Machine Learning Methods for sEMG-based Motion Intention Classification in Upper-Limb Exoskeletons

Abstract

Surface electromyography (sEMG) provides a crucial non-invasive method for early assessment of motor intent, essential for the control of upper-limb exoskeletons operating under stringent real-time constraints. This review compares three prominent classifiers—support vector machine (SVM), linear discriminant analysis (LDA), and random forest (RF)—in the context of sEMG-based motion intention recognition. Employing a standardized methodological framework, this paper analyze performance metrics such as accuracy, latency, robustness, and computational efficiency, while addressing challenges like non-stationarity, inter-subject variability, and class imbalance. Results indicate that LDA is suitable for fast inference in control loops at 50–100 Hz, while SVM offers high accuracy for complex decision boundaries at a greater computational cost. RF stands out for its robustness to noise and cross-user variance, making it viable for real-time applications, albeit with trade-offs in model complexity and interpretability. Overall, LDA and compact RF models present practical options for embedded systems, whereas SVM is preferable for scenarios prioritizing peak accuracy. Future research should focus on compact feature fusion, domain adaptation, and hybrid learning models to enhance applicability in real-world settings.