Correlating Data-Driven Muscle Selection Approaches to Synergies for Gait Prediction.

Optimizing sensors for physiological input is critical to enhance performance as well as minimize the cost and complexity of assistive devices (e.g. lower-limb exoskeletons). Electromyography (EMG) data can trace muscle activation for gait kinematics prediction. However, identifying optimal muscle groups for electrode placement and the potential variance between users has not yet been established. In this study, we use data-driven channel selection techniques on EMG signals to find muscle group combinations that maximize prediction performance. We apply greedy search (Recursive Feature Elimination, RFE) and variance-based (Principal Component Analysis, PCA) methods to select muscle groups during gait, without prior knowledge of musculoskeletal inter-connectivity. The selected muscle subsets are evaluated using the normalized accuracy of a Multi-Layer Perceptron (MLP), mapping muscle activity to knee flexion angle in a one-step-ahead scheme. The RFE selection led to an average predicted knee angle validation accuracy of % higher than the PCA approach, suggesting that dynamic search is more appropriate than a variance analysis of the signals. Whilst the RFE-selected muscle groups differed across subjects, the selected muscles were consistently spread out over more than 80% of the extracted synergy groups. This study underlines the value of incorporating synergistic information when developing gait prediction models, and reveals that maximizing the number of synergy groups could constitute the basis of muscle selection frameworks.