A Compensatory-Topology and Relay-Gating Graph Network for multimodal stroke rehabilitation assessment

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Artificial Intelligence in Healthcare

Background:

• Accurate motor function assessment is crucial for effective stroke rehabilitation.
• Automated assessment faces challenges in differentiating task completion from compensatory movements.
• Skeletal and inertial sensors provide complementary but distinct movement data.

Purpose of the Study:

• To develop a multimodal framework, CTCG-Net, for therapist-assisted rehabilitation assessment.
• To integrate skeletal and inertial sensor data for improved motor function analysis.
• To accurately estimate Primary Outcome (PO) and Control Factor (CF) scores.

Main Methods:

• Proposed CTCG-Net, a multimodal spatio-temporal graph framework.
• Combined a compensation-aware Prior-Guided Spatio-Temporal Graph Convolutional Network (ST-GCN) with Kinematic Relay Gating.
• Utilized synchronized IMU dynamics to modulate skeletal features and a Spatially Decoupled Regression Head for score estimation.

Main Results:

• CTCG-Net achieved the lowest average errors on a public benchmark (MAD: 0.4001, RMSE: 0.5026).
• Ablation studies confirmed the contribution of compensation-aware topology and relay gating.
• Analyses demonstrated complementary information from different components for PO and CF estimation.

Conclusions:

• CTCG-Net effectively integrates multimodal sensor data for motor function assessment in rehabilitation.
• The framework supports quantitative, clinician-supervised monitoring, enhancing rehabilitation assessment accuracy.
• CTCG-Net shows promise for improving the precision of rehabilitation progress tracking.