller peak knee extensor torques, and larger peak extensors torques at hip and ankle joints [21] were reported in older adults during step response in forward perturbations. Conversely, when considering backward perturbations, older adults showed shorter reaction times than during forward perturbations to avoid fall [22].

Age-related postural deficits were also investigated considering different perturbation paradigms such as sinusoidal translating perturbation in the anterior-posterior direction [14], random rotations around sagittal axis [17], slippery floor surface [23], and mixed forward/backward and left/right platform translations [24]. More specifically, Nardone et al. found a greater head stabilization strategy and a looser coupling between head and hip motion in the eyes opened condition in older adults [14]. While, in the study of Cenciarini et al., older adults presented significantly higher active stiffness as compared to young adults to maintain body balance and to counteract destabilizations effects [17]. Moreover, impaired motor patterns in older adults have been also linked to an abnormal postural sway under forward/backward and left/right perturbation directions as attested in the study of Liaw et al. [24].

Although much has been learned about postural responses of older adults exposed to the before mentioned perturbation paradigms, the motor strategies involved in maintaining balance under perturbations around the vertical axis are still unclear. In healthy subjects, reactive postural responses during rotational perturbations around the vertical axis imposed by a robotic platform have been investigated only by few studies [10,11,15]. Earlier postural responses of the distal compared to the proximal body segment were observed both in kinematics [10] and in muscle activity [11]. The evaluation of reactive postural responses in term of COM displacement control, coordination of body segments and body sway response during this specific rotation would offer an ecological analysis of fall-inducing factors in everyday life. In fact, in the elderly, falls frequently occur within a narrowing familiar environment, such as the kitchen and the bathroom, during activities required for basic mobility, such as turning in place [25].

Over the last decade, measuring motor impairments through portable and wearable devices have demonstrated to be of great impact in managing neuromotor and aging deficits [26–28]. In fact, the several advantages of wearable sensors, such as the low cost, the high portability, the limited size and the ease-of-use, encourage the use of these technologies in clinical setting as a useful tool for monitoring assessment [29,30]. In this context, the aim of this study is to investigate the effects of aging on reactive postural responses to rotational perturbations around the vertical axis, by using a rotating platform and wearable inertial sensors in different visual conditions. Clarifying changes of dynamic postural control in older adults during rotational perturbations around the vertical axis would help to understand the mechanisms leading to falls when turning. This in turn, would be useful for the design of effective strategies for falls prevention.

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

#### *2.1. Subjects*

A cohort of ten healthy older adults (six females and four males, mean age 61 ± 4 years old, mean body mass 68 ± 9 kg, and mean height 170 ± 6 cm) and a cohort of ten young adults (five females and five males, mean age 28 ± 3 years old, mean body mass 63 ± 14 kg, and mean height 169 ± 11 cm) were enrolled. All participants were community-dwelling, medically stable, and able to walk and stand independently without aids. Subjects with intellectual, vestibular and/or visual deficits, neuromuscular diseases, orthopedic and/or neurological surgery interventions in the last three years were excluded from this study. All participants gave written consent before being included in the experimental session. The protocol was designed and conducted in accordance with the Ethical Standard of the 1964 Declaration of Helsinki.

#### *2.2. Experimental Setup*

The RotoBit1D was used in this study to provide sinusoidal perturbations around the vertical axis [15]. It is a rigid, round, flat robotic platform with a diameter of 0.5 m that allows a comfortable upright bipedal stance without narrowing feet. The mechanical design consists of a servo motor (SANYO DENKI) with maximum torque of 1.96 Nm, an incremental encoder, a toothed belt (PowerGrip HDT), speed reducer, and a polyethylene rotating disk. The robotic platform was computer-controlled by an ad-hoc LabVIEW software program (v.2014, National Instruments, Austin, TX, USA). Two sinusoidal perturbations around vertical axis were designed with fixed peak amplitude of ± 55◦ C and frequencies of (i) 0.2 Hz and (ii) 0.3 Hz, namely lower (L) and higher (H) frequency, respectively. The peak angular velocity was 80 ◦C/s and 100 ◦C/s for the lower and higher frequency, respectively. To avoid sudden variation in starting/stopping velocity, a sigmoidal wave was added at the start/end of the sinusoidal trajectory. In pilot trials, we chose rotation parameters as those able to provide a high perturbation intensity without requiring stepping response of subjects. Inertial Measurement Units (IMUs) (MTw, Xsens Technologies—NL) including a 3-axes accelerometer (<sup>±</sup> 160 m/s<sup>2</sup> FS), a 3-axes gyroscope (± 1200 ◦C/s FS), and a 3-axes magnetometer (± 1.5 Gauss FS) were used for gathering kinematic data of pelvis, trunk, and head. More specifically, the pelvis-sensor was placed centered on the median sacral crest and just below the anterior sacral promontory. The trunk-sensor was placed under the suprasternal notch on the sternum body; while, the head-sensor was placed on the frontal bone over the superciliary arch. Each subject was instrumented by the same expert operator, to guarantee consistent sensor location on body segment.

IMUs were placed by means of suitable elastic belts to avoid relative movement between sensor and body. Sampling frequency was set at 40 Hz. Equipment was simultaneously triggered at both the beginning and the end of each acquisition. More specifically, the ad-hoc LabVIEW software was designed to simultaneously drive the RotoBit1D servo motor and provide an external trigger, i.e., a square signal ranging from 0 to +5 V, to the IMUs Awinda Station through an NI USB-6212 DAQ board. The Awinda was set to start and end the IMUs' acquisition on the rising and the falling edge of the trigger signal, respectively.

#### *2.3. Experimental Procedure*

The experimental protocol was conducted at the Department of Physical Medicine and Rehabilitation, Sapienza University of Rome, Italy. Before each session, all tested subjects were asked to perform a Functional Calibration procedure advised by an operator. The FC procedure provided sensor orientations with respect to body segment, to complete the body-to-sensor alignment procedure [31]. The FC procedure consisted of a standing and sitting task, each lasting 5 s. Afterwards, all subjects stood in a comfortable upright bipedal position with vertically hanging arms and externally rotated feet at a preferred angle with a symmetrical placement, on the top center of the robotic platform. All subjects were asked to wear heelless shoes. The experimental procedure included two different

perturbation frequencies (lower and higher frequency) and two different visual conditions (eyes opened (EO) and eyes closed (EC)). More specifically, the participants' reactive postural responses were measured under four balance tasks: standing with EO during platform rotation at (i) lower (EO-L) and (ii) higher (EO-H) frequencies; and standing with EC considering both (iii) lower (EC-L) and (iv) higher (EC-H) platform frequency rotations. In the EO condition, subjects were asked to stare at a fixed point placed on the wall at 2 m from the platform.

Each task was performed three times randomizing the task order across subjects to avoid bias in results due to similar task sequences. In addition, subjects were not advised about frequency, to avoid habituation of postural responses or anticipatory strategies due to predictability [32].

#### *2.4. Data Analysis*

All data were analyzed off-line using MATLAB (v.2015b, MathWorks, Natick, MA) program. Angular rotations around the vertical axis, i.e., yaw angles, of the platform (pt), the head (h), the trunk (t), and the pelvis (p) segments were considered for the data analysis. Comparisons between rotation of the platform and body segment rotations in the transversal plane were obtained via the fast Fourier transform analysis, by considering Gain ratio (G) and phase shift (ϕ) indices, akin to [15]. G was computed as the ratio between the maximum amplitude value of the fundamental wave of the first signal and the amplitude value of the second signal at the same frequency. ϕ was obtained as the difference of the phase angles of the Fourier transform of the two signals at the frequencies having the maximum amplitude in the Fourier domain. Before transforming in Fourier domain, each body segment rotation in the transversal plane was demeaned. Among angular rotation around the vertical axis of the three body segments and platform, 6 pairs were considered for the analysis: platform-head, platform-trunk, platform-pelvis, head-trunk, trunk-pelvis, and head-pelvis. The following nomenclatures were chosen for G and ϕ indices to indicate the 6 pairs:

*f Gs*, *<sup>f</sup>* ϕ*<sup>s</sup>* (1)

where f and s represent the first and second sine waves considered, respectively. Perfect agreement in amplitude and timing between the first and the second element of the pair were observed considering a G value close to 100% and a ϕ value close to 0 ◦C. Instead, an anticipation/delay in phase angle of the second sine wave compared to the first one was defined as a positive/negative phase shift. For sake of clearness, *ptGh*, *ptGt*, *ptGp*, *pt*ϕ*h*, *pt*ϕ*t*, and *pt*ϕ*<sup>p</sup>* were addressed as G-absolute and ϕ-absolute, because each of the yaw body angles (second sine wave) is referred to the yaw platform angle (first sine wave). While, *hGt*, *<sup>t</sup> Gp*, *hGp*, *<sup>h</sup>*ϕ*t*, *<sup>t</sup>* ϕ*p*, and *<sup>h</sup>*ϕ*<sup>p</sup>* were addressed as G-relative and ϕ-relative, because both the first and the second sine wave are referred to a body segment sine wave.

To assess inter-joint coordination on the transversal plane, the continuous relative phase (CRP) technique was computed on yaw body angles, i.e., body rotation on transversal plane [33]. The CRP analyzes the differences in phase angles of two body segments during a particular motion task. Differently from ϕ index, which provides the average phase shift of the specific sine wave above described, the CRP refers to all the frequency components of a signal, reporting coupling behaviors of two body segments over the entire motor task. Concerning the CRP analyses, the phase space usually consists of the time-dependent measured signal and its first derivative. Thus, the CRP for a particular task is obtained as the four-quadrant arctangent phase angle from the phase space. Several methods have been developed for the calculation of the phase angles based on phase portrait analysis or the Hilbert transform [33–35]. Among those methodologies, the Hilbert transform-based method has been proved to be more robust than the phase angles in performing the phase portrait, especially regarding none-purely sinusoidal signals [33]. The Hilbert transform allows the transformation of any real signal into complex, analytic signal according to:

$$\mathbb{X}(t) = \mathbb{x}(t) + iH(t) \tag{2}$$

at time *ti*, the phase angle can be computed by:

$$\mathcal{O}(t\_i) = \arctan\left(\frac{H(t\_i)}{\varkappa(t\_i)}\right) \tag{3}$$

The continuous relative phase CRP between two signals can be defined as the differences of the phase angles:

$$\text{CRP}(t\_i) = \theta\_1(t\_i) - \theta\_2(t\_i) = \arctan\left(\frac{H\_1(t\_i)\mathbf{x}\_2(t\_i) - H\_2(t\_i)\mathbf{x}\_1(t\_i)}{\mathbf{x}\_1(t\_i)\mathbf{x}\_2(t\_i) + H\_1(t\_i)H\_2(t\_i)}\right) \tag{4}$$

where H1(t) and H2(t) are the Hilbert transform of signals of the proximal and distal segment, respectively. The CRP index can assume values between 0◦ and 180◦. Values close to 0◦ indicate in-phase coupling of the two segments, while values close to 180◦ represent an out-phase coupling of signals.

To identify differences in inter-joint coordination between young and older adults, the mean absolute relative phase (MARP) was computed by averaging the absolute values of the curve points considering the overall trial duration [36], as in the following equation:

$$\text{MARP} = \frac{\sum\_{i=1}^{p} \left| \overline{CRP\_i} \right|}{p} \tag{5}$$

where *p* is the number of time points in each trial. Similar consideration done for the CRP can be adopted for the MARP value.

Furthermore, the Deviation Phase (DP) was analyzed in order to assess variability among trials in inter-joint coordination [36]. The DP can be assessed by averaging the standard deviation among trials of the CRP(*ti*) over the trial duration, as in the following equation:

$$\text{DP} = \frac{\sum\_{i=1}^{p} SDi}{p} \tag{6}$$

where SDi represents the standard deviation of the CRP among the three trials at the *i*-th time instant. The DP is a measure of the stability organization provided by the neuromuscular system [36]. DP values close to 0 ◦C attest less intra-subject variability of the inter-joint coordination.

