Rehabilitation exercise quality assessment through supervised contrastive learning with hard and soft negatives

Area of Science:

• Artificial Intelligence in Healthcare
• Rehabilitation Medicine
• Computer Vision for Exercise Analysis

Background:

• Exercise-based rehabilitation programs improve patient outcomes but face challenges in remote monitoring.
• AI-driven virtual rehabilitation enables home-based exercise with automated data analysis and clinician feedback.
• Limited data per exercise type in rehabilitation datasets hinders the development of generalizable AI models.

Purpose of the Study:

• To develop a novel AI framework for robust rehabilitation exercise quality assessment.
• To address the challenge of limited sample sizes per exercise type in training datasets.
• To improve the generalizability and efficiency of AI models in virtual rehabilitation settings.

Main Methods:

• Implementation of a supervised contrastive learning framework incorporating hard and soft negative samples.
• Utilization of a Spatial-Temporal Graph Convolutional Network (ST-GCN) architecture for exercise data analysis.
• Training a single, generalized model applicable across diverse exercise types using the entire dataset.

Main Results:

• The proposed framework demonstrated enhanced generalizability across various rehabilitation exercises.
• A significant decrease in the overall complexity of the AI model was achieved.
• Experimental results on three public datasets (UI-PRMD, IRDS, KIMORE) surpassed existing methods.

Conclusions:

• The novel AI framework effectively overcomes data limitations in rehabilitation exercise assessment.
• The ST-GCN based model offers a new benchmark for accurate and generalizable exercise quality evaluation.
• This approach has the potential to significantly advance AI-powered virtual rehabilitation systems.