PCA-Guided Separation of Mixed Motor Unit Sources in High-Density EMG

Objective: Decomposition of high-density electromyographic signals enables non-invasive analysis of individual motor unit (MU) behavior, but reliable interpretation of physiological changes in health and disease depends on accurate MU discharge detection. This accuracy is compromised by mixed source estimates, where high amplitude peaks are associated with discharges from more than one MU. We introduce a post-decomposition framework to identify and separate suspected mixed sources using PCA-guided source refinement. Method: For each suspected mixed source, extended and whitened EMG vectors were extracted at source peaks and projected into a low-dimensional PCA subspace. This subspace highlighted MU-specific differences across candidate discharges, including subtle or spatially localized features of the spatiotemporal MUAP profile. Clusters in the PCA subspace were used to initialize source estimates for the constituent MUs. During iterative source refinement, source peak amplitudes were reweighted according to the distance of their corresponding points from the associated cluster center. Particle swarm optimization selected the reweighting factor that minimized the coefficient of variation of inter-spike intervals (CoVISI). Results: The algorithm separated mixed MU sources in simulated and experimental HDsEMG data. In simulated data, resolving mixed sources increased median rate of agreement (RoA) by >40%. In experimental recordings, MU yield increased by 1.27 per trial and CoVISI decreased by 0.28 (33% RoA improvement). Conclusions: PCA-based representation enhanced separability between MUs with similar MUAP profiles, while distance-based amplitude reweighting reduced re-merging during source refinement. Significance: This framework resolves merged MU discharge trains, improving decomposition accuracy and recovering MUs that might otherwise be excluded by quality thresholds.