From Simulation to Reality: Predicting Torque With Fatigue Onset via Transfer Learning.

Area of Science:

• Biomechanics
• Artificial Intelligence
• Human Movement Science

Background:

• Muscle fatigue significantly affects upper extremity function, yet biomechanical models often neglect this.
• Accurate prediction of joint torques during fatiguing movements is crucial for understanding and mitigating functional limitations.

Purpose of the Study:

• To enhance the prediction accuracy of elbow flexion torque during fatiguing movements using a transfer learning approach.
• To compare the performance of a transfer learning model against a direct learning model and traditional biomechanical simulations.

Main Methods:

• Developed two artificial neural networks (ANNs): one using direct learning on recorded data, and another using transfer learning (pre-trained on simulated data, fine-tuned on recorded data).
• Utilized a musculoskeletal and muscle fatigue model for 1,701 simulations to generate pre-training data.
• Collected static subject-specific features and dynamic muscle activations/torques from 25 healthy adults during sustained elbow flexion.
• Employed a long short-term memory (LSTM) network architecture for torque prediction, integrating simulated and recorded datasets.

Main Results:

• The transfer learning model achieved a 24.9% lower root-mean-square error (6.22 Nm) compared to the direct learning model (8.28 Nm).
• Both ANN models outperformed conventional musculoskeletal simulations, which tended to underpredict elbow flexion torque.
• Transfer learning demonstrated improved robustness and reduced reliance on biomechanical model assumptions.

Conclusions:

• Transfer learning from simulated to recorded datasets is an effective strategy for improving torque predictions in fatiguing upper extremity movements.
• This approach enhances the accuracy and reliability of biomechanical predictions under real-world conditions.
• The findings suggest a promising direction for developing more sophisticated and adaptable models of human muscle function.