In order to assess postural control of body motion in the anterior-posterior and medio-lateral directions in response to external balance perturbations, body displacements of the head, the trunk and the pelvis were estimated via a strap-down integration of the acceleration signal, similarly to [37]. Body displacement of pelvis-sensor obtained through this method can be addressed as an estimation of the COM displacement, as authors reported [37]. However, differently from [37], in our study, we used the rotation matrix obtained from the quaternion output of the IMUs for the strapdown integration. In the original work [36] the authors rather used the gyro output for obtaining an estimation of the rotation matrix, as the system available for them did not provide the quaternion as an output.

To estimate the displacement, the acceleration signal was firstly rotated in the global coordinate frame to remove gravitational acceleration. After the gravitational acceleration removal, the acceleration signal of the inertial sensor in the global coordinate frame was straightforwardly integrated. Velocity was then high-pass filtered and the displacement was obtained through a second integration and filtering. The applied filters were zero-lag first-order Butterworth filters with a cut-off frequency of 0.2 Hz for the anterior-posterior (AP) and medio-lateral (ML) components and 0.5 Hz per the vertical (V) component.

Considering the body displacement, the following kinematic parameters were obtained for the statistical analysis: (i) the range of motion of the body segment displacement in the ML (RoMML) and AP (RoMAP) directions expressed in mm; (ii) the total path length of the body displacement normalized to the task duration (PATH) and (iii) the maximum velocity of the displacement (MV) expressed in m/s. The before mentioned parameters were considered for the head, trunk, and pelvis displacements.

#### *2.5. Statistical Analysis*

All data were tested for normality by means of the Shapiro-Wilk test. Statistical analysis was performed with the SPSS package (IBM-SPSS Inc., Armonk, NY, USA). Two unpaired t-tests were conducted to assess differences in body mass and height between the two groups. In order to test differences in postural strategies induced by aging and different perturbation frequencies, a 2 × 2 two-way mixed ANOVA, with AGE as a between-subject factor (two levels: young adults and older adults) and FREQ (two levels: lower and higher frequency) as a within-subject factor was used separately for the EO and the EC conditions. When the assumption of sphericity was violated, the Greenhouse-Geisser correction was considered. A paired *t*-test within each group and an unpaired t-test between the two groups were performed when the interaction between the main effects was significant. The Bonferroni corrections were considered for all the statistical analysis. The significance level was set at 0.05 for all the statistical tests.

#### **3. Results**

All subjects were able to complete the experimental procedures without losing balance and/or experiencing fatigue.

No statistical differences were found in body mass and height between groups.

Considering all statistical analysis, the 2 × 2 two-way mixed ANOVA reported no significant interactions between main effects AGE and FREQ.

• G- and ϕ-absolute:

In Table 1, mean and standard deviation of G-absolute and ϕ-absolute values and *p*-values are reported.

With regards to AGE main effect in the EO task, older adults showed a significant lower G-absolute value of trunk and head than young adults. The pelvis body segment of young adults reached the highest mean values of G-absolute in both frequency conditions (57.04% and 48.00%), see Table 1 and Figure 1. Compared with younger population, older adults exhibited a smaller amount of the axial body motion. In the EC condition, a similar reduction of motion amplitude was found in G-absolute values of head and pelvis of older subjects. Considering the main effect FREQ, upper body G-absolute values statistically decreased as a function of rotation frequency increment, regardless of visual condition for all body segments.

Concerning AGE main effect, no differences were found in the EO condition related to ϕ-absolute. While the main effect of FREQ was found statistically significant for all the body segments. In the EC task, ϕ-absolute values of the pelvis were statistically different between young and older adults, according to AGE main effect. To face external yaw postural perturbations, older adults anticipated pelvis motion while younger adults adopted a delayed motion strategy. Head and pelvis phase shifts were found statistically significant in the main effect FREQ.

• G- and ϕ-relative:

In Table 2, mean and standard deviation of G-relative and ϕ-relative values and *p*-values are reported. In the EO task, the main effect AGE of the G-relative values were found to be statistically different for all the considered segment-couples. More specifically, young adults reported G-relative values close to 100% attesting a similar amplitude pattern between the rotation around the vertical axis of the proximal and the distal segment. By contrast, older adults exhibited lower values of G-relative attesting a reduction in amplitude pattern of the distal body compared to the proximal. As regard to the main effect FREQ, statically differences were found for all the considered body segment-couples. In the EC condition, similar trend of the EO condition was found for the main effect AGE. While regarding the main effect FREQ, statistical differences were found for the head-pelvis and trunk-pelvis couples, attesting that the amplitude value of the distal segment decreased as a function of frequency perturbation increment.

**Figure 1.** (**a**) Mean (solid line) and standard deviation (dashed lines) of pelvis angle in the transverse plane for the eyes opened condition during the low frequency task. (**b**) Mean (solid line) and standard deviation (dashed lines) of pelvis angle in the transverse plane for the eyes opened condition during the high frequency task. The orange curves refer to young group, the blue ones to the older adults while the black curve is platform trajectory.

Regarding ϕ-relative, significant differences were found in the head-pelvis couple between young adults and older ones, in EO condition. More specifically, the distal segment of older adults appeared delayed with respect to the proximal one, while in the young group, synchronization strategies were adopted between segment couples. Increasing frequency motion, a higher delay was observed in the distal segment, especially in older adults.

• Continuous relative phase: MARP and DP:

In Table 3, mean and standard deviation of MARP and DP indices expressed in [◦] are reported. In the EO task, age-related differences were found for all the segment-couples. Higher MARP values were reached by older populations who preferred a more anti-phase motion strategy in response to balance rotational disturbance. An in-phase coupling strategy was observed in the young subjects. Increasing perturbations intensity, similar in-phase motion strategy was observed for all segment-couples. Considering the EC condition, a similar trend was found both regarding AGE as main effect. While considering FREQ as main effect, a more anti-phase strategy was observed in both groups when increasing perturbation intensity.

As regards coordination variability, older adults exhibited higher values of DP in all the segment-couples compared to young adults, regardless of the task condition (see Figure 2). Although the aforementioned results reported a wide amount of motor engagement, with the increase perturbation intensity, coordination variability of the trunk-pelvis couple increased as well in both visual conditions.

• Body displacement: RoMML, RoMAP, PATH, and MV:


ϕ-absolute values expressed in [◦] concerning head, trunk, and pelvis,

**Table 1.** Mean and standard deviation of G-absolute values expressed as a percentage [%] and

**Table 2.** Mean and standard deviation of G-relative values expressed as a percentage [%] and ϕ-relative values expressed in [◦] concerning head-pelvis, trunk-pelvis, and head-trunk, for both lower (L) and higher (H) frequencies and both visual conditions, i.e., eyes opened (EO) and eyes closed (EC) in young and older adults. *p*-values of the main effects, AGE and FREQ, of the two-way mixed ANOVA are reported. Statistical differences (*<sup>p</sup>* < 0.05) in the main effects are single-starred and reported in bold.


**Table 3.** Mean and standard deviation of the mean absolute relative phase (MARP) and Deviation Phase (DP) index of the continuous relative phase (CRP) technique,both expressed in [◦] concerning head-pelvis, trunk-pelvis and head-trunk, for both lower (L) and higher (H) frequencies and both visual conditions, i.e., eyes opened (EO) and eyes closed (EC) in young and older adults. *p*-values of the main effects, AGE and FREQ, of the two-way mixed ANOVA are reported. Statisticaldifferences (*<sup>p</sup>* < 0.05) in the main effects are single-starred and reported in bold.


**Figure 2.** (**a**) and (**b**) are phase portraits of the trunk and head body segment of one healthy young subject and one older subject during EO condition at low frequency, respectively. (**c**–**f**) Mean (solid line) and standard deviation (dashed lines) of CRP of the head-trunk couple for the EO and EC condition for the low frequency task and EO and EC conditions for the high frequency task, respectively. The orange curves refer to the young group, the blue ones to older adults.

In Table 4, mean and standard deviation of upper body displacement measures and *p*-values of the mixed ANOVA are reported. In EO condition, RoMML, PATH, and MV of the head and pelvis displacements were statistically lower in older adults compared to young adults. Regarding FREQ as a main effect, statistical differences were found considering PATH, MV in all upper body segments. Additionally, frequency increment effect was also found in RoMML and RoMAP in trunk displacement. A similar trend was reported in the EC task for both main effects.


**Table 4.** Mean and standard deviation of

medio-lateral

 (ML) and

anterior-posterior

 (AP) components

 of range of motion (RoM) expressed in [mm], the total path

#### **4. Discussion**

In our study, age-related changes were found in balancing continuous yaw perturbations, regardless of visual condition and perturbation frequency. A more conservative and less destabilizing motion was observed as a postural strategy in the older population, suggesting that older adults compensate for their reduced physical capabilities by becoming more cautious while performing a postural task.

• G- and ϕ-absolute:

As the fast Fourier analysis on yaw segment angles reported, young adults showed a matching strategy between the body displacement and the platform rotation, taking into account the amplitude of the trajectories. Conversely, older adults exhibited a smaller amount of motion amplitude of the upper body to counteract an unbalanced rotational state. This reduction in amplitude was observed in all the upper body yaw angles in older adults, especially regarding the distal segments. As observed in our results, this stabilization strategy progressively decreases from proximal to distal segment, becoming noticeable especially in the head motion of the older adults when perturbation frequency increases. In this context, the different balance response of adults could be caused by different perceptions of the vestibular system. Perceiving risky head rotations, the vestibular system tends to modify body dynamics by minimizing head oscillations, regardless of visual information. Moreover, avoiding visual feedback, a delayed motor strategy of the proximal segment was found in comparison with the platform motion in the older population, attesting the incapability of older adults to manage perturbation with anticipatory adjustment of the axial body.

As reported in literature, continuous and predictable perturbations are more easily managed compared to discontinuous and impulsive perturbations [18]. However, contrarily to sudden perturbations, during continuous perturbation subjects blend two primarily postural mechanisms. Subjects focused on the adjustment of their body motion in accordance with the actual external perturbation [18]. Moreover, subjects anticipate the body motion predicting the mechanical effect of the reflexes triggered by the displacement itself. In this context, pursuing the external movement of the perturbation by complying balance response with the platform displacement was observed to be the most functional and less-expensive postural strategy adopted by healthy younger adults [14,18]. Similar postural strategies were observed in our results by healthy younger subjects in balancing continuous yaw perturbations.

• G- and ϕ-relative:

By considering the fast Fourier analysis on segment-couples, proximal and distal segments of the younger group were in perfect agreement regarding amplitude rotation as well as time shift. Conversely, older population motor response reported a reduction and a delay in distal segment rotation in comparison with the proximal one, especially considering the head-pelvis segment-couple. Basically, younger subjects were able to oscillate in accordance to platform motion. As a consequence, younger subjects adopted a strategy that minimized the active effort by generating a lower torque couple among body segment. Older adults instead reacted differently from young adults, adopting more complex motor control strategies targeted to head stabilization. These findings can be ascribed to the greater difficult of older subjects to manage head-trunk movements, as also demonstrated during walking tasks [38].

• Continuous relative phase: MARP and DP

As regards the coordination pattern, the CRP technique provides a measure of the coupling or the phase relationship between the actions of body segment-couples. Since coordination impairment represents a common sign among musculoskeletal [39] and neurological disease [40], the age-related differences in the coordinative compensatory strategy were investigated in this study. Segment-couple coordination patterns of the upper body were different in the two groups. During EO condition, compared to the older adults, the younger population responded coupling body segments across the trial and reporting a less variable intra-subject trial-to-trial relationship between the actions of the segment-couples. Older adults also tended to balance yaw perturbation moving joints in and out-of-phase opposite fashion with an increased variability among trials. These outcome can be justified by the well-known loss of spatio-temporal coupling of muscles during postural responses, as reported in [41].

