Machine learning–based classification of pathological shoulder motion using phase-specific kinematic features

This study investigated an upper-extremity exoskeleton for machine learning-based discrimination of orthopedic shoulder pathology and identification of discriminative temporal features. Twelve patients with shoulder impairments and thirty healthy controls performed eight standardized tasks. Logistic regression with stratified 5-fold cross-validation was used for classification. Temporal effect sizes were computed using pointwise Cohen's d, and permutation-based phase ablation quantified the contribution of movement phases to AUROC. Classification performance ranged from 0.70 to 1.00, with six tasks achieving AUROC ≥ 0.90. Mid-cycle phases dominated in flexion and abduction tasks, whereas early and late phases were most informative for rotational movements, supporting interpretable, phase-aware ML models.