Comparison Evaluation of Machine Learning Regression Models for EMG-Based Hand Grip Prediction across Multiple MVC Levels

Abstract

Electromyography–based force prediction provides an intuitive control strategy for assistive and rehabilitation hand systems. This study investigates an EMG-based hand grip force prediction framework using machine learning techniques by modeling the relationship between forearm muscle activation and grip force at varying contraction levels. sEMG signals were obtained from the FDS and FCR muscles of ten healthy female participants (aged 20–25 years) during controlled grip tasks performed at five MVC levels ranging from 20% to 100%. The recorded signals were filtered, processed using RMS feature extraction, and normalized to MVC prior to regression modeling. LR, GPR, SVR, and kNN models were evaluated using offline analysis. Performance was evaluated using RMSE and prediction accuracy, defined relative to the measured grip force. The results indicate that GPR demonstrates the most consistent performance, achieving the highest average accuracy (85.84%) and the lowest average RMSE (10.54). In contrast, SVR and kNN exhibit higher prediction errors, particularly at higher MVC levels.