Vision-based Automatic Detection of Compensatory Postures of after-Stroke Patients During Upper-extremity Robot-assisted Rehabilitation: A Pilot Study in Reaching Movement

Abstract

Objectives: Compensations are commonly employed by stroke patients during robot-assisted rehabilitation, leading to suboptimal recovery outcomes. This study investigated the feasibility of vision-based detection on compensatory postures in stroke patients by using kinematic data and machine learning algorithms.Methods: Ten stroke survivors did a scripted reaching movement using the affected side with an end-effector robot assistance while three dimensional(3D) trajectories of upper body joint positions were recorded based on a vision-based tracking system.Algorithms: A set of discriminating features which correlated with various compensatory postures were extracted based on the kinematic data. Two multi-label classifiers were trained to identify and categorize the postures of the participants.Results: The most obvious compensatory postures in the scripted reaching movement are forward trunk displacement and trunk rotation, followed by shoulder elevation, and less insufficient elbow extension. The two multi-label classifiers trained achieved a fair compensation detection performance: Multi-Label k-Nearest Neighbor (ML-KNN) classifier (Hamming loss: 0.16, Macro-average: 0.642, Micro-average: 0.85), Multi-Label Decision Tree (ML-DT) classifier (Hamming loss: 0.15, Macro-average: 0.654, Micro-average:0.85).Conclusions: It is feasible to realize automatic detection of compensatory postures during the upper-limb robot-assisted rehabilitation based on the low-cost vision system. It is expected to complement real time compensatory postures recognition of clinical faculty and be integrated into a rehabilitation robot system to reduce compensations in stroke patients.