A lightweight real-time framework for interpretable rehabilitation motion quality assessment using wearable inertial sensors

Accurate and interpretable evaluation of rehabilitation motion quality remains challenging due to the subjectivity of manual scoring and the latency of existing sensor-based systems. This paper presents RehabNet, a lightweight and interpretable deep framework designed for real-time rehabilitation motion assessment using wearable inertial measurement units (IMUs). RehabNet integrates a Stage-Aware Multi-scale Attention (SAMA) module with temporal convolutional and recurrent encoders to capture both short- and long-term motion dependencies. The proposed SAMA mechanism adaptively highlights phase-specific motion cues, improving interpretability and robustness across diverse exercises. Extensive experiments on rehabilitation datasets demonstrate that RehabNet achieves 94.3% accuracy, MSE of 0.027, and inference latency under 10 ms, outperforming existing methods in both precision and efficiency. Furthermore, visual attention heatmaps reveal interpretable focus on critical movement phases, supporting clinical decision-making. These results indicate that RehabNet offers a practical and deployable solution for real-time, resource-efficient, and interpretable rehabilitation monitoring. λ A new AI system offers real-time, personalized feedback for rehabilitation exercises using wearable sensors. λ The model explains its decisions visually, helping clinicians understand and trust automated motion scoring. λ Designed for mobile and home use, it delivers fast, accurate, and interpretable rehabilitation monitoring.