Increased variability in coupling relationship due to the aging was already reported in Yen et al. [42] during the obstacle-crossing task. This age-related biomechanical modification is associated to a lower ability of older adults to maintain a stable body displacement when crossing obstacles with different heights. A similar trend was observed in our results when subjects shifted from eyes open to the eyes closed condition. The segment-couple coordination of the upper body assumed more in-phase motion behaviors. Avoiding out-phase segment movements might be a coordination strategy adopted in the process of mastering redundant degrees of freedom of the upper body in a more unsafe scenario.

#### • Body displacement: RoMML, RoMAP, PATH, and MV:

In terms of displacement of body segments, the analysis of the upper body unveiled that subjects behave as a double inverted pendulum, mainly in the frontal plane. As Figure 3 shows, both groups tended to assume a more oscillating displacement in the medio-lateral fluctuation of the head and the pelvis body segment with respect to the trunk. In particular, when visual information was allowed, head and pelvis medio-lateral movements were found to be both in a wider range, while the trunk stood noticeably more stable, acting as a center pivot. The upright balance was continuously reach by actively counteracting the head body segment with respect to the pelvis one, as they appear in almost perfect phase opposition (see Figure 3). In Figure 3, results of the eyes opened condition during the high frequency task were reported. However, a similar trend was observed in all the experimental tasks. Despite subjects were challenged with a rotational perturbation in the transversal plane, results highlighted motor behaviors similar to those of the double inverted pendulum already observed in previous studies, in which subjects were exposed to purely translational perturbations. This aspect could be justified considering the effect of the centrifugal force. In our interpretation, the rotational perturbation generated a centrifugal force, which tended to displace the body center of mass in medio-lateral direction. Subjects counteracted by translating head in the opposite direction. Head trajectory, in fact, appeared to be perfectly in opposition with respect to the pelvis one (see Figure 3), similarly to what happens due to the hip strategy in the case of translational perturbations [6].

Despite the before-mentioned trend reported in both groups, a prominent age-dependent stiffening strategy was witnessed by a smaller amount of medio-lateral oscillation, a lower path length and a lower mean velocity of the pelvis and head displacements. As previously reported in the literature, in the upright balance, humans can be sketched as an Acrobot [43], consisting in a series of inverted pendulums related to feet, legs and torso. When a large deflection of the base of support occurs, a quick shift of the COM is required to maintain the upright stance. In a human model, a proper displacement at the hip level modulates the torque created by the gravitational force on the shifted body, allowing the subject to keep the COM over the base of support and the foot flat on the ground [44]. By approximating the human torso as a single link, the head displacement acted in the opposite direction, allowing the generation of the counteractive moment of the gravitational head force.

**Figure 3.** Mean (solid line) and standard deviation (dashed lines) of head, trunk, and pelvis displacement in the eyes opened condition during the high frequency task. The orange curves refer to the young group, the blue ones to older adults. In young group, medio-lateral sway of head is in phase opposition with respect to pelvis. Both have higher amplitude with respect to trunk, which acts as a pivot. This strategy is less evident in older subjects.

Interestingly, a different motor strategy was observed when vision was denied. The absence of visual inflow changed the body motion strategy adopted to keep the balance. Since the hip strategy was apparently preserved, the head displacement in the sagittal plane appeared more evident, overshooting pelvis motion in both groups. As a consequence of visual condition, the counteraction of body inertia was mainly in charge of the somatosensory and vestibular reference control, which reported a less effective role compared to the visual reference control in maintaining head stabilization. Similarly to findings in Nardone et al. [14], in which subjects were exposed to continuous anterior-posterior external perturbation, head oscillations were found more prominent during eyes-closed tasks. On the contrary to [14], older adults stimulated by continuous yaw perturbation presented a lower amount of head displacement compared to younger population, both regarding the eyes-open and the eyes-closed conditions.

By summarizing, our results highlighted age-dependent differences in rotational, translational, and coupling motor behaviors of the upper body of subjects elicited by external yaw perturbations. Future developments could be focused on the design of rehabilitation programs targeted to restore those impaired motor strategies. Those programs will be beneficial in reducing the risk of fall in older adults, especially those occurring during turning tasks, which still represent the 13% of all real-life falls [45].

Although this study provides insight into contributing factors to manage unbalanced conditions, some limitations should be taken into account. The small sample size and the large variability in the phase shift parameters could have affected the statistical analysis biasing results. In a future study, it will be necessary to increase the number of the participants to enforce statistical results regarding age-related differences, and design rehabilitation program aimed to enhance motor deficit of older population.

#### **5. Conclusions**

In this paper, we addressed the question of how aging affects postural control by examining the kinematic response under external yaw perturbation. The fundamental age-related postural change was mainly observed in the head stabilization strategy. Outcomes of older adults reported a decreased amount of the rotational and the translational body motion with a tendency of delayed behaviors and an out-phase and highly-varied coordinative compensatory strategy of the upper segment-couples. During low-frequency perturbation, postural strategies implied an easier following-strategy of the platform motion reporting a more complex reflex response than an anticipatory postural adjustment. During a more demanding frequency yaw rotation, a counteractive motor response was observed by the stiffening of the upper body motion and by strongly acting on the body inertial control.

**Author Contributions:** I.M., J.T., S.R., A.S. and E.P., designed the experiments; I.M., J.T., and E.P. performed the experiments; I.M. and E.P. processed kinematic data; I.M., J.T., S.R., Z.D.P., M.P., A.S. and E.P. analyzed data and interpreted results; I.M. wrote the first draft. I.M., J.T., S.R., Z.D.P., M.P., A.S. and E.P. reviewed and approved the manuscript. 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**


© 2019 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* **Exploring Risk of Falls and Dynamic Unbalance in Cerebellar Ataxia by Inertial Sensor Assessment**

#### **Pietro Caliandro 1, Carmela Conte 2, Chiara Iacovelli 2,\*, Antonella Tatarelli 3, Stefano Filippo Castiglia 4, Giuseppe Reale <sup>5</sup> and Mariano Serrao 4,6**


Received: 6 November 2019; Accepted: 12 December 2019; Published: 17 December 2019 -

**Abstract:** Background. Patients suffering from cerebellar ataxia have extremely variable gait kinematic features. We investigated whether and how wearable inertial sensors can describe the gait kinematic features among ataxic patients. Methods. We enrolled 17 patients and 16 matched control subjects. We acquired data by means of an inertial sensor attached to an ergonomic belt around pelvis, which was connected to a portable computer via Bluetooth. Recordings of all the patients were obtained during overground walking. From the accelerometric data, we obtained the harmonic ratio (HR), i.e., a measure of the acceleration patterns, smoothness and rhythm, and the step length coefficient of variation (CV), which evaluates the variability of the gait cycle. Results. Compared to controls, patients had a lower HR, meaning a less harmonic and rhythmic acceleration pattern of the trunk, and a higher step length CV, indicating a more variable step length. Both HR and step length CV showed a high effect size in distinguishing patients and controls (p < 0.001 and p = 0.011, respectively). A positive correlation was found between the step length CV and both the number of falls (R = 0.672; p = 0.003) and the clinical severity (ICARS: R = 0.494; p = 0.044; SARA: R = 0.680; p = 0.003). Conclusion. These findings demonstrate that the use of inertial sensors is effective in evaluating gait and balance impairment among ataxic patients.

**Keywords:** inertial sensors; cerebellar ataxia; movement analysis; gait analysis; balance; personalized medicine; rehabilitation

#### **1. Introduction**

Patients suffering from cerebellar ataxia exhibit peculiar spatiotemporal and kinematic features that contribute to an unstable gait [1–5]. The gait impairment typically worsens over time, in parallel with the functional decline associated to the neurodegenerative process [6,7]. While stable gait is characterized by repeatable walking patterns [8], steadiness in the case of perturbations [9–13], and effectiveness in maintaining upright balance [14,15], ataxic gait is extremely variable over gait cycles [1] and exhibits inefficient coordination between upper and lower segments of body, even in the absence of external perturbations [16]. Taking into account such conditions, it is reasonable to

hypothesize that when perturbation occurs in ataxic patients, the consequent fall risk increases, and the gait pattern can be defined as unstable [6,7].

The evaluation of gait instability and fall risk is, therefore, pivotal in the study of ataxic gait to prevent further disabilities, and in order to maximize and optimize the information we gather from such evaluation, it should be performed in a real-life environment outside the motion analysis laboratory for a long period of time. In this context, wearable magnetic and inertial measurement units (MIMUs), consisting of a three-axial accelerometer, a gyroscope, and a magnetometer, represent a self-contained alternative to conventional laboratory-based motion capture systems [17–19]. This technology estimates the three-dimensional (3D) orientation of MIMUs with respect to a global coordinate system by specific sensor fusion algorithms, using angular velocity, gravity and magnetic field vectors.

A series of biomechanical stability measures based on MIMU evaluations have been proposed in several studies on neurological gait disorders with dynamic unbalance [3,20–22]. The maximum Lyapunov exponent (λmax) is an available method to evaluate gait instability [4] and fall risk [3] in ataxic patients, but the relationship between λmax and clinical severity has not been definitively established, since it has been demonstrated to be both positively [4] and negatively [3] correlated to International Cooperative Ataxia Rating Scale (ICARS) scores. A possible explanation could be found in the heterogeneous etiologies of the study samples, respectively acquired cerebellar lesions after tumor resection [4], and neurodegenerative ataxia [3]. Another important issue is that λmax properly explores the nonlinear dynamic local stability of the trunk during locomotion when at least 150 continuous strides are recorded [15]. However, such stride numbers are often not practically feasible in ataxic patients, and this could have influenced the correlation analysis between λmax and clinical severity.

To the best of our knowledge, no other studies in the literature have used additional indexes of stability, like harmonic ratio (HR) and the coefficients of variation (CV) based on MIMU data to detect the instability of ataxic patients. Therefore, the aim of this study is to evaluate these indexes of stability and, in particular, examine the ability of each index to detect the instability of ataxic patients compared to healthy controls and determine the fall risk. HR was chosen to evaluate the trunk acceleration patterns, a key feature in determining the severity of the ataxic gait [5,16], while CV was chosen to evaluate the variability of step length, an important compensatory mechanism in ataxic patients.

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

#### *2.1. Participants*

Seventeen patients affected by primary degenerative cerebellar ataxia were enrolled in the study. Table 1 summarizes the patients' clinical features and genotype.

The complete neurological assessment included (1) cognitive evaluation according to mini-mental state examination (MMSE) scale, (2) cranial nerve evaluation, (3) muscle tone evaluation, (4) muscle strength evaluation, (5) joint coordination evaluation, (6) sensory examination, (7) tendon reflex elicitation, and (8) disease severity measured by International Cooperative Ataxia Rating Scale (ICARS) and Scale for the Assessment and Rating of Ataxia (SARA) [23,24]. We excluded patients with gait impairment due to extracerebellar symptoms or orthopedic disorders. Regarding the extracerebellar disorders affecting gait, we excluded patients with spasticity, polyneuropathy, cognitive deficits, and extrapyramidal disorders. Of the recruited patients, no one presented with signs of spasticity, hyposthenia, hypoesthesia, and/or cognitive impairment (MMSE > 24). All patients were able to walk alone without any kind of assistance or aid, and were receiving physical therapy, including active and passive exercises for upper and lower limbs as well as balance and gait re-education. Furthermore, no patient had significant visual deficits according to the Snellen visual acuity test. Almost all of the patients had non-disabling oculomotor abnormalities, such as nystagmus or square wave jerks pursuit movements, because of the underlying disorder. A brain MRI showed that all patients had cerebellar atrophy. Regarding the fall risk assessment, all patients had to complete a

specific questionnaire designed to evaluate the number of falls in the previous year, the characteristics of such falls (side, associated injury), and the circumstances in which they occurred. The number of falls in the last year was used for correlation analysis. Sixteen age-matched healthy adults (age, ataxic patients 53.53 ± 12.12 years, healthy controls, 50.94 ± 8.79 years, p > 0.05) were enrolled as the control group. We obtained informed consent from each patient and healthy subject, which complied with the Helsinki Declaration and was approved by the local ethics committee.


