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 major challenge in neuromuscular modeling, particularly when direct measurements are limited by sensor placement, noise, and accessibility constraints. Traditional approaches such as inverse dynamics and static optimization require high-quality kinematic and kinetic data, while surface electromyography (sEMG) signals are often complex, nonlinear, and difficult to interpret directly. This study evaluated whether deep learning models can learn generalizable intermuscular activation relationships using sEMG alone. Measured muscle activations were treated as prediction targets during a standardized forward-reaching task performed by 30 participants. Three architectures, convolutional neural networks (CNN), long short-term memory (LSTM) networks, and hybrid CNN-LSTM models, were compared across different numbers of predicted muscles, training dataset sizes, and leave-one-muscle-out conditions. No single architecture consistently outperformed the others across all target muscles and prediction configurations. CNN-LSTM achieved favorable performance in selected cases, whereas CNN and LSTM produced comparable or better results in others. Compared with a population-mean baseline, the machine learning (ML) models did not consistently produce lower RMSE or higher correlation, suggesting that the standardized reaching task contained shared population-level activation structure captured reasonably well by a simple average profile. Prediction performance depended strongly on target muscle, prediction configuration, and subject-level variability, with the biceps showing consistently lower predictability. Increasing the number of training participants improved performance, but gains plateaued beyond 20 subjects. Overall, these results support the feasibility of sEMG-based deep learning for simultaneous muscle activation prediction, while indicating that its advantage over simpler baselines is configuration-dependent rather than systematic.