An Approach to Extract Nonlinear Muscle Synergies from sEMG through Multi-Model Learning

How does the Central Nervous System (CNS) controls a group of muscles is an important question in the field of motor control. A common conception is developed over the years that the CNS make use of predefined activation patterns, known as muscle synergies during task execution. These muscle synergies are extracted by applying any of the factorization algorithms such as Non-Negative Matrix Factorization (NNMF), Independent Component Analysis (ICA) or Principle Component Analysis (PCA) on a concatenated surface EMG data set recorded from the target muscles. However, the step to concatenate sEMG signals before they are given as input to these linear algorithm is crucial as the synergistic structure changes significantly based on the number and choice of muscles considered during concatenation step. To address this problem, we propose a new approach of extracting muscle synergies by treating sEMG signals from each muscle as an individual modality and then learning the synergistic structure among them if it exists using multi-view learning. In this study, we propose to use Manifold Relevance Determination (MRD) to find nonlinear synergies from sEMG by assuming the sEMG of a muscle as an individual modality. Results have shown that synergistic patterns extracted using our approach are consistent upon addition of sEMG signals from new muscles.