Balance Therapy With Hands-Free Mobile Robotic Feedback for At-Home Training Across the Lifespan

Providing aging adults with engaging, at-home balance therapy is essential to promote long-term adherence to unsupervised training and to foster independence. We developed a portable interactive balance training system that provides real-world visual cues on balance performance using wobble board tilt angles to control the speed of a robotic car platform in a three-dimensional environment. The goal of this study was to validate this mobile balance therapy system for home use across the lifespan. Twenty younger (18-39 years) and nineteen older (58-74 years) healthy adults performed balance training with and without visual feedback while standing on a wobble board instrumented with a consumer-grade inertial measurement unit (IMU) and optical motion tracking markers. Participants performed feedback trials based on either the robotic car’s movements or a commercially-available virtual game. Wobble board tilt measurements were highly correlated between IMU and optical measurement systems (<inline-formula> <tex-math notation="LaTeX">$\text {R}{>}{0.84}$ </tex-math></inline-formula>), with high agreement in outcome metrics (<inline-formula> <tex-math notation="LaTeX">$\text {ICC}{>}{0.99}$ </tex-math></inline-formula>) and small bias (<inline-formula> <tex-math notation="LaTeX">$\text {mean}{ < }{3}\%$ </tex-math></inline-formula>). Both measurement systems identified similar aging, feedback, and stance type effects including (1) altered movement control when older adults performed tilting trials with either robotic or virtual feedback compared to without feedback, (2) two-fold greater wobble board oscillations in older vs. younger adults during steady standing, (3) no difference in board oscillations during steady standing in narrow vs. wide double support, and (4) greater wobble board oscillations for single compared to double support. These findings demonstrate the feasibility of implementing goal-directed robotic balance training with mobile tracking of balance performance in home environments.


I. INTRODUCTION
36 F ALL prevention programs that target postural control can 37 alleviate the pervasive balance decline associated with 38 aging [1], [2], thereby reducing fall risk between 25-50% 39 [3], [4]. However, balance therapy requires consistent, long-40 term adherence of at least six months to be most effective [4]. 41 This substantial training commitment surpasses the typical 42 six-week clinical treatment period [5], and is challenging to 43 maintain for older adults [6], [7]. Without supervision, indi-44 viduals must rely primarily on intrinsic motivation for train-45 ing, and compliance drops more than 50% at discharge [8]. 46 In addition to a lack of accountability to another individ-47 ual/therapist, unsupervised training often lacks evidence of 48 recovery progress. Without this information, the individual 49 may not appreciate the direct health benefits from training, 50 which reduces motivation and compliance [6], [9]. To address 51 this problem, mobile therapy instruments that convey quanti-52 tative at-home balance performance can empower individuals 53 to make informed healthcare decisions by providing them 54 with objective information on their training progress [10]. 55 In addition, remote access to mobile metrics can extend 56 supervised care without additional clinical visits to promote 57 long-term adherence [6], [11] for more effective treatments [4]. 58 Balance training with a multi-axial wobble board is 59 commonly prescribed for both preventative and rehabilitative 60 programs. Conventionally involving a series of steady stand-61 ing (i.e., minimizing level board oscillations) and dynamic 62 tilting motions [12], wobble board training improved bal-63 ance in young healthy participants [13], [14], [15], people 64 post-stroke [16], and healthy older adults [17], [18], [19]. 65 Wobble board training is also protective for reducing future 66 sports-related injuries [14] and improves ankle function in 67 young adults with chronic ankle instability [20]. Despite   Beyond tracking performance, real-time feedback during 121 training facilitates complex motor learning [24]; however, the 122 type and frequency of feedback affects performance [32], [33]. 123 For example, higher feedback frequency (i.e., every trial vs. 124 every other trial) is more effective for complex motor skill 125 learning and retention [32]. This learning benefit for a complex 126 task such as balance control may be due to the external shift 127 in attention focus that promotes automatic movement control, 128 Fig. 1. The robotic real-time feedback system consisted of (a) a robotic car platform (12 × 10 cm), (b) an IMU-based motion controller (6 × 5.5 cm), and (c) a physical modular maze (120 × 60 cm). retention, and transfer to daily activities [34], [35]. Because 129 feedback perspective also affects motor learning (i.e., third-130 person leads to more effective learning than first-person, [36]), 131 visual physical cues may evoke different balance coordination 132 behavior compared to virtual feedback. That is, mapping body 133 motion to a physical object in a real environment may promote 134 spatial awareness that improves neuromusculoskeletal balance 135 control in activities of daily living and be more intuitive 136 than a screen for older adults [37]. However, the effect of 137 interactive real visual cues on balance training performance 138 remains unknown. Further, screen-based technology has poten-139 tial health consequences with excessive use including links to 140 digital fatigue, attention deficits, and neurodegeneration [38], 141 [39], [40], and may also affect memory recall [41]. Therefore, 142 alternative feedback mechanisms for at-home therapy that do 143 not rely on a screen are needed.

