A multi-resolution physics-informed recurrent neural network: formulation and application to musculoskeletal systems.

Area of Science:

• Biomechanics
• Computational Neuroscience
• Machine Learning

Background:

• Musculoskeletal (MSK) motion prediction from surface electromyography (sEMG) is complex due to differing signal frequencies.
• Accurate mapping requires advanced computational models to capture muscle dynamics and joint motion.

Purpose of the Study:

• To develop a multi-resolution physics-informed recurrent neural network (MR PI-RNN) for simultaneous MSK motion prediction and parameter identification.
• To address the challenge of mapping high-frequency sEMG signals to low-frequency joint motion.

Main Methods:

• Utilized fast wavelet transform to decompose sEMG and joint motion signals into multi-resolution components.
• Employed a gated recurrent unit (GRU) for training on coarse-scale signals, with parameters transferred recursively to finer scales (transfer learning).
• Ensured training satisfied underlying dynamic equilibrium for physics-informed prediction.

Main Results:

• The MR PI-RNN framework demonstrated higher accuracy in predicting elbow flexion-extension motion compared to single-scale training.
• Successfully identified physiologically consistent muscle parameters from subject kinematics data.
• Generated a physics-informed forward-dynamics surrogate model.

Conclusions:

• The proposed MR PI-RNN framework effectively integrates multi-resolution analysis and physics-informed deep learning for MSK motion prediction.
• This approach enhances accuracy and enables reliable parameter identification in complex biomechanical systems.
• Offers a robust method for understanding and modeling neuromuscular control and MSK dynamics.