Curated Residual Decomposition for Increased MU Yield from HD-sEMG

Abstract

High-density surface electromyography (HD-sEMG) decomposition algorithms identify motor units (MUs) from non-invasive recordings, yet even gold-standard methods typically decompose only a subset of all physiologically active MUs and preferentially extract higher-amplitude units, leaving low-amplitude MUs underrepresented. We propose Curated Residual Decomposition (CRD): Through decomposition of the residual signal remaining from a manually edited initial decomposition, we identify additional, mainly lower amplitude MUs that were previously undetected due to action potential superposition and preferential convergence to larger MUs. We applied this approach to three datasets: concurrent HD-sEMG and HD-iEMG recordings from three muscles for two-source validation, and two publicly available datasets. Across all datasets, CRD substantially increased MU yield. In the two-source validation dataset, the first CRD iteration increased yield by 31-50% (FDI: 46.9%, TA: 49.8%, VL: 31.1%), with validation via intramuscular recordings confirming comparable identification accuracy of the CRD compared to the original MUs (RoA original: 0.994 {+/-} 0.005; RoA CRD: 0.992 {+/-} 0.009). In public datasets, yield increases ranged from 35-142% per muscle in the first iteration. Correspondingly, the explained signal power ratio increased, with the first iteration accounting for most gains. A second CRD iteration resulted in an additional, yet smaller yield increase (0-26.7%). Additionally identified MUs had smaller action potential amplitudes and lower recruitment thresholds, indicating the recovery of smaller, lower-threshold MUs. CRD reliably increased MU yield and explained signal power across diverse muscles, datasets, and decomposition implementations. The approach provides an optional, validated step for more complete decomposition results across the full recruitment range in offline analyses.