Design recommendations for automated balance training systems

Introduction: Impaired balance significantly increases fall risk, often resulting in severe injury or death (Kumar et al., 2025). Balance rehabilitation improves outcomes (Badke et al., 2004; Sherrington et al., 2008) and physical therapists (PTs) iteratively individualize care while considering patient deficits, safety, goals, and clinical expertise (Haas et al., 2012). As telehealth and digital health expand (Shaver, 2022), wearable sensors, coupled with machine learning (ML), are powerful tools to support clinicians. However, the creation of clinically viable systems requires aligning ML model capabilities with clinicians’ workflows and decision-making logic beyond focusing on model performance alone. This work presents a human-centered synthesis of three complementary studies integrating insights from kinematic balance data, eye-tracking gaze patterns, and qualitative think-aloud reasoning to establish system design recommendations for automated balance assessment and progression.Methods: Across three IRB-approved studies, adults with and without balance deficits performed standing balance tasks on firm and compliant surfaces while instrumented with full-body inertial measurement units (IMUs). Expert PTs provided balance performance ratings, determined exercise progressions (easier/similar/harder), and provided verbal rationales for their decisions either in-person (while also wearing eye-tracking glasses) or via video. These studies were synthesized to create design-oriented recommendations informing the multidimensional data needed to guide clinically meaningful approaches for automated balance rehabilitation systems.Study 1 (n=47 balance participants, 42 PT participants): Trained a convolutional neural network (CNN) model on data from 13 IMUs to predict PT participants’ performance ratings and conducted a sensor-reduction analysis. The results demonstrated the feasibility of an ML model and whole-body IMU data to predict balance performance with variability comparable to that seen in PTs (RMSE=0.66), with the predictive performance level being maintained with as few as four sensors placed on the thighs, lower back, and upper back (Jabri et al., 2026).Study 2 (n=11 balance participants, 11 PT participants): Analyzed PT participants’ gaze metrics (visit count and duration) to each body segment across performance assessments. The analysis of PTs’ visual behaviors found that PTs, on average, considered the whole body in their decisions and placed a high visual priority on the legs across all balance performance levels (Nigrelli et al., 2025).Study 3 (n=47 balance participants, 46 PT participants): Used a mixed-methods approach to evaluate the frequency and underlying thematic drivers of PT agreement/disagreement on exercise progressions. The quantitative analysis identified that PTs disagreed on the difficulty of the next exercise to be performed over 25 % of the time. A thematic analysis identified that disagreements were often driven by differing interpretations of the same observations and that, even in agreement, PTs sometimes discussed differing goals for a progression (in progress).Results: This synthesis yielded four core design recommendations for automated balance clinical systems:1. Wearable sensors are valid tools to capture balance observations: Whole-body movement observations are foundational to PTs’ decisions (Nigrelli et al., 2025) and have been used with ML to successfully predict balance ratings (Jabri et al., 2026). IMUs combined with ML are viable tools to capture and learn motion features representative of these clinical observations.2. Model inputs should consider observations, but do not need to directly mirror human visual inputs: While eye-tracking reveals that PTs visually consider the entire body (with emphasis on the legs) (Nigrelli et al., 2025), ML models achieve optimal predictive stability with a minimal subset of just four sensors (thighs, upper back, lower back) (Jabri et al., 2026). Model sensor data inputs can represent PTs' whole-body observations, but do not need to include sensors on every body segment to achieve acceptable results.3. System architectures should be context-aware to explain variability in decisions: PTs may disagree on progression difficulty (>25 % of the time) and may express differing interpretations of identical observed movements or divergent therapeutic goals (in progress). Models using IMU data can achieve variability comparable to PTs (Jabri et al., 2026), but kinematic data alone is insufficient to capture the variability in PTs’ decisions. Model features should incorporate patient-specific contexts and individual clinician mental models because there may be patient or clinician factors that inform a final decision beyond the observed performance alone.4. Models should support probabilistic recommendations: To further capture PTs’ recommendation variability, future systems should avoid providing a single deterministic recommendation and rather provide multiple or probabilistic options. Providing options or probabilities in model outputs can capture the various decisions different PTs may consider and support decisions in patient-clinician balance training scenarios, as different progression strategies can achieve desired outcomes.Conclusions: This body of work guides data collection and implementation for the use of wearable motion sensing to support balance rehabilitation decision making. Fully automated balance rehabilitation systems cannot rely on kinematic data alone. Integrating multi-modal insights on how clinicians gather, prioritize, and interpret movement will allow future automated systems to shift from rigid classifiers to flexible and context-aware decision support tools that safely align with clinical reality.Acknowledgements: This work is based upon material supported by the U.S. National Science Foundation under Grant No. 2125256 and the University of Michigan AI & Digital Health Innovation Precision Health Investigators Award. This synthesis builds upon published work by W. Carender ((Jabri et al., 2026), (Nigrelli et al., 2025), in progress), C. DiCesare ((Jabri et al., 2026)), J. Hauth ((Jabri et al., 2026)), X. Huan ((Jabri et al., 2026)), S. Jabri ((Jabri et al., 2026)), E. Nigrelli ((Nigrelli et al., 2025), in progress), D. Shin (in progress), K. H. Sienko ((Jabri et al., 2026), (Nigrelli et al., 2025), in progress), L. Stirling ((Jabri et al., 2026), (Nigrelli et al., 2025), in progress), L. Ojeda ((Jabri et al., 2026)), and J. Wiens ((Jabri et al., 2026)).