144
The purpose of this study was to evaluate an affordable 145 (<$500) mobile balance therapy system with robotic motion 146 control to assess its potential for evidence-based, customized 147 at-home balance training. We tested the instrumented therapy 148 system's accuracy and validity for home use across the lifespan 149 by assessing its ability to (1) alter training performance 150 through goal-directed feedback, (2) detect balance perfor-151 mance during steady standing, and (3) quantify balance metrics 152 compared to gold standards to guide targeted therapy. The 156 mobile robotic feedback system (Fig. 1) included a motion 157 controller with an inertial measurement unit (IMU) and a 158 miniature robotic car with a microcomputer (Table I). 159 Euler angles were calculated in real-time with an embedded 160 system using the IMU signals (BNO055 9-DOF, Adafruit, 161 New York, New York) and transmitted wirelessly to a micro-162 computer (Raspberry Pi 3B+, Cambridge, United Kingdom) at 163 an unequal sampling rate (average: 40 Hz, range: 10-110 Hz). 164 Fig. 2. Human-in-the-loop control schematic of the robotic real-time feedback balance training system. Wobble board pitch (θ P ) and roll (θ R ) angles are calculated using the IMU motion-controller signals and transmitted wirelessly to the robot microcomputer. Pitch and roll angles are linearly mapped (β 1 =slope, β 0 =intercept) to car linear (v) and angular (ω) velocities, respectively, and converted to motor velocity for closed-loop proportional-integral (PI) speed control of each wheel. The central nervous system perceives the robotic real-time feedback, plans, coordinates and executes movements using the neuromusculoskeletal and sensorimotor systems. Brain image courtesy of [42].    (Sensbalance Miniboard Wireless, Sensamove, Groessen, 179 Netherlands) included a motion controller that mapped 180 wobble board tilting to on-screen maze movements (Fig. 3). 181 This system served as a clinically relevant comparison to the 182 robotic feedback system, with improved balance and therapy 183 engagement reported in older adults [17], [18], [24] and 184 functional balance reported in individuals post-stroke [43]. commercially-available wobble board (Wobblesmart, Wob-187 blesmart International, Denmark, Europe) was set to level one 188 (i.e., most stable setting) for all trials to eliminate the effect 189 of geometric board variations (Fig. 4). This setting provided 190 approximately 15 degrees of multi-axial tilt with a half-round 191 adjustable pivot (diameter=14 cm, height=6.5 cm, [44]).

192
Four retroreflective spherical markers (14-mm diame-193 ter) were placed on the wobble board's standing platform 194 (diameter=39 cm) and were tracked at a sampling rate 195 of 150 Hz (n=2) or 200 Hz (n=37) with a seven-camera 196 Fig. 4. The wobble board was instrumented with four reflective markers and tracked with an optical motion capture system. The IMU-based robot motion controller and virtual feedback sensor were aligned with the wobble board axes (+x-axis points forward) and secured to the board.    (Tables II,III). Participants prior to starting the study procedures, which were approved 226 by the Colorado Multiple Institutional Review Board (Protocol 227 21-2971).

