Decomposing single-channel intramuscular electromyography signal sampled at a low frequency into its motor unit action potential trains with a generative adversarial network

Abstract

Conventional methods decompose single-channel intramuscular electromyography (iEMG) signals into their constituent motor unit action potential trains (MUAPTs) by detecting and clustering individual motor unit action potentials (MUAPs). However, these methods are not applicable for iEMG signals recorded by electrodes with a large sensing areas or iEMG signals sampled at a low frequency, in which detecting and clustering individual MUAPs are difficult due to superimpositions of the MUAPs and the loss of MUAP morphological characteristics. In this study, we propose an approach based on a generative adversarial network to decompose iEMG signals, which does not depend on detecting and clustering individual MUAPs from the iEMG signal. The proposed approach decomposes the iEMG signal into its MUAPTs based on Bayes’ law and a Wasserstein generative adversarial network with gradient penalty (WGAN-GP). MUAPTs generated by the WGAN-GP were used to decompose the iEMG signal to maximize the posterior probability of the generated MUAPTs given the iEMG signal. The accuracy of the proposed approach is analysed directly by decomposing the simulated iEMG signal with seven gold-standard motor units. The results showed that the proposed approach achieved a 53% accuracy in capturing the firing regularities of the MUs, while the conventional method achieved a 37% accuracy on the same task.