Emg-Based Force Estimation Using Deep Learning Models

Abstract

The estimation of force through electromyography (EMG) assumes paramount importance in diverse domains, including neurorehabilitation, myoelectric control, and neurofeedback systems. The intricate relationship between muscle contraction and force, characterized by linear associations in small muscles with narrow motor units and nonlinear relationships in larger muscles with wider motor units, underscores the complexity of this physiological interplay. Against the backdrop of a global demand for advanced technologies to address limb loss limitations, with an estimated 100 million individuals worldwide in need of prosthetics, there arises an urgent need for sophisticated solutions. Meeting the diverse needs of prosthetic users emphasizes the crucial role of EMG-based force prediction, striving to provide adaptive and personalized solutions for an inclusive and effective approach to limb rehabilitation. This comprehensive study explores the dynamic interplay between surface electromyography (sEMG) and intramuscular electromyography (iEMG) signals and force estimation. Leveraging a diverse set of machine learning and deep learning models, the research aims to predict forces in both healthy individuals and those with trans-radial amputations. Across sEMG and iEMG modalities, deep learning models, including Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and the hybrid LSTM-TCN, consistently exhibit remarkable efficacy. These models, boasting R² values surpassing 0.90 in force prediction, offer promising advancements in refining force estimation through electromyography. Notably, the TCN emerges as an exemplary performer, yielding R² values of 0.98 for able-bodied individuals and 0.87 for amputees in sEMG. Simultaneously, the hybrid TCN-LSTM model maintains strong performance with R² values of 0.98 for able-bodied individuals and 0.85 for amputees in sEMG. The LSTM model also upholds notable performance, showcasing R² values of 0.99 for able-bodied individuals and 0.80 for amputees in sEMG. Beyond unraveling the intricacies of EMG-based force estimation, this study sheds light on the unique challenges posed by amputations, contributing substantively to the ongoing quest for enhanced precision and effectiveness in rehabilitation interventions.