xAARA: Explainable and Augmented Automated Rehabilitation Assessment

Stroke rehabilitation assessments compress clinicians’ nuanced observations into coarse scores, obscuring the relationship between movement quality and functional recovery. Through the development of xAARA , an explainable augmented assessment system, we present a three-stage methodology for augmenting expert assessment in complex embodied clinical work. In development , co-design with ten clinicians over 30 months produced an assessment ontology decomposing tasks into segments and movement-quality elements, with multi-view capture and annotation interfaces. In deployment , embedding these tools in routine assessment of 107 stroke survivors produced a clinical-scale corpus and trained computational experts whose recommendations reached over 90% agreement with high-confidence expert ratings and resolved 92% of ambiguous cases. In adoption , clinicians external to development reviewed, accepted, or refined these recommendations, both validating the system and supplying the signal that tunes it. Throughout, assessment entropy is the measure by which augmentation is designed and evaluated: observability increases while uncertainty is reduced and localized where it can guide therapy.