An enhanced EMG-driven Musculoskeletal model based on non-negative matrix factorization

Abstract

In recent years, electromyography (EMG) signals have been widely used in human–machine interfaces (HMIs). Musculoskeletal models (MMs) have been employed to decode human movement intentions from EMG signals in HMIs. Non-negative matrix factorization (NMF) has been adopted to extract neural control information from surface EMG signals. The aim of this study was to investigate if the NMF could extract the control information from surface EMG signals to substitute muscle activations of the deep muscles of the MM. An enhanced MM based on NMF (NMF-MM) was proposed to predict wrist pronation/supination (Pro/Sup) joint angles from surface EMG signals. We extracted the muscle activations of a pair of virtual agonist–antagonistic muscles with NMF from multi-channel EMG signals and inputted the extracted signals into the MM. Eight able-bodied subjects were recruited and tested in four different upper limb postures. Eight bipolar electrodes were attached to the upper forearm to record EMG signals. The Pearson’s correlation coefficient (r) and the normalized root mean square error (NRMSE) between measured and predicted joint angles were adopted to quantify the performance of the decoding methods. The performance of the proposed NMF-MM was compared with that of the NMF, linear regression (LR), artificial neural network (ANN), LR based on NMF (NMF-LR), ANN based on NMF (NMF-ANN), LR based on time-domain features (TD-LR), and ANN based on time-domain features (TD-ANN). The results demonstrated that the NMF-MM outperformed the other methods with higher prediction accuracy and better robustness. This study proposed a promising method for predicting continuous and coordinated movements in HMIs.