AI Tool Enhances Stroke Rehab Assessment with Uncertainty-Aware Fusion

Tailoring effective stroke rehabilitation programs requires a nuanced assessment of movement organization, not just whether a movement is completed. Current assessment tools, such as the Action Research Arm Test (ARAT), often condense rich behavioral observations into single scores, losing critical details about movement quality that differentiate true recovery from compensatory strategies. Existing automated solutions typically prioritize accuracy on noisy labels, leading to opaque scores that rarely integrate into clinical practice. To address these limitations, researchers developed xAARA, an AI engine designed to enhance, rather than replace, clinical judgment. xAARA processes multi-view video to generate ARAT assessments, crucially providing calibrated uncertainty estimates and detailed explanations across task, movement-phase, and movement-quality levels. It achieves this by fusing 692 calibrated multimodal models via a Dynamic Bayesian Network with entropy-based gating, qualifying results against clinical validity rules and deferring low-confidence cases. In a study involving 105 stroke survivors, xAARA demonstrated 94.2% task accuracy and 81.3% movement-phase accuracy, reducing predictive uncertainty by 96.1% compared to single-clinician scoring. Clinicians validated the assessments and expressed willingness to adopt the system, highlighting the importance of principled uncertainty quantification and clinician-aligned explainability for deployable clinical tools.