Machine Learning Methods to Predict Unmeasured Muscle Activation in Upper Limb Reaching Task for Assessing Population-Level Cross-Subject Generalizability.

Abstract

Accurately estimating muscle activation remains a fundamental challenge in neuromuscular modeling, particularly when measurements are limited by sensor placement, noise, and accessibility constraints. Traditional methods, such as Inverse Dynamics and Static Optimization, require high-quality kinematic and kinetic data. However, interpreting muscle activation directly from surface electromyography (sEMG) signals can be challenging due to their complex and non-linear characteristics. In this study, we propose a deep learning-based framework to investigate the feasibility of learning generalizable inter-muscular activation relationships using sEMG alone. To establish a controlled validation framework, measured muscle activations are treated as prediction targets during a standardized forward-reaching task performed by 30 participants. We evaluated three model architectures: Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and hybrid CNN-LSTM models, to determine their effectiveness in capturing both spatial and temporal dynamics in EMG data. We systematically varied the number of unmeasured output muscles, assessed the impact of training dataset size, and conducted Leave-One-Muscle-Out (LOMO) analysis to evaluate the importance of input muscles. The CNN-LSTM model outperformed the standalone CNN and LSTM models, particularly in multi-muscle prediction tasks, achieving the highest accuracy (RMSE=0.103 & r=0.866) for the triceps long head muscle. These results demonstrate the feasibility and benefits of using deep learning for scalable, non-invasive and simultaneous prediction of muscle activation patterns for superficial as well as provide a foundation for future proxy prediction of muscles that are difficult to measure directly.