**Table 1.** Ataxic patients' clinical and anthropometric characteristics.

SAOA: sporadic adult onset ataxia of unknown etiology; SCA: spinocerebellar ataxia; FRDA: Friedreich's ataxia.

#### *2.2. Gait Analysis*

We acquired data with an inertial sensor (BTS GWALK, BTS, Milan, Italy), attached to an ergonomic belt placed around the pelvis at the level of the L5 vertebra, connected to a portable computer via Bluetooth. The sampling rate was 100 Hz, and the sensor, endowed with a tri-axial accelerometer (16 bit/axes), a tri-axial magnetometer (13 bit), and a tri-axial gyroscope (16 bit/axes), measured the linear trunk accelerations and the trunk angular velocities in three space directions (i.e., AR: anterior-posterior; ML: mediolateral; VT: vertical direction).

#### *2.3. Task Description*

Before starting the experimental session, participants were asked to walk along a predetermined route in order to familiarize themselves with the procedure. Recordings of all the patients were obtained during overground walking. We asked participants to walk along a corridor (3 m wide and 20 m long) at their preferred speed. Control subjects were asked to walk at a low speed in order to match the two groups for speed (ataxic patients, 0.939 ± 0.195 m/s; controls, 0.924 ± 0.239 m/s; p > 0.05).

#### *2.4. Inertial Sensor Data Processing*

The 'walking protocol' of the inertial sensor (G-STUDIO, BTS, Milan, Italy) was used to detect: (1) trunk acceleration patterns, (2) right and left heel strikes, and (3) toe-off. The HR and the CV were calculated using MATLAB software (MATLAB 7.4.0, MAthWorks, Natick, MA, USA).

Harmonic ratio. The harmonic ratio (HR), initially described by Gage [25] and later modified by Smidt et al. [26], provides an indication of the acceleration patterns, smoothness, and rhythm. Since the unit of measurement from a continuous walking trial is a stride (two steps), a stable, rhythmic gait pattern should be characterized by multiples of two repeated acceleration patterns within any given stride. Accelerations patterns that do not repeat in multiples of two generate out of phase accelerations, reflecting irregular accelerations during a walking trial and, therefore, an unstable gait pattern. The harmonic content of the acceleration signals can be analyzed in each spatial direction using stride frequency as the fundamental frequency component. Based on each stride time, 20 harmonics were calculated. Trunk accelerations of each stride were broken down into individual sinusoidal waveforms using discrete Fourier transform (DFT).

Since a stable smooth gait pattern is characterized by acceleration signals in VT and AP directions that repeat in multiples of two during a single stride, HRs in the VT and AP directions were calculated as the ratio of the sum of the amplitudes of the first 10 even harmonics divided by the sum of the amplitudes of the first 10 odd harmonics. In the ML direction, acceleration signals were repeated once for any given stride, so HRs in the ML direction were calculated as the sum of the amplitudes of the odd harmonics divided by the sum of the amplitudes of the even harmonics. We used a high-pass filter with cutoff at 20 Hz to eliminate noise signals.

HRs per stride were determined and averaged across a steady walk, resulting in a mean HR. HR in AP and VT, and in the ML direction, were calculated as below [19]:

HR in anterior–posterior and vertical directions

$$HR = \frac{\sum\_{i=1}^{10} A\_{2i}}{\sum\_{i=1}^{10} A\_{2i-1}}$$

HR in the medio-lateral direction

$$HR = \frac{\sum\_{i=1}^{10} A\_{2i-1}}{\sum\_{i=1}^{10} A\_{2i}}$$

where A2i denotes the amplitude of the first 20 even harmonics and A2i–1 indicates the amplitude of the first 20 odd harmonics. The higher the HR value, the smoother the walking pattern.

Coefficient of variation. In order to compute the step length CV, the step length was estimated using the upward and downward movements of the trunk, as proposed by Zijlstra and Hof [27]. Assuming a compass gait type, the body's center of mass (CoM) movements in the sagittal plane follow a circular trajectory during each single support phase. In this inverted pendulum model, changes in height of CoM depend on step length [27]. Thus, step length can be deduced by known height changes and predicted from geometrical characteristics as follows: step length = 2 2lh <sup>−</sup> h2.

In this equation, h is equal to the change in height of the CoM, and l represents the pendulum length. Changes in vertical position were calculated by a double integration of the vertical acceleration. A high-pass filter (fourth-order zero-lag Butterworth filter at 0.1 Hz) was used in order to avoid integration drift. The difference between highest and lowest position during a step cycle was used to determine the amplitude of changes in the vertical position (h). Leg length was considered as pendulum length (l). Step length was calculated as the mean of step lengths observed during seven subsequent steps of each subject.

Then, the step length coefficient of variation (CV) was computed as follows: *CV* = 100 *SD mean* where mean is the mean step length and SD is the standard deviation over the entire step length for each subject [1]. The CV is a measure of the variability of a data set; the closer to 0 the CV is, the less variable the data are.

#### *2.5. Statistical Analysis*

We used the SPSS 17.0 software (SPSS Inc. Chicago, IL, USA) for statistical analysis. All data were expressed as mean ± standard deviation; p < 0.05 was considered statistically significant. We assessed the normality of distributions using the Shapiro-Wilk test.

Mean and standard deviation within subjects were computed for speed and stability indexes. We used the independent-samples t test to look for differences between the stability indexes of ataxic patients vs. controls. Cohen's d index was used to assess the effect size of the stability indexes in the three spatial directions [28,29]. We used the Pearson's test to investigate any correlation We used the Pearson test to investigate any correlation of acceleration HR and step length CV with (1) age, (2) height, (3) weight, (4) disease duration, (5) total ICARS and SARA scores and (6) number of falls in the last year.

#### **3. Results**

Looking at the low scores of ICARS and SARA, the recruited patients mainly showed cerebellar symptoms (see Table 1).

HR in all three directions and step length CV were all significantly different when compared to the controls (Table 2). Briefly, the HR of patients was lower than the HR of healthy subjects, meaning a less harmonic and rhythmic acceleration pattern of the trunk, while the CV of step length was greater in patients than in the controls, indicating a more variable step length in ataxic patients. Both HR and CV of step length showed a high effect size in distinguishing patients and controls, but HR in all three directions showed a higher effect size score when compared to the CV (Table 2).

**Table 2.** Comparisons of the stability indexes between 17 ataxic patients and 16 controls at matched gait speed.


Mean ± standard deviation values, the results of the independent samples t-test and Cohen's d are reported. Values of p lower than 0.05 were considered statistically significant. HR-AP: harmonic ratio in the anterior–posterior direction; HR-ML: harmonic ratio in the mediolateral direction; and HR-VT: harmonic ratio in the vertical direction.

Surprisingly, no correlation was found between HR in all directions, falls/year, and clinical severity (ICARS and SARA scores) (Table 3), while a significant positive correlation was found between the CV of step length and the falls/years and ICARS and SARA scores (Figure 1).

**Table 3.** Correlation analysis between HR in all directions and ICARS, SARA, and falls/year.


The reported values represent Pearson correlation value (R) and statistical significance value (p). HR-AP: harmonic ratio in the anterior–posterior direction; HR-ML: harmonic ratio in the mediolateral direction; and HR-VT: harmonic ratio in the vertical direction.

**Figure 1.** Correlations between the maximum step-to-step coefficient of variation and the falls/year, ICARS-total, and SARA-total scores in 17 ataxic patients. Pearson's R coefficient (R) and significance (p) are reported.

#### **4. Discussion**

In the present study, we found that trunk acceleration smoothness, as described by HR values, and the variability of step length, as described by the CV, may provide insights about gait stability in subjects with degenerative ataxia. Furthermore, the variability of step length correlated with both clinical severity and fall risk.

Regarding the acceleration patterns of the trunk, the HR of patients significantly differed from that of healthy controls in all three spatial planes. Moreover, it showed a high effect size, according to Cohen's d index (Table 2). This means that ataxic patients, compared to healthy subjects, exhibit a substantial reduction of trunk movement smoothness. When discussing these findings, we should bear in mind that the trunk has a great functional importance in minimizing the magnitude of linear and angular displacement of the head, ensuring clear vision [30,31], facilitating the integration of vestibular information [32], contributing to the maintenance of balance [5,6,16,33,34], and acting as a driving force for locomotion [35]. Consequently, investigating upper body stability in patients with degenerative cerebellar ataxia is essential, since the lack of motor control [5] and coordination [16] makes the trunk itself generate perturbations in a sort of vicious circle in parallel to the clinical decline [2]. In this context, trunk acceleration smoothness, as described by the HR values, provides a deeper insight into gait disturbances [14,36]. From the literature, we know that trunk acceleration smoothness during walking is predictive of gait dysfunction [37,38] and fall risk in older people [14,39]. Moreover, HR has already been found to be abnormal in patients who have suffered a stroke, Parkinson's disease, or multiple sclerosis [19,20,22,40].

Overall, these findings suggest that HR can substantially describe trunk accelerative behavior abnormalities among patients with degenerative ataxia [41]. On the other hand, we did not find any correlation between HR, the number of falls, and clinical severity. This last result is apparently in contrast with previous studies that found a relationship between clinical severity, increased range of motion of trunk [5] and trunk–thigh coordination deficit [16]. Considering the small sample size of our study, we cannot exclude a type II error. Nevertheless, another possible explanation might come from the different implemented technologies and protocols. In fact, previous studies assessed the kinematic patterns of the upper segment of the head and the trunk via optoelectronic systems [5,16]. This means that the body markers were located on body segments (i.e., the head and upper trunk) whose range of movements was wider than the lumbar one, as investigated by a BTS GWALK device located on L5 vertebra. Further studies will assess such differences, evaluating the role of the ergonomic belt placed around the thorax just underneath the axilla, and will validate inertial sensor findings against optoelectronic systems.

The other parameter we considered was the CV of step length, which has been reported to be significantly different in subjects with cerebellar ataxia when compared to healthy subjects [42]. During the progression of the disease (i.e., >4 years from the onset, as in our sample), subjects with degenerative ataxia tend to lose the ability to both enlarge their step width and fasten their walking speed and—maintaining the same step width and speed—they shorten their step length in order to reduce their single support time [43], with a significant increase in step length CV that can lead to an

increased risk of falls. In fact, we found that the CV of step length was higher in patients with ataxia than the controls and, unlike HR, the CV of step length significantly correlated with the ICARS and SARA scores and with the number of falls per year. These findings differ from those from a previous study, where a correlation between the CV of step length and clinical severity was not detected [1]. This difference might be due to both the use of different movement analysis technologies (inertial sensor vs. optoelectronic system) and different investigated samples (sporadic adult onset ataxia of unknown etiology (SAOA)/ spinocerebellar ataxia (SCA) vs. SCA/SAOA/Friedreich's ataxia (FRDA)). Since a camera-based optoelectronic system can capture a smaller change of gait than MIMUs, our data should be interpreted with caution. However, the investigation of a large number of patients with FRDA in Serrao et al. [1] might explain, at least in part, such discrepant results. In this view, patients with FRDA and those with SCA and SAOA may show a different relationship between clinical features and gait stability; further studies are needed to explore this issue. However, our aim was not to obtain an alternative measure of step length, but to detect the relationship between the multifactorial gait impairment [5,6,16,33,34], clinical severity, and the fall risk. Because the MIMU-measured CV of step length is influenced by movements of the trunk [27], and trunk–thigh coordination is impaired in ataxic patients [16], our MIMU-measured CV might reflect trunk–thigh coordination variability. In this respect, the aforementioned limitation might come in handy, being such a multifactorial parameter able to summarize factors that, put together, explain gait instability in ataxic patients.

