Hand Gesture Recognition from Myoelectric Signals in Individuals with Incomplete Spinal Cord Injuries Using Machine Learning and Automatic Feature Selection

Spinal cord injuries (SCIs) are a major cause of upper limb impairment. In individuals with incomplete SCI, limitations of traditional myoelectric control have motivated artificial intelligence-based pattern recognition to better detect movement intent. This study aimed to develop and evaluate a system for the acquisition and classification of 8 hand gestures from myoelectric signals (MESs) in three participants with incomplete SCI. The 4-channel system captured myoelectric signals using individual sensors, from which 22 time-domain and frequency-domain features were extracted. Three feature selection techniques were applied: ReliefF, sequential feature selection (SFS), and genetic algorithms (GA). Eight machine learning models were trained using cross-validation: random forest (RF), support vector machines (SVM), multilayer perceptron (MLP), k-nearest neighbors (KNN), linear discriminant analysis (LDA), logistic regression (LR), decision trees (DT), and Gaussian Naïve Bayes (GNB). Statistical analysis confirmed RF as the best-performing classifier (p < 0.05), while KNN and MLP showed equivalent performance across subjects. Offline accuracies ranged from 71.78% to 81.60%. Real-time evaluation of RF-SFS yielded 76.67% accuracy with a mean end-to-end processing time of 330.42 ± 6.65 ms. Given the small and heterogeneous cohort, these findings should be interpreted as preliminary, participant-specific evidence of the feasibility of myoelectric hand gesture classification in individuals with incomplete SCI rather than as population-level estimates of performance.