Extracting muscle synergy patterns from EMG data using autoencoders

Abstract

Muscle synergies can be seen as fundamental building blocks of motor control. Extracting muscle synergies from EMG data is a widely used method in motor related research. Due to the linear nature of the methods commonly used for extracting muscle synergies, those methods fail to represent agonist-antagonist muscle relationships in the extracted synergies. In this paper, we propose to use a special type of neural networks, called autoencoders, for extracting muscle synergies. Using simulated data and real EMG data, we show that autoencoders, contrary to commonly used methods, allow to capture agonist-antagonist muscle relationships, and that the autoencoder models have a significantly better fit to the data than others methods.