Finally, our results cannot be generalized as representative of the ataxic population because they refer to patients with a disease duration of at least 8 years, preserved walking ability, and without extracerebellar symptoms as disabling oculomotor abnormalities. Moreover, our findings highlight the need to investigate the relationship between each MIMU-measured index and the corresponding ones measured by traditional optoelectronic systems in order to have proper validation.

#### **5. Conclusions**

In conclusion, the present study highlighted that both HR and CV differed between ataxic patients and healthy subjects. However, when considering the correlation with clinical severity and fall risk, only MIMU-measured CV of step length was able to describe the burden of ataxic symptoms and to draw clinical attention towards a possible increased fall risk. These MIMU-based parameters might provide real-world information on patients' disabilities and falls, since they are obtained through wearable and comfortable devices.

**Author Contributions:** Conceptualization, P.C., M.S., C.I., C.C. and A.T. Methodology, P.C., C.C., M.S. and C.I.; Investigation, C.C., S.F.C., C.I.; Data Curation, C.I., C.C. and S.F.C. Writing—Original Draft Preparation, P.C., C.I., C.C. and S.F.C. Writing—Review & Editing: M.S. and G.R. Supervision: M.S.

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

**Acknowledgments:** Thanks to Matteo Gratta for language editing.

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

#### **References**


© 2019 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* **High-Specificity Digital Architecture for Real-Time Recognition of Loss of Balance Inducing Fall**

#### **Daniela De Venuto and Giovanni Mezzina \***

Department of Electrical and Information Engineering, Politecnico di Bari, 70125 Bari, Italy; daniela.devenuto@poliba.it

**\*** Correspondence: giovanni.mezzina@poliba.it; Tel.: +39-0805963562

Received: 1 November 2019; Accepted: 30 January 2020; Published: 31 January 2020

**Abstract:** Falls are a significant cause of loss of independence, disability and reduced quality of life in people with Parkinson's disease (PD). Intervening quickly and accurately on the postural instability could strongly reduce the consequences of falls. In this context, the paper proposes and validates a novel architecture for the reliable recognition of losses of balance situations. The proposed system addresses some challenges related to the daily life applicability of near-fall recognition systems: the high specificity and system robustness against the Activities of Daily Life (ADL). In this respect, the proposed algorithm has been tested on five different tasks: walking steps, sudden curves, chair transfers via the timed up and go (TUG) test, balance-challenging obstacle avoidance and slip-induced loss of balance. The system analyzes data from wireless acquisition devices that capture electroencephalography (EEG) and electromyography (EMG) signals. The collected data are sent to two main units: the muscular unit and the cortical one. The first realizes a binary ON/OFF pattern from muscular activity (10 EMGs) and triggers the cortical unit. This latter unit evaluates the rate of variation in the EEG power spectrum density (PSD), considering five bands of interest. The neuromuscular features are then sent to a logical network for the final classification, which distinguishes among falls and ADL. In this preliminary study, we tested the proposed model on 9 healthy subjects (aged 26.3 ± 2.4 years), even if the study on PD patients is under investigation. Experimental validation on healthy subjects showed that the system reacts in 370.62 ± 60.85 ms with a sensitivity of 93.33 ± 5.16%. During the ADL tests the system showed a specificity of 98.91 ± 0.44% in steady walking steps recognition, 99.61 ± 0.66% in sudden curves detection, 98.95 ± 1.27% in contractions related to TUG tests and 98.42 ± 0.90% in the obstacle avoidance protocol.

**Keywords:** near falls; loss of balance; pre-impact fall detection; activities of daily life; bio-signals; EEG; EMG

#### **1. Introduction**

Recently, freezing of gait (FOG) and falls received increasing recognition as strongly debilitating features of Parkinson's disease (PD) [1,2]. By contrast with the tremor, which dominates the early stage of PD [2], falls and FOG are most common in advanced PD stages. The two phenomena seem related to each other, according to the study in [2], in which it is shown how sudden FOG can disturb the balance and, thereby, represents a common cause of falls in PD. Epidemiologic prospective studies conducted with 1-year, or 6-months, follow up [1–6], showed that the 45–68% of people with PD experience at least one fall per year, with a large portion (50–86%) falling recurrently [3–5]. It is not surprising that including the "near falls" this rate increases up to ~90% [6]. In this context, a near fall is a situation in which, despite a loss of balance, the body-ground impact could be avoided by grasping a support [6].

The clinical presentation in [2] shows that in PD patients most falls result from sudden changes in posture (in particular, turning movements of the trunk), rapid changes in the walking tasks (curve, transfers from the bed or the chair, etc.) or because they try to perform more than one activity simultaneously with walking or balancing.

In this context, the advances in wireless sensors networks, wearable acquisition devices, and new and more reliable digital signal-processing approaches for kinematic and biosignals analysis prompted the scientific community to develop technological solutions for early fall detection (FD).

The systematic review of FD solutions in [7] showed that body-worn accelerometers can be used to detect impacts and changes in orientation associated with falls. In the same context, the authors in [7] conclude that the accuracy of these FD systems may be improved by jointly using multiple sensors, e.g., signals from smartphone gyroscopes or barometers to define the height changes associated with falls [7–9]. These technologies aim to provide fast detection of falls but, at the moment, they are still not able to fully prevent injuries resulting from falls (e.g., hip fractures and traumatic brain injury) [8,9]. For this purpose, the focus of research contextually moved on fall risk assessment. This area of interest oversees the identification of the people's risk of falling, facilitating in this way early interventions via FD systems. Currently, fall risk-assessment procedures take into account the clinical evaluation of multiple domains such as balance control, mobility, physiology (strength, vision), psychology (fear of falling), cognition and environmental risk [7].

In this context, detecting near-falls (or recoverable imbalances) provides new opportunities to identify people with a high risk of falling before an actual fall occurs [10]. Near falls are defined as loss of balance that does not result in a fall because corrective action is taken to recover balance. They typically consist of slips, trips, and missteps. Moreover, since older people who frequently experience near falls are at increased risk of future falls [7,10], remote monitoring of these events during daily life could provide useful information to target falls and related circumstances as part of fall prevention initiatives [7].

Ultimately, an accurate algorithm for the detection of near falls could enhance the quality of existing fall detection systems by reducing false alarms [7].

In this respect, Table 1 summarizes state-of-the-art solutions [11–15] declared to be able to recognize near falls. The table reports the architectures in terms of used acquisition equipment, fall indicators (i.e., the feature(s) to be monitored and classified) and chosen classification method. Table 1 also dedicates a field to the Activities of Daily Life (ADL) and near-fall scenarios included in the discrimination. Finally, the last two rows summarize the declared system performance (i.e., accuracy and efficiency) and the applicability of proposed systems to daily-life and/or ambulatory contexts, as well as their suitability in the context of real-time near falls detection and fall prevention. All the studies selected for the comparison analyze unexpected slippages, classifying them as near falls because all the perturbations analyzed in [11–15] led to balance recovery.

Table 1 shows that the most used technologies in the loss of balance detection are motion capture systems (MCS) [12,13,15] and inertial measurement units (IMU) [11,14]. Solutions based on MCS are classified as context-aware and typically consists of a set of reflective markers and fixed cameras. For this reason, MCS-based fall detection systems present two limits: they are expensive and only suitable for ambulatory applications [12,13]. Pointing at daily-life applicability, authors in [11,14] propose wearable solutions mostly based on IMU sensors. More in detail, the authors in [11] analyze acceleration and angular velocity from 7 IMU sensors via a machine-learning (ML) approach. In a similar way, the authors in [14] exploit acceleration data from a single device, placed on the waist, to record vertical velocity from trials belonging to the chosen classification clusters (i.e., ADL vs. near falls). In terms of adopted classification methods, the result is that the most used approaches are still based on thresholds, in order to preserve a good speed in system response [12,14,15].



threshold, SVM: support vector machine, ARIMA:

Activities of Daily Life 5

**Applicability**

 **Acronyms:** OL: ordinary (or daily) life applicability,

autoregressive

 integrated moving average, ICA: independent component analysis, AR:

 Clin.: Ambulatory applicability,

 FD: Fall-detection

 system suitability.

Autoregression.

 4 **Classes Acronyms:** ADL:

Machine-learning based solutions have been also investigated by authors in [11,13]. In this respect, noteworthy is the approach in [13], where the authors use an artificial neural network (ANN) to classify acceleration independent components, providing an interesting tradeoff between overall accuracy and fall-detection timing.

Table 1 reports the most investigated performance in FD and near-fall detection applications: the fall/imbalance recognition accuracy and the detection time [16]. The system accuracy parameter is composed of the sensitivity and the specificity. For the sake of comparison, all the analyzed works [11–15] share the same sensitivity and specificity definition. The first (i.e., sensitivity, Se (%) in Table 1) is defined as the ratio between the number of successfully detected falls/losses of balance over the total number of recorded perturbations. While, the specificity (Sp (%) in Table 1) is determined by the ratio between the number of successfully detected ADL over the total number of ADL-related trials. Finally, the detection time (DT in Table 1) characterizes the system efficiency. It is defined as the time difference between the fall initiation (that can be uniquely defined) and fall detection. This parameter gives an idea on how rapid the fall-detection system responds to a fall.

Concerning the applicability field of Table 1, the device's wearability and the proper specificity characterization relate to the suitability of the application to ordinary-life contexts. While, the detection time under a balance recovery limit (i.e., ~550 ms) [17] determines the system suitability for pre-impact FD strategy improvement.

In this paper, we propose and preliminarily validate a digital architecture for the loss of balance recognition during unexpected slippages potentially inducing fall. The main contributions of the paper concern:


The architecture proposed here exploits medical evidence from recent studies [18–23], according to which the cerebral cortex can regulate the postural stability according to environmental demands [18]. Specifically, the authors in [19,20] proved that low-frequency cortical rhythms (f < 13 Hz) are related to perception and cognitive control. In the loss of balance context, the modulation of the bands θ (4–7 Hz) and α (8–12 Hz) seems to be related to the visual field stabilization and active decoding of data from the vestibular system. Contextually, authors in [21–23] concluded that high-frequency cortical rhythms (f > 13 Hz) are commonly related to highly specific motor functions. Specifically, the β bands (i.e., β I, β II, β III), play a main role in muscle firing operations to compensate balance.

Besides the cortical dynamics' characterization, the muscular behavior could also be uniquely characterized. In this respect, the authors in [22–25] demonstrated that for accelerations or decelerations of the supporting surface (e.g., slippage) a low latency response (70–300 ms) occurs in the muscles near the ankles. It results in a muscular pattern characterized by co-contractions between agonist and antagonist muscle bundles [24].

Keeping this evidence in mind, the novel architecture exploits electrophysiological measurements from 10 EMG electrodes, to assess the muscular activity, and 13 EEG channels, to analyze the subject's cortical involvement during reactive response or normal motor planning.

Electrophysiological signals (i.e., EEG/EMG) are synchronously acquired via a central gateway. The gateway streams data to two computational units that distinctly analyze muscular and cortical activity. The unit dedicated to the muscular characterization has two main roles: realizing a binary ON/OFF pattern from muscular activity and triggering the cortical analysis unit. Once triggered, this latter unit quantifies the cortical involvements as the rate of variation in the EEG power spectrum density (PSD), considering the five bands of interest identified by authors in [18–23]. The parameters extracted from these units define some neuromuscular features of the subject under monitoring. As a final step, these features are sent to a logical network, which embeds a set of dynamic thresholds from the system calibration phase. In this application, the system calibration progressively builds a conservative range in which the neuromuscular features can be considered as "standard" and, thus, safe for the balance. The expectancy is that the ADL do not strongly affect the cortico-muscular parameters as, instead, happens during a loss of balance.

The paper is structured as follows. Section 2 outlines the experimental protocols, the setup and the implemented algorithm. Section 3 is dedicated to experimental results. Section 4 proposes a discussion about the system outcomes and Section 5 concludes the paper, presenting future perspectives.

