Electromyography-based assistive and rehabilitation devices have shown potential for restoring mobility, especially for post-stroke patients. However, the variability of biological signals and the processing delays caused by signal acquisition and feature extraction influence myoelectric control systems’ real-time functionality and robustness. This study evaluates the classification performance of electromyographic (EMG) signals for six distinct hand gestures in healthy individuals and post-stroke patients. Different feature extraction methods and machine learning algorithms are employed to analyze the impact of acquisition time (0.5–4 s) and the number of channels (1–4) on model accuracy, robustness, and generalization. The best results are obtained using power spectral density and dimensionality reduction, reaching a classification accuracy of 94.79% with a 2 s signal and 95.31% for 4 s. Acquisition time has a greater effect on accuracy than the number of channels used with accuracy stabilizing at 2 s. We test for generalization using post-stroke patient data, evaluating two scenarios: intra-patient validation with 90% accuracy and cross-patient validation with 35–40% accuracy. This study contributes to developing effective real-time myoelectric control systems for neurorehabilitation.

Electromyography-based assistive and rehabilitation devices have shown potential for restoring mobility, especially for post-stroke patients. However, the variability of biological signals and the processing delays caused by signal acquisition and feature extraction influence myoelectric control systems’ real-time functionality and robustness. This study evaluates the classification performance of electromyographic (EMG) signals for six distinct hand gestures in healthy individuals and post-stroke patients. Different feature extraction methods and machine learning algorithms are employed to analyze the impact of acquisition time (0.5–4 s) and the number of channels (1–4) on model accuracy, robustness, and generalization. The best results are obtained using power spectral density and dimensionality reduction, reaching a classification accuracy of 94.79% with a 2 s signal and 95.31% for 4 s. Acquisition time has a greater effect on accuracy than the number of channels used, with accuracy stabilizing at 2 s. We test for generalization using post-stroke patient data, revealing a substantial accuracy drop from 90% (intra-patient validation) to 35–40% (cross-patient validation). This study paves the way for developing effective real-time myoelectric control systems for neurorehabilitation.