LONG SHORT-TERM MEMORY IN POST-REHABILITATION EXERCISE CLASSIFICATION

The field of artificial intelligence (AI) and computer vision presents significant opportunities to advance physical rehabilitation. However, their application in rehabilitation assessment still remains underexplored. This study addresses the need for accessible and precise real-time evaluation of post-rehabilitation exercises, contributing to the United Nations Sustainable Development Goal 3: Good Health and Well-Being. The proposed system integrates BlazePose, a human pose estimation model, alongside a Long Short-Term Memory (LSTM) model to classify movement correctness and provide real-time feedback. The system was trained on three low back pain rehabilitation exercises from the KinesiothErapy and Rehabilitation for Assisted Ambient Living (KERAAL) dataset. The researchers found that a 13-keypoint skeletal configuration yielded the most optimal performance, achieving an F1-Score of 80% for flank stretch, 78% for torso rotation, and 77% for hiding face. Furthermore, with the same skeletal keypoint configuration and multiple independent training sessions, a Repeated Measures One-Way ANOVA confirmed statistically significant differences (p < .05) in F1-Scores across exercises, indicating varied model performance among the exercise types. The evaluation focused on assessing the robustness and consistency of the model across multiple training runs to ensure reliable performance measurement. The results provide empirical evidence supporting the applicability of deep learning–based pose analysis for automated rehabilitation exercise assessment.