Physiologically constrained neuromuscular synergy extraction using a deep off-policy dynamic neuro-fuzzy system in wheelchair propulsion

Abstract

Manual wheelchair propulsion (MWP) is a repetitive activity that risks upper limb injuries, necessitating analysis of intramuscular coordination for effective intervention. The state-of-the-art synergy extraction methods struggle with MWP's nonlinear, dynamic nature and often overlook biomechanical constraints. This study introduces a deep reinforcement learning-based dynamic neuro-fuzzy (DRL-DNF) system that models complex electromyography (EMG) patterns and refines synergy structures in real time. Data from 24 manual wheelchair users included EMG signals from six muscles, joint kinematics, and handrim force. A musculoskeletal model was incorporated to account for joint dynamics and external forces, ensuring physiologically meaningful synergy extraction. Statistical analysis showed that DRL-DNF outperformed conventional methods, achieving a mean variance accounted for (VAF) of 94.12 ± 4.12 % in noise-free and 90.1 ± 4.18 % in noise-presence for three synergies, indicating strong robustness to noise and superior modeling of agonist-antagonist interactions. Significant differences in third synergy activations revealed autoencoder (AE) had higher values than independent component analysis (ICA) and non-negative matrix factorization (NMF), while ICA had lower values than NMF and DRL-DNF. Statistical parametric mapping indicated temporal differences, with ICA underactivating during the push and recovery phases, and AE overactivating in late recovery phase. Synergy coefficients also differed significantly across methods. AE consistently assigned lower weights to biarticular muscles, while NMF emphasized shoulder flexors and DRL-DNF provided a balanced representation. DRL-DNF identified physiologically relevant synergies, enhancing insights into intramuscular coordination strategies and showing potential for synergy-based control in assistive devices and rehabilitation. Future work will aim to enhance computational efficiency and expand dataset diversity for real-time application.