An integrated wireless sEMG framework with Bayesian-optimized machine learning for multi-task locomotor classification and neuromuscular characterization

Wearable surface electromyography (sEMG) systems show strong potential for real-time gait analysis, yet few frameworks combine wireless acquisition, standardized normalization, and optimized machine learning within a single validated pipeline, leaving their joint contribution to subject-independent locomotor recognition unclear. This study developed and validated a custom wireless sEMG framework combining hardware design, normalization evaluation, and Bayesian-optimized machine learning for locomotor-task classification and neuromuscular characterization. An ESP32-based system recorded tibialis anterior, soleus, and gastrocnemius lateralis activity in 25 healthy participants during level walking, stair ascent, and ramp ascent. Signals were bandpass filtered (20–500 Hz), rectified, and used to extract 120 time-, frequency-, and time–frequency-domain features. Three normalization strategies, three feature-selection methods, and eight Bayesian-optimized classifiers were evaluated via an exploratory participant-level split, nested 5 × 5 cross-validation, leave-one-subject-out (LOSO) validation, and an independent test set. Task-dependent EMG differences were significant with large effect sizes (η 2 = 0.183–0.708, p < .001). Level walking was the most electromyographically distinct task, while stair and ramp ascent were similar yet separable; ramp ascent showed multi-muscle co-activation at mid-stance and peak soleus activity at toe-off (50.7 ± 6.2% MVC). Genetic Algorithm feature selection with Z-score normalization gave the best single-split performance (logistic regression: 98.33% accuracy; SVM: 96.33%). Under the stricter LOSO regime, random forest reached 89.02% accuracy with 44 features, while a correlation-pruned 17-feature subset matched or exceeded this across all classifiers—best with SVM (92.40%)—achieving an 85.8% dimensionality reduction from the original feature space (61% relative to the 44-feature set) with no loss in mean accuracy. Two deep-learning baselines (1D-CNN, BiLSTM) achieved comparable LOSO accuracy (91.39% and 87.72%), confirming the feature-based pipeline captures the signal’s discriminative structure without end-to-end deep architectures. These findings support an interpretable, compact, subject-independent framework for multi-task locomotor classification, though broader clinical validation is still needed.