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

#### *2.1. Participants*

Nine young and healthy volunteers (8 males, 1 female, 26.3 ± 2.4 years old, 64.5 ± 9.8 kg, 1.71 ± 0.06 m) were enrolled for this study. Six of them contributed to a near-fall scenarios test, while three subjects were actively involved in the system robustness test via ADL-like tasks. Before starting the experimental sessions, all the participants signed the informed consent. Research procedures were in accordance with the Declaration of Helsinki and was approved by the Local Ethical Committee (Protocol n. 2019\_0025904).

#### *2.2. Architecture Overview*

Figure 1 shows a block diagram of the proposed loss of balance detection architecture. According to the figure, the system can be divided in four main sections: the acquisition unit, the muscular and cortical units and, finally, the classification block. As depicted in Figure 1, the proposed digital architecture synchronously operates on a STM32L4x microcontroller for the muscular analysis, and by means of Simulink real-time modeling to assess the cortical involvement. The Simulink model has been fully realized by blocks from the Digital Signal Processing (DSP) library in order to be implemented on a microcontroller.

The system working principle is inspired by our previous works [26,27], which laid the methodological bases for the joint analysis of EEG and EMG signals in the fields of gait analysis and involuntary movements detection. The overall processing chain is detailed in Figure 1.

**Figure 1.** Proposed architecture block diagram. The figure shows the electroencephalography/ electromyography (EEG/EMG) experimental setup, as well as the graphical representation of the working flow of each involved block.

#### 2.2.1. Acquisition Unit

The acquisition unit consists of a multi-sensing interface that jointly collects data from 10 surface EMGs and an EEG headset. The acquisition equipment has been selected to be fully wireless and wearable, allowing the subject complete freedom of movement.

In more detail, during the test and data collection phases, subjects wore a 32-channels EEG wireless headset (g.Nautilus Research by g.Tec [28]) and a set of 10 wireless surface EMG electrodes (Cometa WavePlus by Cometa srl [29]). According to the experimental measurement setup sketch in Figure 1, thirteen EEG sites have been monitored: F3, Fz, F4, C3, Cz, C4, Cp5, Cp1 Cp2, Cp6, P3, Pz, P4, according to the international 10–20 system. The O2 electrode was used for noise suppression, AFz as ground and the A2 (right earlobe) as the reference electrode. The EEG data were sampled at 500 Hz with 24-bit resolution.

On the muscular side, 10 surface EMG channels were monitored from following bilateral muscle groups: Anterior tibialis (AT), Lateral gastrocnemius (LG), Vastus medialis (VM), Rectus femoris (RF), and Biceps femoris (BF). The EMG signals were recorded with a sample rate of 2048 Hz and down sampled to 500 Hz (@16-bit resolution) before the transmission.

Data from the 10 EMG nodes are wirelessly streamed to a dedicated gateway, which is mounted on a Nucleo STM32L476RG board via a dedicated Printed Circuit Board (PCB) shield. Then the Muscular Unit algorithm runs on the microcontroller, analyzing the signals sample-by-sample.

Data from the EEG headset are sent to a base station connected via USB to a central computation unit that runs the Simulink model. The base station is also equipped with a 26-pin D-SUB connector used for the parallel reception of 8 digital input pins (DIN). These DINs will be used to receive data and triggers from the microcontroller. On the Simulink model side of the cortical unit, data from the monitored channels are continuously sent to nch = 13 circular registers, waiting for the enable signal from the muscular block. In this application, the central computation unit that runs the Simulink model consist of a HP Y5L00AE computer embedding an AMD A10-9600P processor (Hewlett-Packard—Palo Alto, CA, USA).

Pre-processing. The EEGs were progressively band-filtered between 1 Hz and 40 Hz by using a built-in 8th order Butterworth filter before the transmission [30]. The EMG node band-pass filters the signal between 15 Hz and 250 Hz before to be sent data to microcontroller [31]. Finally, a numeric notch filter 48–52 Hz has been implemented via the Simulink model for both EEG and EMG signals.

#### 2.2.2. Experimental Protocols

To test the robustness of the algorithm proposed here and to ensure system suitability for daily-life contexts, the system response was assessed during four different ADL-like tasks. Figure 2 shows, through a snapshot grid, the experimental protocols carried out by the participants. Each row in the figure is composed of 6 frames, realizing a demonstrative sequence of the four experimental tasks:


**Figure 2.** Experimental protocols grid. Each row in the panel represents a 6-frame demo sequence of the experimental protocol carried out by the participants. (**a**) Steady walking to near fall (slip) protocol; (**b**) Chair transfer via timed up and go test; (**c**) Balance-challenging obstacle avoidance.

#### 2.2.3. ON/OFF Muscular Pattern Extraction

The muscular unit operates on the collected EMGs, generating an ON/OFF binary pattern of muscular activation (OOM—Figure 1) starting from the electrophysiological signal of each monitored muscle. Briefly, the implemented algorithm set OOM = 1 when the muscle is contracted, otherwise reset OOM = 0. This binarization procedure is entrusted to a moving threshold approach detailed in [26,33], because it demonstrated to be able in following the muscle tone changes (e.g., due to fatigue).

The ON/OFF muscular pattern extraction routine implemented on the STM32L4 μC can be briefly summarized by the following steps: the system progressively stores, for each muscle, a time-window containing the last M = 250 samples received.

It then extracts two data blocks: the first one containing the full EEG time-window (M = 250 samples, i.e., ~500 ms) and the second one that includes only the last N = 125 samples (i.e., ~250 ms). The algorithm squares these two vectors, averaging their elements. The resulting EMG power value for the longer time window (PM) acts as adaptive threshold, while the same parameter for the shorter time window (PN) as the instantaneous power. Finally, the two values are compared: if PN > PM, the system set OOM = 1, otherwise OOM = 0. The PM and PN values refresh and progressively adapt to each sample.

This ON/OFF muscular pattern digitization step (Figure 1) generates 10 parallel OOMs (one per muscle), which are sent to the muscular activity pattern (MAP) step according to the block diagram in Figure 1.

Two OOMs from both the Gastrocnemii are selected to trigger the Cortical Unit. These OOMs will be named master trigger (MT, Figure 1) hereafter.

To exclude, from the computation, the cortical activity that it is not strictly related to the specific movement, protecting from false alarms in the EEG unit, we selected as MT the gastrocnemius because it uniquely intervenes during the midstance gait phase.

#### 2.2.4. Cortical Involvement Assessment

Once enabled via MT (side independent), the cortical unit extracts from the circular buffers 13 time-windowed EEGs of 400 samples (~800 ms) preceding the MT onset.

As a first step, these subset of EEG data undergo the on-line Riemannian artifact subspace reconstruction (rASR) [34]. The rASR is an online/offline artifacts attenuation method for mobile EEG data based on an ASR with Riemannian geometry.

The cortical unit analyzes these artifacts-free brain signals, quantifying the rate of variation in the EEGs power within the five bands of interest identified by authors in [18–23]: θ (4–7 Hz), α (8–12 Hz), β I, β II, β III (13–15, 16–20, 21–40 Hz). In more detail, the power spectrum density measurements are done by applying a sliding-window fast Fourier transform (FFT) on the considered EEG subset. For the purpose, the artifacts-free EEG subset is split in 20 overlapped windows long 200-samples with a step of 10 samples, covering the entire length of the subset.

Considering a single EEG window, the application of the FFT leads to a spectral resolution of 2.5 Hz (considering fsEEG = 500 sps and Lwin = 200 samples), which is suitable for the band multiplexing [26,35].

For each evaluated window, the system extracts a matrix named **SBoI** <sup>∈</sup> **<sup>R</sup>**nch, nBoI, with nch = 13 and nBoI = 5 number of bands involved in the multiplexing.

Each SBoI element is the sum of the spectral contents falling within the selected j-th band due to the multiplexing:

$$\mathbf{S\_{Bel}(i,j)} = \frac{\sum\_{\mathbf{k} = (\text{jth band})} (\mathbf{S(k)}) \big|\_{\text{dB}}}{\text{Rg}} \qquad \mathbf{i} = 1 : \mathbf{n\_{ch}} \text{ } j = 1 : \mathbf{n\_{Bol}} \text{ } k = 1 : \mathbf{Rg} \tag{1}$$

where the j-th band can mean the θ (k = 2:3), α (k = 3:5), β I (k = 6:7), β II (k = 8:10), β III (k = 11:16) band range, while Rg is the maximum k index (i.e., length of the j-th band).

The **SBoI** is then extended to the 20 overlapped windows, generating a 3D matrix: **<sup>Y</sup>** <sup>∈</sup> **<sup>R</sup>**nch, nBoI, nW with nW = 20 number of measurements.

For the sake of clarity, considering a single band of interest (e.g., α band), data from the 20 FFT steps undergo a linear fitting via an ordinary least squares (OLS) estimator according to the equation:

$$\left. \hat{\mathbf{p}}(\mathbf{i}) \right|\_{\alpha} = \mathbf{A} \backslash \mathbf{Y}(\mathbf{i}, \alpha, \mathbf{1} \; : \; \mathbf{n} \mathcal{W}) \tag{2}$$

where **p**ˆ(**i**) <sup>α</sup> is the OLS-based parameter vector for the i-th channel on the α band. It contains, in the order, the estimated linear model intercept *q*ˆ = pˆ(i) α[1] and the estimated straight-line slope *m*ˆ = pˆ(i) α[2]. In the same equation (i.e., Equation (2)), **<sup>A</sup>** is the matrix of the basic functions containing a column of 1 and column of time vector (t = 20:800 ms, step 20 ms). Finally, **Y**(i, α, 1 : nW) is the vector that contains the FFT measurements on the i-th channel and the α band. The resulting linear models (OLS estimation—Figure 2) permit to approximate the cortical involvement parameter as the straight-line slope, *m*ˆ . More details about the EEG computation branch implementation has been provided in our previous works [26,33]. The OLS-based estimation procedure is contextually applied to 13 channels and 5 band of interests, generating 65 *m*ˆ values.

#### 2.2.5. Muscular Activity Pattern Extraction

The muscular unit hosted by the microcontroller operates in parallel with the cortical involvement analysis. In this frame, the muscular unit analyzes the 10 parallel OOMs via the MAP extraction routine. This stage aims to analyze the contraction status of each analyzed muscle "in correspondence" of the MT rising edge. Specifically, is time windows of 20 ms (11 samples), 10 ms before and 10 ms after the MT rising edge, is considered. The resulting OOM observation is named **wOOM** <sup>∈</sup> **<sup>R</sup>**nEMG, Lw, where nEMG is the number of monitored EMG nodes and Lw is the number of samples composing the subset. In this application nEMG is 10, while Lw is 11. Also, the element wOOM(i,j) corresponds to the j-th sample of the i-th OOM observation window. In view of this, the **MAP** vector could be mathematically extracted as follows:

$$\mathbf{MAP(i)} = \frac{\sum\_{j=1}^{Lw} \mathbf{w} \mathbf{OOM(i,j)}}{Lw} = \begin{cases} 1 & \mathbf{MAP(i)} > 0.5\\ 0 & \text{otherwise} \end{cases} \tag{3}$$

According with Equation (3), the outcome of this computation block consists of a 10-element vector (i.e., **MAP**). Each vector element corresponds to a muscle and it is 1 if the considered muscle is active (contracted) for more than half of the observation time, otherwise 0 (i.e., time predominance rule).

All the MAPs collected during a first brief stage of unperturbed gait or other ADL allow the system to build a first muscular behavior statistic. Specifically, they are used to extract a set of weights. These weights are based on the occurrence of a specific muscle contraction in correspondence of the MT contractions comprising the database. Two weights vectors are derived, one for the right leg (RL) movements and one from the left leg ones (to avoid asymmetry issues). In this way, it is possible to extract the most probable muscular pattern and, thus, a scoring method able to provide a high score if the incoming MAP is similar to the standard pattern, otherwise a low score (in presence of anomaly such as a perturbation). The weights vectors are continuously updated when requested by classification block, according to changes in user rhythms.

