The effect of Non-Negative Matrix Factorization initialization on the accurate identification of muscle synergies with correlated activation signals

The quantification of muscle coordination through muscle synergy analysis has been shown to represent a useful tool for inferring motor control strategies and extracting indexes of motor impairment and recovery. The main goal of this work has been the assessment of the performance of different initializations techniques for Nonnegative Matrix Factorization (NNMF) for the extraction of muscle synergies. Previous research has shown that NNMF performance might be affected by different kinds of initialization. The present study aims at optimizing the traditional NNMF initialization for decomposing EMG data with temporal dependencies, typical of pathological conditions such as stroke. For this purpose, three different initializations have been implemented: Random, SVD-based, and sparse. Traditional NNMF update rules have been used to identify muscle synergies from experimental EMG signals recorded during pedaling from 11 subjects. Synthetic data have been generated from real EMG data, whose activation coefficients have been corrupted by simulating different degrees of correlation. By measuring the quality of identification of the original synergies underlying the data it has been possible to compare the performance of different initialization techniques. Simulation results demonstrate that sparse initialization performs significantly better than all the other kinds of initialization in accurately estimating muscle synergies when the activation coefficients are characterized by high levels of correlation.