Compensatory movements frequently occur during upper-limb rehabilitation for patients with stroke, potentially impeding effective motor recovery. Vision-based systems offer practical solutions for monitoring such compensations, but their application has often been limited by tracking inaccuracies and inadequate integration of trunk and arm movements. To address this limitation, we developed a real-time compensation detection and feedback system for upper-limb rehabilitation robot, which integrates upper-body reconstruction based on sensor-fusion with a machine learning classifier. The system estimates joint kinematics of both the trunk and arm to detect compensatory movement and delivers audio-visual feedback to reduce compensatory movement. We validated the system with 18 patients, who were divided into an experimental group and a control group, during robot-assisted planar reaching tasks. The classifier achieved reliable binary classification performance, with an F1-score 0.85 during classifier training and an F1-score 0.77 during real-time application with stroke. Importantly, the experimental group (n=9) significantly reduced compensatory movement occurrence from 58.9% at baseline to 38.2% post-training (p=0.016), showed improved trajectory mean distance from the theoretical path (p=0.0052), and exhibited decreased trunk movement magnitude (p<0.05). These results demonstrated effectiveness in reducing compensation and enhancing movement quality, highlighting its potential clinical utility for stroke rehabilitation.
Hwang et al. (Thu,) studied this question.