Recursive Feature Elimination With Cross-Validation for Experimental Surface Electromyography Signal Decoding

Surface electromyography (sEMG) plays a crucial role in decoding neuromuscular activity, with applications in prosthetic control and human–computer interaction. While extensive research has focused on hand gesture classification, discriminating identical hand movements performed at varying wrist angles remains underexplored. This study evaluates five feature reduction techniques, namely, principal component analysis (PCA), linear discriminant analysis (LDA), chi-square test, mutual information (MI), and recursive feature elimination with cross-validation (RFECV) for classifying five wrist-based hand movements. Two independent sEMG datasets were analyzed. Experimentally collected Dataset 1 comprised 17 subjects performing 5 repetitions of hand movements at wrist angles (+45°, +90°, 0°, −45°, and −90°) using the Delsys Trigno electromyography system. Dataset 2 (Ninapro DB1, Exercise B, Movements 13–17) comprised 27 subjects performing 10 repetitions of 5 distinct wrist and hand movements. Gradient boosting, LightGBM, random forest, SVM-RBF, and a 1D-CNN classifier were evaluated under different validation strategies, including subject-wise hold-out, subject-wise cross-validation, and movement-level cross-validation. RFECV consistently delivered strong performance across both datasets. Mean accuracies ranged from 77.73% under movement-level cross-validation to 90.11% under subject-wise hold-out, with best subject-wise accuracies reaching 95.38% on Dataset 1 and 96.35% on Dataset 2. In contrast, the 1D-CNN on raw EMG windows yielded only 36.64% mean accuracy, confirming the importance of feature engineering. These findings demonstrate that RFECV is a reliable and generalizable feature selection technique for sEMG-based movement classification, particularly for challenging tasks involving angular variations of identical hand movements.