Spatio-temporal graph attention network for rehabilitation movement classification

The automated assessment of home-based rehabilitation benefits from skeleton-based methods, which provide a robust and privacy-preserving means of monitoring patient movements. While such approaches effectively capture local body pose and spatial joint configurations, they often fall short in modeling the complete temporal coherence of an entire exercise sequence. To address this limitation, we construct a novel cascaded framework that combines a Spatio-Temporal Graph Convolutional Network (ST-GCN) for capturing local joint dynamics with a Transformer-based attention mechanism to model global, long-range dependencies. Evaluated on the UI-PRMD and KIMORE datasets under a rigorous 5-fold cross-validation protocol with independent test sets, STGA-Net achieved mean accuracies of 92.71% and 95.24%, respectively, surpassing existing state-of-the-art methods. These results demonstrate its strong potential as a quantitative and privacy-preserving tool for personalized rehabilitation assessment in both clinical and home settings.