C. Experimental Protocol 229
Participants completed wobble board balance training with-230 out real-time visual feedback and were randomly assigned to 231 also perform training with either robotic or virtual feedback. 232 The conventional wobble board training approach without 233 real-time feedback involved (1) controlled forward/backward 234 and side-to-side tilting maneuvers while avoiding touching the 235 board's edge to the floor, and (2) steady standing with the 236 board horizontal [12]. Participants focused their gaze on an 237 X marked on the floor 1-m away from their base of support 238 during trials without feedback. Data were simultaneously 239 collected for each trial with optical and IMU-based systems. 240 Two handrails were secured to the floor on either side of the 241 participant to mitigate fall risk; however, participants were 242 encouraged to perform the tests with arms out to their sides. 243 A standard arm position was not enforced to improve comfort 244 level during testing. All trials were performed unshod with or 245 without socks according to participant preference. 246 1) Dynamic Balance Training: For the tilting trials with-247 out real-time visual feedback (No Feedback), participants 248 performed two repeated 40-second trials per direction (for-249 ward/backward and side-to-side) while standing in narrow 250 double support with the medial borders of their feet touch-251 ing. Participants then performed seven minutes of structured 252 practice either driving the robot or moving the ball with board 253 tilting motion (Feedback). Next, they navigated the robot/ball 254 through a real (Fig. 1) or virtual (Fig. 3) maze as far as 255 possible during three repeated one-minute trials (Fig. 5). Par-256 ticipants inherently received performance feedback based on 257 their ability to progress the robot/ball toward the goal. Also, all 258 participants were given similar verbal encouragement between 259 trials, which was unrelated to their individual performance.  cross-correlation (Fig. 7).  We compared the optical and IMU-based motion capture 294 systems' quantification of balance performance outcome met-295 rics across feedback type, stance type, and aging groups. All 296 outcome metrics were calculated across a 30-second duration. 297 Angular path length was the sum of Euclidean distance (in 298 degrees) between successive samples of wobble board pitch 299 and roll angles (Fig. 7b). Mean absolute velocities were also 300 calculated as the average of the absolute value of individual 301 pitch and roll components of angular velocity in the laboratory 302 reference frame. 303 We performed linear mixed effects analysis with R (R Core 304 Team, 2021, v4.1.2) and lme4 [47] to assess the relation-305 ship between outcome metric and dynamic tilting balance 306 training with feedback type (no feedback, robotic or virtual 307 real-time feedback) and age group (older, younger) as fixed 308 effects. Participant intercepts were the random effects with 309 feedback types as the nested factor. Similar mixed effects 310 models were performed for the steady standing tasks with age 311 group and stance type (wide double support, narrow double 312 support, single support) as fixed effects and stance types as 313 the nested factor. Visual inspection of residual plots supported 314 homoscedasticity and normality assumptions. Estimated (i.e., 315 least-square) marginal means and model differences were used 316 to calculate the percent differences between age groups, stance 317 types, and feedback. 318 We quantified Pearson correlation coefficients of the 319 time-matched, zero-meaned signals to assess the correlation 320 between wobble board angles and angular velocities mea-321 sured with the optical and IMU-based systems during steady 322 standing and tilting movements with and without real-time 323 feedback in younger and older adults. We also performed 324 Bland-Altman analyses [48] and calculated Intraclass Corre-325 lation Coefficients (ICC) to study the agreement between the 326 optical and IMU-based outcome metrics.  Fig. 9; roll: 42-60% 332 Fig. 10) compared to the younger participants ( p<0.05). How-333 ever, when older participants performed robot or virtual feed-334 back, their angular path lengths and mean angular velocities 335 decreased to similar levels as the younger cohort ( p>0. 1,336 Tables S1-S2). The IMU and optical motion capture systems 337 detected similar differences in balance training performance 338 related to aging and feedback (within 2% difference across 339 metrics). type effects (Fig. 11). The older participants produced 54-57% 344 higher angular path lengths and 47-52% (pitch) and 57-63% 345 (roll) faster mean absolute angular velocities compared to the 346 younger cohort across all stance types ( p<0.01, Table S3).

347
Across age groups, single support stance on the non-dominant 348 limb produced 42-43% higher angular path lengths and  Effective balance therapy requires long-term adherence 374 to a customized training program. However, compliance to 375 unsupervised training is challenging, especially if the train-376 ing is insufficiently challenging and the health benefits are 377 unclear. Interactive mobile therapy instruments can motivate 378 participation in home training programs through real-time 379 feedback and remote monitoring of recovery progress by 380 storing performance metrics (e.g., angular path lengths and 381 velocities) over the course of a training session. These balance 382 scores could automatically be uploaded to a mobile device 383 and shared with a clinician to extend supervised training. 384 Understanding how feedback type and narrowing base of 385 support affect wobble board balance training in aging adults 386 provides valuable insights towards tailoring therapy programs 387 for more effective treatments. Therefore, this study validated 388 the performance of a portable, interactive balance therapy 389 system with hands-free robotic feedback by comparing it to 390 an established virtual feedback system and testing agreement 391 with lab-based measurements. The older adults performed isolated dynamic tilting tri-394 als without real-time visual feedback with potentially more 395 difficulty controlling rotation compared to younger adults, 396 Fig. 9. Mean absolute pitch velocity for dynamic tilting trials box and whisker plots [49]. The older participants performed higher mean absolute pitch velocity during isolated tilting trials without real-time feedback (No Feedback), which may indicate more difficulty controlling sagittal plane rotation compared to the younger group (p<0.05). With either robot or virtual feedback (Feedback), the older group slowed their pitch velocity (p<0.05) to similar levels as the younger participants (p>0.1). The IMU and optical motion capture systems demonstrated similar aging and feedback effects. training, similar to prior study [24]. However, additional 407 research is necessary to establish the relationship between 408 board tilting movements and improved dynamic balance con-409 trol. Additional benefits may also be realized because this 410 training approach combines anticipatory (i.e., route planning) 411 and compensatory (i.e., balance correcting) postural adjust-412 ments, which are affected by aging [25]. Through real-time adjustments [25], [52]. Further, cognitive engagement with 417 balance training can stimulate neuroplasticity for improved 418 balance recovery [53]. Therefore, goal-directed feedback bal-419 ance therapy has great potential to promote immediate and 420 long-term balance control in older adults.

