Physics-Informed Deep Learning for Muscle Force Prediction With Unlabeled sEMG Signals

Area of Science:

• Computational biomechanics
• Human movement analysis
• Deep learning applications

Background:

• Physics-based models offer insights into neural drive, muscle dynamics, and joint kinematics but are computationally intensive.
• Data-driven methods are faster but typically require difficult-to-obtain labeled data for training.
• Existing methods face challenges with computational latency and data acquisition for accurate biomechanical analysis.

Purpose of the Study:

• To develop a novel physics-informed deep learning method for predicting muscle forces without requiring labeled training data.
• To enable the identification of personalized muscle-tendon parameters.
• To improve the efficiency and accuracy of computational biomechanical modeling.

Main Methods:

• Embedded a Hill muscle model-based forward dynamics simulation within a deep neural network as an additional loss function.
• Utilized a fully connected neural network (FNN) architecture.
• Validated the method on wrist joint data from six healthy subjects.

Main Results:

• The proposed method accurately predicted muscle forces, achieving comparable or lower root mean square error (RMSE) than baseline methods using labeled surface electromyography (sEMG) data.
• Demonstrated high coefficient of determination for muscle force predictions.
• Successfully identified personalized muscle-tendon parameters.

Conclusions:

• The physics-informed deep learning approach effectively predicts muscle forces and identifies muscle-tendon parameters without labeled data.
• This method offers a promising solution to overcome the limitations of traditional biomechanical modeling techniques.
• The findings highlight the potential of integrating physical principles into deep learning for advancing human movement analysis.