A Feature-Encoded Physics-Informed Parameter Identification Neural Network for Musculoskeletal Systems.

Area of Science:

• Biomechanics
• Computational Biology
• Machine Learning

Background:

• Musculoskeletal (MSK) digital twins require accurate muscle-tendon force and activity identification from physiological data like motion and surface electromyography (sEMG).
• Existing data-driven machine learning models for motion prediction from sEMG are often black-box, lack physical consistency, and have limited generalizability.

Purpose of the Study:

• To develop a novel Feature-Encoded Physics-Informed Parameter Identification Neural Network (FEPI-PINN) for simultaneous motion prediction and parameter identification in human MSK systems.
• To improve the accuracy and generalizability of MSK digital twin models by incorporating physical laws into machine learning.

Main Methods:

• Projecting high-dimensional, noisy sEMG signal features into a low-dimensional, noise-filtered embedding space to enhance forward dynamics prediction.
• Training the FEPI-PINN model to establish relationships between sEMG signals and joint motion while simultaneously identifying key MSK parameters.
• Utilizing physics-informed neural networks to ensure predictions adhere to underlying biomechanical principles.

Main Results:

• The FEPI-PINN framework effectively identified subject-specific muscle parameters.
• The trained physics-informed forward-dynamics surrogate model accurately predicted elbow flexion-extension motion.
• Predicted motion aligned well with measured joint motion data, demonstrating the model's efficacy.

Conclusions:

• The proposed FEPI-PINN offers a robust approach for subject-specific MSK digital twin construction.
• This method enhances the prediction of human motion and the identification of crucial biomechanical parameters.
• The integration of physics principles improves the reliability and generalizability of data-driven MSK models.