425
This comparable outcome was achieved using the common 426 goal to move the robot/ball as far as possible, despite subtle 427 controller differences between systems. The robot also pro-428 vides interactive balance therapy without the use of a screen. 429 Therefore, performing this alternative therapy in a home 430 setting will not contribute to the increasing detrimental effects 431 of digital fatigue [54]. Translating the current system for home 432 use would require a mobile device application that provides 433 seamless operation, training outcome assessments, and data 434 uploading capabilities. In a home setting where handrails may 435 be unavailable, a doorway, furniture, or counter edge can be 436 used to ensure safety. Also, a specific track may not be needed 437 in the home, but rather any course could be established using 438 everyday objects as obstacles and/or following longer paths.

439
Whereas this study revealed similar dynamic balance train-440 ing performance between the two feedback systems, further 441 research is required to establish how feedback affects the 442 underlying biomechanics during balance training. Specifically, 443 subtle differences in muscular demand and kinematic strate-444 gies may exist between virtual and robotic feedback, and 445 these effects may differ across the lifespan. In addition, this 446 experiment was controlled to a subset of challenges for all 447 users. The highly variable response to feedback across users 448 suggests benefits may exist from tailoring the initial challenge 449 level to balance skill. For example, a more uniform response 450 across users may be realized by incorporating the spectrum of 451 available games, controller modes, and wobble board levels. 452 Therefore, further study is needed to assess the full potential 453 of each system and to guide progressively demanding therapy 454 Fig. 10. Mean absolute roll velocity for dynamic tilting trials box and whisker plots [49]. The older participants performed higher mean absolute roll velocity during isolated tilting trials without real-time feedback (No Feedback), which may indicate more difficulty controlling frontal plane rotation compared to the younger group (p<0.05). With either robot or virtual feedback (Feedback), the older group slowed their roll velocity (p<0.05) to similar levels as the younger participants (p>0.  25), which provides perspective on the 480 relative task demand for each group. That is, the muscular 481 demand may be similar for older adults in double support as 482 younger adults in single support on a wobble board, just as 483 increasing balance task demand by reducing the radius of the 484 half-round adjustable pivot corresponds to greater center of 485 pressure path lengths, muscle activity, and co-activation [44]. 486 However, further research on muscle demand during steady 487 wobble board standing in older adults is required because 488 aging likely elicits additional compensations. These findings 489 serve as a comparative tool to guide periodic at-home assess-490 ments for early detection of balance deficits with healthy aging 491 and monitoring training efficacy.

492
Reduced balance control due to narrowing base of support to 493 a single leg during steady standing was also detected by the 494 IMU-based wobble board metrics, with an increase of 45% 495 for both age groups. Surprisingly, narrowing double support 496 from hip width to feet touching did not increase the wobble 497 board oscillations, which implies similar whole body balance 498 control is required for both tasks. This finding contrasts the 499 increase in postural sway that occurs with narrowing double 500 support during level ground quiet standing [60] and implies 501 that narrowing double support stance width on a wobble board 502 does not progress the task demand as it does on level ground. 503 However, underlying muscle contributions may be affected by 504 stance width and not reflected by wobble board kinematics. 505 The presented results suggest that when used as a screening 506 tool, wide or narrow double support stance widths can be 507 used interchangeably. Therefore, individuals can self-select 508 a comfortable stance width, which may improve balance 509 Fig. 11. Steady standing outcome metrics (angular path length and mean absolute angular velocities across 30 seconds) were similar when measured by optical motion capture or IMU-based systems (<3± difference). Wobble board oscillations were greater for older compared to younger adults across stance types (p<0.01). Single support stance elicited greater board oscillations compared to both double support stances across age groups (p<0.01).

519
The two systems also demonstrated excellent agreement in 520 outcome metrics during steady standing and active tilting 521 motions, with Intraclass Correlation Coefficients that were 522 greater than 0.99 and mean system bias that was less than 3%.

523
The IMU and optical systems' outcome metric Intraclass 524 correlations were higher (+0.23 across metrics) than pre-525 viously reported correlations between IMU and force plate 526 center of pressure mean velocities [31], which is likely because 527 both systems in this study directly measured wobble board tilt 528 angle. The small negative bias indicates that the IMU slightly 529 underestimated tilt angle compared to the optical measure-530 ments, likely influenced by intermittent signal dropout from 531 radio transceiver interference. Collecting the IMU signal at a 532 consistently higher sampling rate (>10 Hz) with a more robust 533 communication protocol (e.g., Bluetooth) could mitigate this 534 bias. Despite signal dropouts (t<0.5 s), the IMU-based 535 system demonstrated similar ability to detect differences in 536 balance associated with aging, feedback type, and narrowing 537 base of support. In addition, robot speed was maintained 538 until the next available IMU command; therefore, the dropped 539 signals were imperceptible to the user.