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

Machine learning is a powerful tool for modeling complex movements. However, the high cost of acquiring biomechanical data often limits observations and therefore the generalizability of machine learning models. Meanwhile, there is untapped utility in presently published musculoskeletal simulations for informing these predictions. This work evaluated whether a transfer learning model trained on both musculoskeletal simulation and recorded data outperforms a direct learning model trained on recorded experimental data alone. For the transfer learning model, we leveraged a simulation dataset with 6,594 lateral pinch simulations to train a long short-term memory neural network to predict three-component thumb-tip forces from muscle activations. We concatenated this pre-trained neural network with a feedforward architecture and used recorded data (n = 12 subjects) to fine-tune the model. The feedforward architecture used the outputs of the long short-term memory neural network, as well as static subject-specific features (e.g., anthropometric measurements). We then trained a direct learning model with a similar architecture using just the recorded data. The transfer learning and direct learning models were parameterized via a random search and underwent five-fold cross-validation. To elucidate the impact of pre-training with simulations, we validated both neural networks and compared their test errors using a leave-out set (n = 3 subjects). The transfer learning model outperformed the direct learning model, as suggested by statistically significant lower absolute errors (p = 0.04). Our results suggest the viability of using musculoskeletal simulations to bolster machine learning models that can expand the utility of simulated datasets and potentially improve predictions of real-world biomechanics.