Muscle synergy-driven ensemble learning framework for individualized stroke gait rehabilitation.

This study proposes a novel ensemble machine learning (ML) framework integrating neurophysiological principles from muscle synergy analysis to support clinical decisions in stroke gait rehabilitation. The framework leverages spatial and temporal features of muscle synergies using a hierarchical ensemble model to improve classification performance and interpretability. Muscle synergies were extracted using non-negative matrix factorization from surface EMG recordings of 380 participants, comprising 120 healthy and 260 post-stroke individuals, each contributing one leg. Feature vectors were derived through a bidirectional decomposition process, wherein either the spatial (W) or temporal (H) synergy matrix was fixed to normative patterns obtained from healthy controls. This enabled the extraction of patient-specific deviations in coordination and activation timing, serving as interpretable indicators of stroke-related neuromuscular impairment. Separate classifiers trained on each feature domain were integrated via meta-regression, achieving classification accuracies above 98% across all configurations on the internal test dataset. After training, performance-weighted feature importance values from tree-based models validated the clinical relevance of learned classification criteria. SHAP (SHapley Additive exPlanations) values quantified sample-specific feature contributions on the test dataset, ensuring individual-level interpretability of predictions. The proposed framework bridges stroke-related neuromuscular impairments and clinical insights, laying a foundation for integrating neurophysiologically grounded ML models into rehabilitation.