Proof-of-concept: Differentiating upper trapezius muscle with myofascial trigger point using deep learning model on a small prospective sEMG dataset.

Myofascial pain syndrome commonly arises from myofascial trigger points (MTrPs). While needle electromyography (iEMG) reveals spontaneous activity at trigger points, surface EMG (sEMG) has not been consistently validated. This study provides a proof-of-concept to investigate whether sEMG signals alone contain discriminative information for identifying myofascial trigger point. sEMG signals were recorded from healthy controls (n = 9) and patients with myofascial pain (n = 13). After preprocessing, the contraction segments from pre-intervention (N = 140) and post-intervention (N = 139) tests were extracted using an automated segmentation algorithm. A 1D convolution neural network (CNN) was trained with stratified five-fold cross-validation at the subject level. Binary cross-entropy, pairwise logistic, and their combined loss were compared, as well as the performance between pre-, post-, and combined pre/post-intervention signals. A latent-only sensitivity analysis was performed. Segment-level predictions were aggregated for subject-level analysis and evaluated with ROC-AUC and complementary metrics. Best subject-level performance was achieved using pairwise logistic loss on pre-intervention segments (AUC 0.940, 95% CI: 0.88-0.98; accuracy 0.96, 95% CI: 0.93-0.99). In latent-only analysis, combined pre- and post-intervention signals yielded the highest AUC (0.916, 95% CI: 0.86-0.96). This study suggests upper trapezius sEMG contain discriminative information for distinguishing primarily latent MTrPs subjects from healthy controls, motivating large-scale validation studies.