Decoding muscle force from individual motor unit activities using a twitch force model and hybrid neural networks

• Published in: Biomedical signal processing and control Vol. 72; p. 103297
• Main Authors: Li, Xinhui, Zhang, Xu, Tang, Xiao, Chen, Maoqi, Chen, Xiang, Chen, Xun, Liu, Aiping
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2022
• Subjects: Deep learning; Motor unit; Muscle force; Neural networks; Surface EMG decomposition; Deep learning; Surface EMG decomposition; Neural networks; Muscle force; Motor unit
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.103297

•A muscle force estimation method using microscopic neural drives is proposed.•The method can differentiate individual motor unit activities for decoding force.•A focus on the advance of mining spatial information from motor unit activities.•Investigation of hybrid neural networks incorporated with a physiological model. A deep learning method involving hybrid neural networks incorporated with a physiological twitch force model is presented for decoding muscle force from individual motor unit (MU) activities. A series of MU action potential (MUAP) trains were obtained using a progressive FastICA peel-off algorithm through high-density surface electromyogram (HD-sEMG) decomposition. The twitch force model was adopted to transform MUAP trains into twitch force sequences. The hybrid convolutional neural networks (CNNs) consisting of a spatial attention model (SA-CNN) and a bidirectional gated recurrent unit (Bi-GRU) were fed by the twitch force sequences to predict the muscle force. The performance of the proposed method was evaluated in predicting the thumb abduction force using both simulated data and experimental HD-sEMG recordings from abductor pollicis brevis muscles of ten subjects. The simulation approach assisted our method design in confirming a simplified relationship between MUAP amplitude and force contribution of each MU. When processing experimental data, the root mean square deviation (RMSD) values between the predicted and the actual forces were 6.57% ± 1.26% and 7.18% ± 0.96% under the user-specific testing scheme and the user-independent testing scheme, respectively. The proposed method outperformed other common sEMG amplitude-based or MU-based methods with significantly decreased RMSD values (p < 0.001). These findings indicate the advance of mining spatial information from individual MU activities using deep neural networks in decoding the muscle force. This study provides a useful tool for muscle force estimation, with wide applications in biomechanics and sports rehabilitation.