This study proposed a synergy-informed evaluation framework that integrates muscle synergy features derived from non-negative matrix factorization (NNMF) of surface EMG signals with deep feature fusion and machine learning techniques. EMG data were collected from nine lower-limb muscles during both anticipated and unanticipated landing tasks in 30 CAI patients and 30 healthy controls. Synergy-derived features were embedded into multidimensional representations and processed through a convolutional neural network (CNN) to achieve hierarchical feature integration, followed by classification using a random forest algorithm. The results revealed that individuals with CAI demonstrated altered synergy structures and a compensatory shift toward proximal muscle recruitment, particularly under unanticipated tasks. Among the four evaluated models, the CNN-RF-Unant configuration yielded the highest classification accuracy (0.96) and F1-score (0.95), outperforming models based on anticipated tasks or without CNN-based fusion. The feature importance analysis further highlighted the superior informativeness of unanticipated-task-derived features. These findings demonstrate that combining NNMF-based synergy analysis with CNN-driven fusion substantially enhances the detection of neuromuscular deficits in CAI, particularly when assessments involve unanticipated perturbations. This approach offers a clinically valuable and ecologically valid framework for identifying latent functional impairments and tailoring individualized rehabilitation strategies.
Zhou et al. (Thu,) studied this question.