Multi-Dimensional Assessment of Rehabilitation Exercises with Multi-gate Mixture-of-Experts

Physical exercises prescribed in-home rehabilitation programs are crucial for individuals with physical disabilities to regain muscle strength and improve balance.However, patients often struggle to assess their motor performance without the guidance of medical professionals.Recent advances in vision-based sensors and deep learning have led to the development of AI-driven methods for automatically assessing rehabilitation exercises.Despite these advances, most existing approaches provide only a single Total Score (TS) to evaluate exercise quality, which offers limited practical guidance.To address this issue, we propose a novel multi-task evaluation model that generates multi-dimensional assessment results.Our model uses Graph Convolutional Networks (GCN) to effectively extract general motion features.Additionally, a Mixture of Experts (MoE) framework is employed to explicitly capture relationships between TS, Primary Outcome Scores (POS), and Control Factor Scores (CFS), generating score-specific motion features.A multi-gate fusion mechanism dynamically integrates general and score-specific motion features to predict diverse assessment scores for each exercise.By simultaneously generating TS, POS, and CFS, our method provides a more detailed and accurate evaluation.Extensive experiments on the KIMORE dataset demonstrate that our approach not only improves TS evaluation accuracy but also offers detailed, multi-dimensional assessment scores, providing valuable insights for personalized rehabilitation.