In recent years, research into optimization simulations and feedback training using musculoskeletal models has progressed to improve athletic performance. These models mathematically represent the human body’s structure and mechanics, enabling estimation of ideal muscle activation patterns and reduced joint loads. However, simulated “muscle activity” differs from measurable surface electromyograms (EMGs), which are noisy and nonlinear. This study aims to estimate model-based muscle activity from EMGs using a recurrent neural network (RNN) trained on time-series data of optimized muscle activity from musculoskeletal simulations. Due to difficulties in directly estimating accurate muscle activity from EMGs, joint angles and angular velocities were used as error functions. To enable error backpropagation, an ordinary differential equation (ODE) model was reproduced within the neural network. Training data included simulated pull-up movements with corresponding muscle activity, joint angles, and angular velocities optimized via a genetic algorithm (GA). Actual surface EMGs, joint angles, and angular velocities were measured during participants' pull-ups using electromyographs and motion sensors. As a result, an RNN was developed that estimates muscle activity from EMG signals during pull-ups. This enables more accurate and intuitive feedback on muscle output based on musculoskeletal model simulations, potentially improving motor learning and training outcomes..
SUZUMURA et al. (Wed,) studied this question.