#### 2.2.6. Muscular Activity Pattern (MAP)-Based Scoring Section

In a real-time application context, the MAP-based scoring block (Figure 1) analyzes the incoming MAP binary vector by dot-multiplying it by the related weight vector. For instance, MAP coming from the right Gastrocnemius contraction is dot-multiplied by the right leg-related weight vector and, finally, normalized. The score assignment outcome tends to 1 if the incoming MAP is similar to the muscular standard, otherwise, it tends to 0.

Naming **WR** <sup>∈</sup> **<sup>R</sup>**1, nEMG the weight vector from right leg movements, and **MAPR** the resulting vector from Equation (3) when the MT is the right Gastrocnemius, the contraction score can be mathematically derived as:

$$\text{Score RL} = \frac{\sum\_{i=1}^{\text{nEMG}} \mathbf{MAP\_{R}(i)} \mathbf{W\_{R}(i)}}{\sum\_{i=1}^{\text{nEMG}} \mathbf{W\_{R}(i)}} \tag{4}$$

where score RL is the score related to a generic MT contraction from the right leg. Equation (4) can be easily extended to a MT contraction of the left leg MT, by changing the subscript R with L. In this latter case, the score is named score LL (LL, left leg) and it is derived via Equation (4) by considering the proper weight vector **WL** and MAPs from the left leg, **MAPL**. A demonstrative example of the general muscular score during experimental walking to slip test is shown in Figure 3. The general score includes scores from right leg contractions (score RL) and, also, from left ones (score LL). The figure also shows a preview of the dynamic threshold extracted during the Calibration phase, detailed in Section 2.2.8 (red dotted line).

**Figure 3.** General muscular score (score right leg (RL) ∧ score left leg (LL)) during experimental walking to slip test (Sub. 3 – Trial 4).

#### 2.2.7. Cortical Scoring Section

The score assignment embedded in the cortical unit passes through two main steps: the generalization and the lateralization assessment. The generalization step aims to reduce the data to be analyzed (65 vectors of *m*ˆ values from 13 channels and 5 bands of interest), providing a qualitative control about the subject's general cortical involvement. In this respect, the generalization step considers the *m*ˆ values on four cortical groups, which roughly identify functional macro areas:


This means that, considering an incoming i-th contraction, the system extracts 20 *m*ˆ values (one per each functional group extended to 5 bands of interests). To clarify the concept, let us consider the α band involvement on the SMA. The generalized *m*ˆ value on the functional group SMA, considering the α band on the i-th contraction can be derived by the following equation:

$$
\hbar\_{\text{SM}A,a}(i) = \left(\hbar\_{\text{C}\lambda,a}(i) + \hbar\_{\text{C}4,a}(i) + \hbar\_{\text{Cz},a}(i)\right) \tag{5}
$$

The notation can be easily extended to the other formula parameters.

By contrast with the generalization step, the lateralization one evaluates the incidence of the power increment on a specific side (i.e., left or right) by analyzing the ratio between two specific macro areas: the left side containing {F, C, P}3 and the averaged {Cp1, Cp5}, and the right side that involves {F, C, P}4 and the averaged {Cp2, Cp6}.

The double-check implementation (i.e., generalization and lateralization) is justified by literature findings [19–23], which demonstrated that a reactive response leads to a widespread cortical involvement, while during unperturbed steps or non-challenging ADL, the cortical response remains more lateralized according to the limbs involved in the movement.

The Cortical Scoring block provides 20 values from the generalization step (one per each functional group on 5 bands of interests) and 5 values from the lateralization one (ratio between left and right-side involvement).

#### 2.2.8. Logic Network-Based Classification

The logic network-based classification block concludes the system workflow according to Figure 1. It consists of two phases: the system adaptive calibration and the logic-network based classifier.

The system adaptive calibration oversees extracting dynamic thresholds (Thr—Figures 1 and 3) for every neuromuscular parameter involved in the classification, i.e., muscular score and the 25 values from the cortical generalization and lateralization steps.

Since the proposed architecture does not embed a learning phase, providing an auto-adaptive turnkey solution, these thresholds are continuously refreshed, contraction by contraction, by means of a sliding observation time window. In this window, the system checks the presence of thresholds lowering via statistical methods (some ADL can drag down the thresholds more than others). If a lowering is recorded, the thresholds are automatically adapted to the next value.

The role of these thresholds is to make the neuromuscular values as handleable as possible, for example, by associating a binary alert to each unexpected behavior. For instance, if the muscular score is below its dedicated threshold (red arrows in Figure 3) the muscular alert goes ON.

In a similar way, the procedure can be applied to the resulting *m*ˆ values.

The main goal of the implemented classifier is to cross-relate, among each other, these binary alerts from muscular and cortical sides. Specifically, the classification stage implements a logic network developed on 3 levels as shown by Figure 4.

**Figure 4.** Logic network-based classifier. The flags that contribute to the classification stage are reported in red dotted boxes.

The 1st level considers the binary alerts from the 4 macro cortical areas (e.g., *m*ˆ *SMA*,α(*i*)- Figure 4)), all over the 5 bands of interest. The system verifies the presence of a widespread increase in brain

signal power. If more (>) than 2 cortical areas, for each evaluated band, are involved in the power increment, the architecture sets a generalization flag (GFα—Figure 4) to 1, otherwise 0.

Generalization flags (GFs) from all the evaluated bands are further analyzed. If more (>) than 2 bands are involved in the brain power increase, the 1st level flag, F1 (i) in Figure 4, goes to 1.

The 2nd level analyzes the ratio between the left and the right cortical side (x/y—Figure 4) as described in Section 2.2.7. If the ratio is higher than 1+ε or lower than 1-ε, with ε specific tolerance (~), a lateralized increment is formally recorded. Similarly, the system generates a binary flag, named LFα in Figure 4, which is equals to 1 if a lateralized brain activity is recognized. The system checks the number of lateralization flag as shown in Figure 4: if less (<) than 2 lateralization flags are active, the 2nd-level outcome (F2 (i) in Figure 4) goes to 1.

According to Figure 4, both the 1st-level output (i.e., F1 (i)), from generalization assessment, and the 2nd-level output (i.e., F2 (i)), from lateralization check, are sent to a final AND gate. If both the flags are ON, it means that the system recognized a not-lateralized increment of the cortical involvement. Finally, the classifier runs the 3rd level. This level considers the outcome from the AND gate, toggling the presence of a muscular alert (MA(i)—Figure 4) from the MAP-based scoring block (Figure 1).

Ultimately, if a not standard muscular behavior, jointly with a widespread and not lateralized cortical behavior, is found, the system classifies the i-th contraction as a potential loss of balance.

The classification output of this logical network can be used to enable a fall-prevention strategy (e.g., through wearable robotics and exoskeletons).

#### **3. Results**

The proposed system has been validated in near-fall scenarios and ADL-like tasks. During the walking-to-slip test protocol all the participants were secured by a safety harness attached to an overhead as shown in Figure 2a and no falls were reported during the trial. Participants were able to perform multistep recovery reaction to find back their balance.

Before starting the experimental sessions, all participants signed informed consent. Research procedures were in accordance with the Declaration of Helsinki and were approved by the Local Ethical Committee (Prot. no. 0028266/2019).

Section 3.1 briefly recaps general performance: sensitivity, detection time and specificity concerning steady walking steps versus near fall scenarios. Section 3.2 focuses on the daily life suitability of the system, discussing the method robustness against ADL. Section 3.3 briefly outlines the acquisition equipment features.

#### *3.1. Architecture Performance: Loss of Balance versus Steady Walking*

As already stated in the state-of-the-art comparison, the performance of a near-fall detection strategy is usually quantified in terms of accuracy and efficiency. According to all the evaluated works [11–15], the accuracy can be evaluated by considering the sensitivity and specificity parameters. Mathematically, the sensitivity can be defined as:

$$\text{Se } (\%) = (\# (\text{TrNF}) / \text{N}\_{\text{LoB}}) \cdot 100 \tag{6}$$

where #(TrNF) is the number of correctly detected near fall events (i.e., induced slippages) and NLoB represents the total number of evaluated loss of balance situations.

In a complementary way, the specificity is identified as:

$$\text{Sp} \left( \% \right) = \left( \# \text{(TrADL)} / \text{N}\_{\text{ADL}} \right) \cdot 100 \tag{7}$$

where #(TrADL) is the amount of successfully detected ADL-like actions (i.e., walking steps, sudden curves, TUG and obstacle avoidance) and NADL is the total number of the evaluated ADL related trials.

In this section, we analyze results recorded during a real-time application of the walking-to-slip protocol. Two data pools were built: the first dataset is composed of 60 contractions from near-fall scenarios (10 perturbations per 6 subjects), while the second dataset includes 2091 contractions from the steady walking of all the subjects.

The experimentally extracted sensitivity, specificity (related to walking steps) and the detection time values are summarized in Table 2. The proposed multi-sensor architecture shows a sensitivity of 93.33 ± 5.16% and a walking steps vs. loss of balance (slip) specificity of 98.91 ± 0.44%.



<sup>1</sup> **SpWS**: specificity strictly related to walking steps as not loss of balance actions; <sup>2</sup> **Avg**: averages among all the analyzed subjects (Sub. 1-6); <sup>3</sup> Values refer to the highest maximum and lowest minimum values among the reported data.

The system detection time is about 370.62 ± 60.85 ms, of which—on average—only 21.75 ms are dedicated to the overall computation chain for muscular and cortical units. The computation time comprises: (i) muscle ON/OFF pattern extraction (ii) sliding window FFT, (iii) band multiplexing, (iv) generalization and lateralization step (v) logic network-based classification and (vi) re-calibration of thresholds. Table 2 also shows that in the worst case (i.e., Sub 2, Trial 5) the system demands about 634 ms to intervene, while in the best case (i.e., Sub 6, Trial 4), the system recognizes the loss of balance in about 160 ms.

#### *3.2. Architecture Performance: System Robustness against Activities of Daily Life (ADL)*

To ensure the daily-life applicability of the proposed architecture, the wearability of the device is not the only constraint. Another important applicability limit lies in the system's robustness against movements that usually a subject does in his/her domestic environment. These are generally named Activities of Daily Life (ADL). In this respect, this section focuses on the system specificity characterization considering three ADL-like actions as: (i) sudden (and tight) curves, (ii) chair transfers (via TUG test) and (iii) obstacle avoidance.

For the sake of completeness, distinct datasets have been created starting from a real-time application of the system to the three tasks. Each test trial shown in the following consisted of a mixed pattern of these three tasks: walking with sudden curve, TUG and obstacle avoidance. Overall, the offline extraction of specific contractions resulted in a first dataset of 331 contractions (3 subjects) related to tight curves, a second dataset of 512 contractions (3 subjects) from the TUG test and, finally, a dataset that includes 352 contractions (3 subjects) related to the obstacle avoidance.

The steady walking specificity has been evaluated in the previous section by using a dataset of 2091 unperturbed steps (no recovery and near perturbed steps), leading to a value of 98.91 ± 0.44% that we assume as a final characterization parameter for the sake of readability.

Figure 5 provides a graphical characterization of the system robustness against the ADL.

**Figure 5.** Cortico-muscular involvement planes for the 3 analyzed subjects. The figure merges contractions from three different Activities of Daily Life (ADL): curves (black), timed up and go (TUG) test (red) and obstacle avoidance (blue). The panels show the false alarms (red circles) and a comparison group (blue crosses) that represents the typical values occurring during a fall-related master trigger (MT) contraction.

The figure shows three panels per subject, except for Sub. 2 that did not perform the obstacle avoidance during the third test trial. Each single panel shows a 2D plane in which the x-axis reports the muscular score (MS in the following) as defined in Section 2.2.3, while the y-axis refers to the *m*ˆ *SMA*,<sup>α</sup> values, obtained according to Equation (5). Each point on the plane has coordinates {MS(i), *m*ˆ *SMA*,α(i)} where "i" is the i-th MT contraction that led to features extraction. Each single point identifies two features that contributes to the final classification of the specific contraction.

