Effect of SNR normalization on the estimation of muscle synergies from EMG datasets

Muscle Synergies Analysis (MSA) has gained significant attention in the last two decades as a method to derive motor control strategies and biomarkers of motor impairment and recovery. MSA is usually performed using the non-negative matrix factorization algorithm (NMF) and critically depends on the selection of the order of the model used in the analysis. The most common methodologies for the selection of the model order utilize arbitrary thresholds on the quality of reconstruction. However, the quality of the EMG data analyzed critically affects the model order selection. Here we present a simulative study on the effect of the SNR of the EMG in the selection of the model order for MSA using an objective criterion based on the Akaike Information Criterion (AIC). We show that low data quality leads to an overestimation of the muscle synergies. We then present a methodology for normalizing the SNR of the EMG before MSA. This methodology is based on the detection of the muscle activation intervals followed by a modification of the variance of the interburst noise. By applying this technique to a dataset collected during isometric reaching movements, we show that SNR normalization leads to a decrease in overall AIC values and in the optimal model order for synergies extraction.