RehabMate: an explainable framework for action detection and corrective feedback in pediatric stroke rehabilitation

Pediatric stroke is a sudden cerebrovascular condition with a disability rate of up to 70%, often requiring long-term, structured rehabilitation. However, in home and community environments, limited access to professional rehabilitation resources poses significant challenges for recovery. To address this gap, we present RehabMate, an explainable framework specifically designed for lower-limb pediatric stroke rehabilitation, aimed at providing clear and actionable support to physiotherapists in delivering effective rehabilitation training. Unlike traditional ”black-box” AI models, RehabMate integrates a multimodal graph-based architecture with a spatial-temporal attention mechanism to enable interpretable assessment of movement quality and detection of incorrect rehabilitation actions. This interpretability allows physiotherapist to trace AI decisions back to specific spatial-temporal features and body segments, providing auditable and rehabilitation-meaningful insights to support individualized training planning and adjustments. In addition, RehabMate incorporates an enhanced language model that generates real-time, personalized corrective feedback and motivational support by combining action recognition results with a corpus curated by rehabilitation professionals. The feedback is adapted according to the child’s pain level, mobility, and age, ensuring safe and effective rehabilitation guidance tailored to individual needs. When evaluated on our custom PSP2 dataset, RehabMate achieved 93.3% top-1% accuracy in cross-view action recognition, significantly outperforming baseline models. By bridging the gap between model interpretability and practical applicability in physiotherapy rehabilitation, RehabMate demonstrates the potential of explainable AI to deliver trustworthy and implementable lower-limb rehabilitation support for pediatric stroke patients.