Machine learning analysis of surface electromyography for quantifying neuromuscular compensatory profiles and...

Neck pain presents significant clinical heterogeneity, while current assessment methods lack systematic approaches for quantifying muscle function. This study established a multi-dimensional surface electromyography (sEMG) framework to quantify the predictive value of muscle function features for clinical outcomes and identify neuromuscular compensatory states.

Eighty-nine participants were enrolled, including 69 chronic neck pain patients and 20 age-, sex-, and BMI-matched healthy volunteers. Surface EMG signals from six bilateral cervical muscles were recorded during functional postural control tasks in four directions. Following preprocessing and quality control, seven features spanning coordination (co-contraction index [CCI], deep-to-superficial ratio [DSR], asymmetry index [ASI]), control (sample entropy [SampEn], coefficient of variation trend [CoV_trend]), and fatigue (initial fatigue rate [IFR], median frequency slope [MDF_Slope]) dimensions were extracted. Six regression models (multiple linear regression, Ridge, Lasso, ElasticNet, Random Forest, and XGBoost) were compared using Repeated K-Fold cross-validation (5 folds × 10 repeats), with SHapley Additive exPlanations (SHAP) quantifying feature contributions. K-means clustering with multi-index validation identified distinct neuromuscular states.

Data quality reached 95.3% (263/276 segments). Six of seven EMG features significantly discriminated between healthy controls and patients (p < 0.05), with five showing large effect sizes (|d| > 0.8). Comparative model evaluation revealed XGBoost as the superior predictor for NDI scores (CV R ² = 0.449 ± 0.182, MAE = 4.66 ± 0.74), outperforming all linear models by 13%-16%; a permutation test confirmed statistical significance (p = 0.001). SHAP analysis identified CCI (mean |SHAP|: 3.68), CoV_trend (1.74), and MDF_Slope (1.56) as primary NDI predictors. Four features exhibited significant severity gradients: CCI progressively increased from mild (0.57 ± 0.06) to severe groups (0.73 ± 0.10, p < 0.001); DSR was significantly higher in mild versus moderate patients (1.73 vs. 1.37, p = 0.033); CoV_trend decreased with severity (p = 0.013); MDF_Slope showed progressive changes (p < 0.001). Multi-index clustering validation (Silhouette, Calinski-Harabasz Index, Davies-Bouldin Index, Gap Statistic) supported K = 2 as the optimal cluster number. Two candidate neuromuscular states were identified: Asymmetric Compensation (63.8%, high ASI: 26.7%, preserved DSR: 1.744, low CCI: 0.635) and Global Stiffening (36.2%, high CCI: 0.771, low DSR: 1.077, steep MDF_Slope: -0.408), consistent with established models of pain-motor reorganization. Cluster membership was associated with disease severity (χ² = 6.06, p = 0.048). EMG-NDI associations were independent of demographic confounders (partial correlations virtually identical to zero-order values after controlling for age, sex, and BMI; maximum change <0.1%).

This interpretable multi-dimensional sEMG framework quantifies muscle function features' predictive value for functional disability. Severity-related neuromuscular changes may enable disease progression monitoring, while the identification of candidate compensatory states provides preliminary stratification criteria for personalized rehabilitation pending prospective validation.