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

Accurate assessment of motor function is central to stroke rehabilitation. However, automated assessment remains challenging because task completion must be distinguished from compensatory movement, and skeletal and inertial sensors capture different aspects of movement. We propose CTCG-Net, a multimodal spatio-temporal graph framework for therapist-assisted rehabilitation assessment. The framework combines a compensation-aware Prior-Guided ST-GCN with Kinematic Relay Gating, which uses synchronised IMU dynamics to modulate skeletal features. A Spatially Decoupled Regression Head estimates Primary Outcome (PO) and Control Factor (CF) scores separately. On a public multimodal rehabilitation benchmark, CTCG-Net achieved the lowest average errors among the evaluated methods, with a MAD of 0.4001 and an RMSE of 0.5026. Ablation, agreement, robustness, topology, and saliency analyses indicated that the compensation-aware topology and relay gating contributed complementary information to PO and CF estimation. CTCG-Net is intended to support quantitative, clinician-supervised rehabilitation monitoring rather than autonomous diagnosis.