Unsupervised machine learning applied to decomposition of high-density myoelectric signals

Abstract

Decomposition of surface High-Density Electromyographic signals (HD EMG) into motor unitspike trains opens a non-invasive window to neural mechanisms underlying movement control. The signals produced by EMG decomposition have a scientific (to understand the physiology of motor system) and technological (to the control of human-machine interfaces, such as myoelectric prosthesis) relevance. Classical blind source separation algorithms have been used to decompose HD EMG; however, these algorithms cannot completely solve the decomposition problem of non-stationary signals with a low latency (online applications). Some recente algorithms try to disentangle these limitations by including a supervised deep learning algorithm in the data pipeline for online decomposition. In the present doctoral research project, the aim is to establish a mathematical-computational framework to the solution of decomposition problem of HD EMG signals. The framework will be based on unsupervised machine learning algorithms. Autoencoders and generative adversarial networks (and combination of these two classes of algorithms) will be evaluated as potential candidates tosolve the blind source separation problem applied to the decomposition of HD EMG. We expectto formulate an efficient algorithm to the solution of the problem defined above and, at the sametime, the algorithm would be included as part of a technological platform designed to providea biomimetic neural control of prosthetic and orthotic devices. (AU)