Hand Gesture Recognition Based on Motor Unit Spike Trains Decoded from High-Density Electromyography

Chen,Yang Yu,Shihan Ma,Xinjun Sheng,Chuang Lin,Dario Farina,Xiangyang Zhu

DOI: https://doi.org/10.1016/j.bspc.2019.101637

IF: 5.1

2019-01-01

Biomedical Signal Processing and Control

Abstract:Objective: Methods for surface electromyographic (EMG) signal decomposition have been developed in the past decade, to extract neural information transferred from the spinal cord to muscles. Here, we characterize the accuracy in the identification of motor unit activities during hand postures from high-density EMG signals and we propose a mapping approach between these neural signals and hand gestures. Methods: High-density EMG signals were recorded during 11 hand gesture tasks from 11 able-bodied subjects. EMG signals were offline decomposed into motor unit spike trains (MUSTs) with a blind source separation algorithm. A gesture recognition approach based on motor unit classification was proposed. MUSTs were first pooled into groups corresponding to the 11 motions. Then the activation level of the neural drive to each motion was estimated as the summed discharge timings of MUSTs in each group. The output gesture class was determined by comparing the estimated activation level of each motion. Results: On average, 29 +/- 8 MUSTs were identified for each motion with an estimated decomposition accuracy >90%. The average classification accuracy for 11 hand gestures based on the proposed approach was >95% and outperformed the classic approach of using global EMG features. Conclusion and significance: These results indicate the possibility of identifying motor unit activities during intended motor tasks and demonstrate high classification accuracy of the hand gestures, with perspectives for human-machine interfacing. (C) 2019 Elsevier Ltd. All rights reserved.