Optimized assessment of physical rehabilitation exercises using spatiotemporal, sequential graph-convolutional...

Area of Science:

• Rehabilitation Medicine
• Artificial Intelligence
• Biomedical Engineering

Background:

• Physical rehabilitation requires continuous monitoring of patient progress to optimize recovery.
• Current clinical assessments can be time-consuming and may not capture the nuances of exercise execution.
• Automated systems are needed to provide objective and efficient feedback on rehabilitation exercises.

Purpose of the Study:

• To develop a lightweight, graph-based deep learning model for automatic assessment of physical rehabilitation exercises.
• To provide a continuous quality score for patient movements, aiding in supervision and complementing clinical evaluations.
• To improve the efficiency and accuracy of rehabilitation exercise assessment.

Main Methods:

• A deep learning model utilizing two sequential graph convolutional networks (GCNs) to process 3D skeleton sequences.
• The first GCN learns spatial features of joint relationships, while the second learns temporal features from frame correlations.
• An added classification phase ensures scores are exercise-specific and based on complete movement demonstrations.

Main Results:

• The proposed model achieved state-of-the-art performance on the KIMORE and UI-PRMD datasets for quality score prediction.
• Demonstrated superior efficiency compared to existing methods.
• Successfully generated continuous quality scores for assessing rehabilitation exercise performance.

Conclusions:

• The developed graph-based deep learning model offers an effective and efficient solution for automatic physical rehabilitation exercise assessment.
• This approach can serve as a valuable tool for patient supervision, potentially reducing the need for frequent clinical examinations.
• The model's performance highlights the potential of AI in enhancing rehabilitation outcomes.