Abstract
Stroke often leads to persistent upper limb dysfunction that impairs activities of daily living, yet objective biomechanical indicators for precise assessment remain limited. This study aimed to characterize phase-specific impairments in energy output, torque stability, and muscle coordination during the hand-to-mouth (HTM) task and to explore their potential for improving rehabilitation evaluation. Motion data from 20 stroke patients and 20 healthy controls were recorded using wearable surface electromyography and inertial measurement unit systems. A musculoskeletal model was applied to calculate joint torque, mechanical work, torque smoothness, and a novel torque-based co-contraction index across four movement subphases. These phase-specific metrics demonstrated significant correlations with clinical motor impairment scores, confirming their clinical validity. Significant dynamic features were then selected to construct machine learning models for group classification. Stroke patients showed reduced output capacity, increased torque fluctuations, and abnormal co-contraction patterns that varied across subphases. Among the classifiers, the quadratic support vector machine achieved the best performance, with an accuracy of 84.6% and an AUC of 0.853, surpassing models based on whole-task features. These findings demonstrate that phase-specific biomechanical features sensitively capture neuromuscular deficits in stroke survivors and highlight the potential of phase-specific biomechanics to inform future individualized rehabilitation assessment and treatment planning.