Continuous Prediction of Human Joint Mechanics Using EMG Signals: A Review of Model-Based and Model-Free Approaches

Abstract

This paper reviews model-based and model-free approaches for continuous prediction of human joint motion using surface electromyography (EMG) signals. The review focuses on approaches that solve the regression problem of continuously decoding motion from EMG as opposed to classification or discrete gesture identification methods. Model-based approaches use physics-based models to capture the dynamics of the muscles; and musculoskeletal geometry to estimate muscle forces and joint torques. Model-free machine learning and deep learning architectures employ training to learn the statistical regularities in the data to make predictions. Model-free approaches work well for real-time applications with better scalability for multiple muscle inputs and lower computational costs while model-based approaches are insightful and capture the essence of underlying muscle activation and contraction dynamics. Integrating model-free techniques to complement parametric models is a good practice for improving generalization of EMG-based predictions. The strengths and limitations of the two paradigms are compared based on scalability, performance, computational efficiency, and robustness for applications in robotics and biomechanics. We expand on previous review works with recent studies to draw inferences about the knowledge gaps and opportunities for future research. We hope this review serves as a comprehensive guide for current and prospective researchers interested in joint mechanics prediction using EMG.