EGCN++: A New Fusion Strategy for Ensemble Learning in Skeleton-Based Rehabilitation Exercise Assessment

Area of Science:

• Computer Science
• Biomedical Engineering
• Rehabilitation Medicine

Background:

• Skeleton-based exercise assessment analyzes movement correctness using skeleton data.
• Existing methods underutilize position and orientation features.
• Previous ensemble-based graph convolutional network (EGCN) improved performance but lacked comprehensive fusion.

Purpose of the Study:

• To present EGCN++, an advanced framework for rehabilitation exercise assessment.
• To introduce a novel fusion strategy, MLE-PO, for enhanced data and model-level integration.
• To evaluate the performance and clinical consistency of EGCN++ with the MLE-PO strategy.

Main Methods:

• Developed EGCN++, an enhanced framework building upon EGCN.
• Implemented the MLE-PO fusion strategy for data and model-level integration.
• Conducted cross-validation experiments on UI-PRMD, KIMORE, and EHE datasets.

Main Results:

• MLE-PO demonstrated superior performance compared to other EGCN ensemble strategies and baselines.
• EGCN++ with MLE-PO achieved higher quantitative consistency with clinical evaluations.
• The framework effectively leverages both position and orientation skeleton data features.

Conclusions:

• EGCN++ with the MLE-PO fusion strategy represents a significant advancement in skeleton-based exercise assessment.
• The proposed method offers improved accuracy and better alignment with human clinical judgment in rehabilitation.
• This framework holds potential for more objective and reliable evaluation of therapeutic exercises.