Transfer Learning With Simulated and Recorded Data Improves Predictions of Lateral Pinch Thumb-Tip Forces

Area of Science:

• Biomechanics
• Machine Learning
• Computational Modeling

Background:

• Biomechanical data acquisition is costly, limiting machine learning model generalizability for complex movements.
• Published musculoskeletal simulations offer a valuable, yet underutilized, data source for improving predictive models.

Purpose of the Study:

• To evaluate if a transfer learning model, pre-trained on musculoskeletal simulations and fine-tuned on recorded data, outperforms a direct learning model trained only on recorded data.
• To assess the efficacy of leveraging simulation data to enhance machine learning predictions in biomechanics.

Main Methods:

• A long short-term memory network was trained on 6,594 lateral pinch simulations to predict thumb forces from muscle activations.
• This pre-trained network was integrated into a transfer learning model, fine-tuned with recorded data (n=12 subjects).
• A direct learning model with a similar architecture was trained exclusively on recorded data; both models underwent cross-validation.

Main Results:

• The transfer learning model demonstrated statistically significant lower absolute errors (p = 0.04) compared to the direct learning model.
• Validation using a leave-out set (n=3 subjects) confirmed the superior performance of the transfer learning approach.

Conclusions:

• Musculoskeletal simulations can effectively augment machine learning models for biomechanical predictions.
• This transfer learning strategy enhances the utility of simulated datasets and improves the prediction accuracy of real-world biomechanics.