Enhancing Clinician Decision-Making via Uncertainty-Aware Multi-Expert Fusion for Stroke Rehabilitation

Researchers present xAARA, an AI system that enhances stroke rehabilitation assessment by analyzing multi-view video to provide ARAT scores with calibrated uncertainty and clinical explanations. The system achieved 94.2% task accuracy while reducing predictive uncertainty by 96.1% compared to single clinicians, with four independent clinicians validating its potential for clinical deployment.

xAARA addresses a critical gap between technical AI capability and clinical utility in stroke rehabilitation assessment. Traditional instruments like the ARAT collapse nuanced movement quality into ordinal scores, while existing automated systems prioritize accuracy metrics over clinical integration, preventing real-world adoption. This work reframes the problem: instead of replacing clinical judgment, xAARA augments it through principled uncertainty quantification and explainability aligned with clinician workflows.

The approach reflects a maturing understanding in medical AI that clinical deployment requires more than high accuracy. By composing 692 calibrated multimodal models through a Dynamic Bayesian Network with entropy-based gating, xAARA generates assessment confidence estimates and defers low-confidence cases to human review. This human-in-the-loop framework addresses the legitimate skepticism clinicians have toward opaque algorithmic outputs. The system's 100% agreement with at least one rater on subjective cases and perfect adherence to clinical validity rules demonstrates domain-specific robustness beyond aggregate metrics.

The validation by four independent clinicians and reported willingness to adopt signals genuine clinical readiness, distinguishing this from academic demonstrations. This matters for the broader healthcare AI ecosystem, where assessment automation—particularly in rehabilitation—remains a bottleneck limiting patient access to standardized, frequent evaluations. xAARA's architecture provides a replicable model for other clinical domains requiring ordinal or subjective judgments where single-clinician assessment introduces variance and resource constraints.

Looking forward, clinical deployment success will depend on real-world integration with existing electronic health record systems and sustained clinician engagement. The framework's emphasis on uncertainty and explainability may become a standard expectation for regulated medical AI, influencing how developers approach clinical tool design across rehabilitation, mental health, and other subjective assessment domains.

• →xAARA achieved 94.2% task accuracy with calibrated uncertainty, reducing prediction variance by 96.1% compared to single clinician scoring
• →The system augments rather than replaces clinical judgment through uncertainty quantification and task/phase/quality-level explainability
• →Dynamic Bayesian Network with 692 calibrated models enables principled deferral of low-confidence cases to human review
• →Independent clinician validation and stated adoption willingness suggest genuine clinical readiness beyond typical academic demonstrations
• →Human-in-the-loop design with domain-aligned explainability represents a replicable model for clinical AI deployment across subjective assessment domains