Identification of spatial-temporal muscle synergies from EMG epochs of various durations: a time-warped tensor decomposition

Extraction of muscle synergies from electromyography (EMG) recordings relies on the analysis of multi-trial muscle activation data. To identify the underlying modular structure, dimensionality reduction algorithms are usually applied to the EMG signals. This process requires a rigid alignment of muscle activity across trials that is typically achieved by the normalization of the length of each trial. However, this time-normalization ignores important temporal variability that is present on single trials as result of neuromechanical processes or task demands. To overcome this limitation, we propose a novel method that simultaneously aligns muscle activity data and extracts spatial and temporal muscle synergies. This approach relies on an unsupervised learning algorithm that extends our previously developed space-by-time decomposition to incorporate the identification of linear time warps for individual trials. We apply the proposed method to high-dimensional spatiotemporal EMG data recorded during performance of whole-body reaching movements and show that it identifies meaningful spatial and temporal structure in muscle activity despite differences in trial lengths. We suggest that this algorithm is a useful tool to identify muscle synergies in a variety of natural self-paced motor behaviors.