A Quality-Guided Extension of Convolution Kernel Compensation for High-Density Surface EMG Decomposition

High-density surface electromyography (HD-sEMG) decomposition provides a non-invasive way to identify motor unit (MU) discharge patterns. However, the number of MUs extracted by blind source separation-based methods is often limited by repeated convergence to dominant sources, particularly when signal quality decreases or contraction level increases. In this study, we propose QE-CKC, a quality-guided extension framework built upon gradient convolution kernel compensation (gCKC), to recover more reliable MU discharge trains while maintaining decomposition accuracy. The proposed method expands candidate search, evaluates candidates using multiple quality measures, and removes redundant outputs through global ranking and spike-train-based deduplication. QE-CKC was evaluated using 150 simulated HD-sEMG trials covering five signal-to-noise ratio levels and three contraction levels, together with 66 real forearm HD-sEMG recordings from 11 participants. Across all simulation conditions, QE-CKC identified $14.06 \pm 4.79$ matched MUs per trial, compared with $11.72 \pm 3.18$ for gCKC and $6.51 \pm 6.76$ for KMCKC. F1 score and rate of agreement remained comparable to gCKC. In the real recordings, QE-CKC identified $16.17 \pm 10.59$ valid MUs per recording, compared with $13.29 \pm 8.23$ for gCKC and $8.53 \pm 9.39$ for KMCKC. After accounting for repeated measurements, QE-CKC identified 2.88 more valid MUs per recording than gCKC (95\% CI: 1.22--4.54, Holm-adjusted $p<0.001$). Among MUs independently reproduced from non-overlapping channel subsets, QE-CKC-only MUs showed discharge-train agreement comparable to that of MUs shared with gCKC. These results indicate that QE-CKC increases MU yield while maintaining comparable decomposition accuracy and reproducibility.