Error-Guided Pose Augmentation: Enhancing Rehabilitation Exercise Assessment through Targeted Data Generation

Effective rehabilitation assessment is vital for tracking patient progress, especially in home-based settings. However, current systems struggle with data imbalance and detecting subtle movement errors. This paper presents Error-Guided Pose Augmentation (EGPA), a novel method that generates synthetic skeleton data by simulating clinically relevant movement mistakes, unlike random augmentations. Integrated into an attention-based graph convolutional network, EGPA significantly improves performance across multiple rehabilitation tasks. Experiments show Mean Absolute Error reductions of up to 27.6% and error classification accuracy gains of 45.8%. Attention maps further highlight EGPA’s ability to guide the model toward clinically important joint relationships, offering interpretability. EGPA thus enhances movement quality assessment, with strong potential for clinical and home-based rehabilitation use.