Simultaneous classification of hand gestures and elbow angles using sEMG signals: a comparison of machine learning performances

This study evaluated the preliminary feasibility of simultaneously classifying hand gestures and corresponding elbow angles using only surface electromyography (sEMG) signals and lightweight machine learning models. Conventional human-machine interface (HMI) systems often separate gesture recognition from angle estimation or require additional sensors, such as inertial measurement units, thereby increasing hardware complexity and cost. sEMG signals were collected from eight participants using a five-channel configuration while performing four hand gestures at two discrete elbow-angle conditions, yielding a total of eight classes under controlled experimental settings. The signals were preprocessed, segmented using a sliding-window approach, and converted into five standard time-domain features, which were used to train support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) classifiers. Among the evaluated models, the RF classifier achieved the highest mean accuracy, reaching 96.72% (±0.58%) across five-fold cross-validation. Misclassification analysis showed that most errors occurred between elbow-angle conditions within the same gesture rather than between different gestures. These findings demonstrate the preliminary feasibility of simultaneously classifying hand gestures and elbow-angle conditions using only sEMG signals and lightweight machine learning models under controlled experimental conditions. This sensor-efficient framework may support the future development of simplified multidimensional motion-intent recognition systems for applications in prosthetic control, rehabilitation robotics, and other HMI systems.