The panel also shows 3 thresholds that are constant along the x-axis (solid red, blue and black lines). These thresholds are those that operate on the muscular score (i.e., MS), acting as shown in Figure 5 by considering the median on an observation window of 15 consequent contractions.

This means that the system slowly adapts to the worst thresholds among the evaluated ones.

For example, considering The Sub.1–Test: 2, the worst MS-related threshold is linked to the ADL3: obstacle avoidance (Figure 5). It means that during its steady functioning, the system will recognize as a real dangerous situation only MS below the obstacle avoidance related threshold (i.e., ADL3 Musc. Thr.—Figure 5).

The panels also provide y-axis constant thresholds that refers to virtual upper limits for *m*ˆ *SMA*,α. These thresholds act in a similar manner of the previously presented ones. Considering for example data from Sub.3–Trial 2 in Figure 5, this means that during its steady functioning the system will recognize a dangerous situation only for *m*ˆ *SMA*,<sup>α</sup> above the obstacle avoidance threshold (i.e., ADL3 *m*ˆ *SMA*,<sup>α</sup> Thr.—Figure 5). Ultimately, real dangerous situations for both muscular and cortical involvements lie in the top-left rectangles delimited by the leftmost MS-related threshold and the highest *m*ˆ *SMA*,α-related threshold. On these panels, some contractions wrongly recognized as losses of balance highlighted via red circles. In this context, we must consider it as incorrect classification, reducing the specificity as per Equation (5).

Finally, for the sake of comparison, Figure 5 reports as blue crosses some slip-related coordinates {MS (i), *m*ˆ *SMA*,α(i)}. These coordinates have been extracted from the walking versus slip dataset (Section 3.1) considering two points per each analyzed subject. This shows how the two groups ADL1,2,3 and slips could be easily divided in clusters.

To give a complete overview of the system robustness against ADL, Table 3 summarizes the experimental results in terms of specificity from each protocol carried out (ADL 1, 2 and 3 Sp. (%)), as well the single test (Tasks 1, 2, 3 Sp. (%)) and subject-related (Sub. Sp. (%)) specificities. Analyzing data in Table 3 and in Figure 5 it is possible to state that, overall, the system showed a specificity of 98.91 ± 0.44% in steady walking steps' recognition (see Table 2), 99.62 ± 0.66% in sudden curves successfully detection, 98.95 ± 1.27% of correct recognition in contractions related to TUG tests. Finally, during the balance-challenging obstacle avoidance protocol the specificity reached 98.43 ± 0.88%.


In order to provide a comparison with the state-of-the-art solutions, already presented in the introduction, Table 4 summarizes some specific features for each analyzed work. In particular, Table 4 focuses on recognized classes, the system performance in terms of accuracy and efficiency and the applicability tabs. Data from Table 4 show how the proposed system ensures very competitive specificity value (i.e., 98.9%).


**Table 4.** Comparison of proposed architecture's performance.

<sup>1</sup> The Sp(%) value has been evaluated as the average between the four specificity values (i.e., steady walking, curves, TUG, obstacle).

#### *3.3. Acquisition Equipment Features*

Once the algorithm robustness against ADL recognition is verified, the system's applicability to ordinary life imposes another constraint: wearability. The chosen equipment should address the wearability constraints, which according to [36] can be summarized briefly in three macro-categories: (i) encumbrance (ii) biomechanical effects and (iii) comfort.

Considering the former constraint, the physical dimensions of the wearable will be paramount. These dimensions include the size, weight and the distribution of the weight of the wearable on the body.

Secondly, the functional placement of the sensor nodes may affect the posture and musculoskeletal loading of the wearer. Finally, the sensors' node placement must avoid discomfort, favoring regular movements (e.g., walking or sitting) and non-biased postures.

To continuously analyze and characterize the subjects' cortical and muscular dynamics in several different ordinary life scenarios, the here-proposed experimental setup consists of a 32-channel wireless EEG headset (g.Nautilus Research) by g.Tec [28] and 10 wireless surface EMG nodes (Cometa Wave Plus) by Cometa Systems srl [29]. Table 5 provides information about the acquisition equipment comprising the set-up. For each device, the table reports the number of monitored nodes or channels, equipment features such as the size and weight, as well as the electrode characteristics and device parameters: wireless transmission range and protocol, resolution and sampling frequency. Table 5 demonstrates how the equipment choice ensures a fully wireless and low-encumbrance solution, validating the applicability in an indoor monitoring scenario. Despite this, the use of gel-based or pre-gelled electrodes could be considered uncomfortable for long-time acquisition. In this respect, the system can be considered reliable for 4 hours' acquisition, before the need to refill the gel to ensure the right input impedance to the amplifier.



#### **4. Discussion**

The detection of near falls is an emerging area of research that is contextually growing with the development of an increasing number of miniaturized and power-efficient wearable devices [7]. Supported by accumulated evidence on fall detection [7–23], the clinical utility of this investigation involves the unobtrusive and continuous monitoring of activities of daily life in populations at high risk of falling. This kind of strategy (i.e., near fall detection) can be useful to identify issues to be further addressed to prevent falls, associated injuries or simply improve the efficiency of already existing pre-impact fall-detection architectures.

In this study, we have proposed a novel wearable architecture that exploits electrophysiological signals from brain and lower limb muscles to discriminate a near-fall scenario (i.e., unexpected slippages) from an activities of daily life. The proposed system realizes a turnkey solution, which can adapt its function to the user neuromuscular rhythms, without any long and fatiguing learning stage.

Results in Section 3.1 showed how the proposed architecture demands about 370.62 ± 60.85 ms to carry out a binary classification (i.e., ADL vs. near fall). As stated in the same section, the overall computation chain of muscular and cortical units requires, on average, 22 ms to be completed.

The remaining time, i.e., ~350 ms, with its high variability (see Table 2), is related to the muscle that has been selected as a master trigger. In fact, it should be reminded that the system starts working from the contraction onset of the gastrocnemius (right or left independently).

The times related to this physiological process remain hard to determine with certainty. In this respect, the response times of the gastrocnemius constitute unavoidable delays in recognizing losses of balance and largely determine the efficiency of the system. Further investigation should be conducted in order to find another muscle bundle that can: (i) uniquely define a gait phase, (ii) activate itself faster in a perturbation context, (iii) ensure repeatability during the contraction timing when the near-fall scenario occurs.

The detection times achieved are competitive with respect to the state-of-the-art solutions, highlighting the system applicability in contexts of postural recovery strategies implementation [17].

Concerning the results in Section 3.2 that analyze the system robustness against the ADL, an interesting evaluation should be undertaken into the losses of balance detected below the worst thresholds, such as Sub1-Test1 and Sub2-Test2 in Figure 5. In these cases, offline checks verified that the threshold was slowly adapting to the final value, causing transitional "false alarm" (wrongly loss of balance detection). Within this, the threshold adapting procedure should be improved and speeded up, while keeping high sensitivity and specificity.

Another noteworthy case is that shown in Sub3- Test 2 (Figure 5). In that case, it seems that a loss of balance is detected below the cortical thresholds. It is important to remember that these panels show only the *m*ˆ *SMA*,<sup>α</sup> values, nevertheless, we must consider that the analyzed problem is hyper-dimensional, because we should take into account other four bands of interest and the remaining 3 cortical groups.

Moreover, in Figure 5 it is notable that the Sub. 1 experienced a high number of false alarms with respect to the following two subjects. This result could be related to the protocol improvement asked on-going to the last two participants. This improvement mainly concerns the sit-down and stand-up movements during the TUG. In fact, since the rejection algorithm rASR has not been optimized to reject the muscular artifacts from strong contractions of the deltoids, the EEG acquisitions were spoiled by unpredicted artifacts. In this respect, further investigations are still ongoing aiming to extend the range of applicability of the implemented rASR algorithm.

#### **5. Conclusions**

In this paper, we proposed and validated a novel architecture for the losses of balance recognition. The proposed system, optimized for unexpected slippages, addressed some still open challenges related to the daily life applicability of this kind of system. Design and verification constraints mainly concern the need for high specificity and system robustness against ADL. In this respect, the proposed algorithm has been tested on five different tasks: sudden curves, chair transfers via the timed up and go test, balance-challenging obstacle avoidance and, of course slip-induced loss of balance. To ensure the ordinary life suitability, the proposed architecture has been fully based on wearable and wireless acquisition devices. Specifically, the architecture exploits electrophysiological measurements from 10 EMG electrodes and 13 EEG channels. The collected data are analyzed by the muscular unit, hosted by a STM32L4 microcontroller, and the cortical unit, which is implemented on a central computation unit via Simulink modeling. The first realizes a binary ON/OFF pattern from muscular activity (10 EMGs) and triggers the cortical unit that evaluates the contraction-related cortical involvements in terms of EEG responsiveness.

This parameter is evaluated as the variation behavior in the EEG PSD, considering five bands of interest. The neuromuscular features from both the computation units are sent to a clinical evidence based logical network. It embeds a set of automatically adaptive thresholds, which follow the user rhythms. Experimental validation on 9 healthy subjects showed that the system could react in a time compliant with fall-detection architectures constraints (i.e., 370.62 ± 60.85 ms). It also ensures a fall detection sensitivity of the 93.33 ± 5.16%. During the ADL tests the system showed a specificity of 98.91 ± 0.44% in steady walking steps' recognition, 99.61 ± 0.66% in successful sudden curves detection, and 98.95 ± 1.27% of correct recognition in contractions related to TUG tests. Finally, during the balance-challenging obstacle avoidance protocol the specificity reached the 98.42 ± 0.90%.

These preliminary results show promising accuracy values that, jointly with the system wearability (wireless acquisition devices), make the system potentially suitable for daily life application. Moreover, the achieved detection time (i.e., ~371 ms) is conservatively below 550 ms, which is considered as the maximum intervention limit for the implementation of countermeasures aimed at restoring the balance of the subject [17]. It ensures the system applicability to improving a fall-detection strategy.

Some drawbacks that need future larger and higher quality studies concern the acquisition devices and the generalization of the implemented method. In fact, the use of wireless sensors (EEG/EMG) theoretically ensures the system's wearability. Nevertheless, future perspectives concern the study of more comfortable solutions able to provide the same electrophysiological patterns, e.g., by using textile-based sensors arrays. The second weak point under investigation is the muscle to be selected as the master trigger to provide a quasi-deterministic delay, improving the system efficiency as a logical consequence.

**Author Contributions:** Conceptualization, G.M. and D.D.V.; methodology, G.M. and D.D.V.; software, G.M; validation, G.M.; formal analysis, G.M.; investigation, G.M. and D.D.V.; resources, D.D.V.; data curation G.M.; writing—original draft preparation, G.M. and D.D.V.; writing—review and editing, G.M and D.D.V.; supervision, D.D.V.; project administration, D.D.V. All authors have read and agreed to the published version of the manuscript. **Funding:** This work was supported by the project AMICO (Assistenza Medicale In COntextual awareness, AMICO\_Project\_ARS01\_00900) by National Programs (PON) of the Italian Ministry of Education, Universities and Research (MIUR): Decree n.267.

**Acknowledgments:** The authors would like to thank Vito Monaco, Federica Aprigliano, Silvestro Micera and the staff of Locomotion Biomechanics Lab at The BioRobotics Institute of the Scuola Superiore Sant'Anna (56025 Pontedera, Pisa, Italy), for their valuable contribution and assistance in data collection, for making available equipment for the experimental validation, and for the time/resources spent in the experimental investigation.

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

#### **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/).

MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel. +41 61 683 77 34 Fax +41 61 302 89 18 www.mdpi.com

*Sensors* Editorial Office E-mail: sensors@mdpi.com www.mdpi.com/journal/sensors

MDPI St. Alban-Anlage 66 4052 Basel Switzerland

Tel: +41 61 683 77 34 Fax: +41 61 302 89 18