Machine learning-based decomposition of the activity of individual motor units from synthetic and experimental data

Within this project, we aim to use synthetic data obtained from simulations to develop novel machine learning-based decomposition algorithms for identifying the activity of individual motor units in skeletal muscles. Compared to existing methods, we expect significant computational speed-ups and decomposition algorithms that identify with much better accuracy more motor units. The results of applying our newly developed algorithms to experimentally measured electromyographic (EMG) recordings will improve the overall understanding of the control of the neuromuscular system. These novel machine learning-based decomposition algorithms will be achieved by developing a 3D multi-domain model to simulate the activity of selected motor units during iso- and non-isometric contractions in muscles of arbitrary shape. The use of a conditional generative adversarial network (cGAN) will provide a flexible and powerful framework for EMG to motor unit activity translation, even in a nonlinear environment. As in silico experiments firing times of individual neurons are no longer unknown, computing the resulting EMG signal provides a basis for designing, training and validating novel machine learning-based decomposition methods. Since the exact motor unit distribution is not known, new measures for comparing synthetic and actual EMG data will need to be developed. A further aim of this proposal is to use these findings to extend the algorithms in such a way that it also can decompose motor units during non-isometric contractions – something that cannot properly be decomposed with existing methods yet. One path to achieve this is to utilise (continuum-mechanical) models to predict motion, and, hence, deformation and the shift of the motor units during contraction. As such, this project directly links to the vision of a digital human model.