Predicting Lower Limb Muscle Forces from Ground Reaction Forces During Gait Using Sequence and Attention-Based Deep Learning Models

Description

This project aims to develop a data-driven model of the relationship between ground reaction forces and muscle forces using deep learning techniques. This model will be trained on a data from ten healthy individuals walking on a force-instrumented treadmill in a motion capture laboratory that has been collected, processed, and open-sourced by Uhlrich et al., 2022.

Four models were explored, including LSTM, CNN-LSTM, LSTM with Attention, and a Transformer. All models achieved strong predictive performance, with the Transformer model consistently outperforming others in accuracy across most muscle forces and overall. These results highlight the potential of deep learning to capture the complex relationships and patterns between GRFs and lower limb muscle forces during gait. This work provides the groundwork for advancing data-driven approaches in robotic cadaveric gait simulation, enabling more reliable and flexible control strategies to replicate physiological motions in cadavers.

Getting Started

Dependencies

• Python 3.11.10
• PyTorch 2.5.1 or later
• NumPy
• SciPy
• Matplotlib
• OpenSim 4.5 or later
• OpenSim Python API
• OpenSim MATLAB API

Data

1. Download dataset from the Coordination Retraining Project on SimTK
2. Extract the dataset to the 'data' directory

Execution

1. Run experimental gait data through the custom static opimization pipeline developed by Uhlrich et al., 2022
2. Batch process and segment ground reaction forces and estimated muscle forces using the OpenSim API
3. Preprocess data by resampling and normalizing data
4. Save data to a .npy file
5. Load .npy file and randomly split data into training, validation, and test sets
6. Implement, tune, and train deep learning models
7. Evaluate model performance on test set
8. Calculate and visualize model performance metrics

Additional Documentation

• OpenSim API Documentation
• PyTorch Documentation

Contributing

1. Create your branch
2. Commit your changes
3. Push to the branch
4. Open a pull request

License

This project is licensed under the Creative Commons Zero License.

Authors

• Anthony Le (anth0nyhle, anthony.le@utah.edu)

Verison History

• 0.1
  • Initial Release

Acknowledgements

1. Uhlrich SD, Jackson RW, Seth A, Kolesar JA, Delp SL, 2022. Muscle coordination retraining inspired by musculoskeletal simulations reduces knee contact force. Scientific Reports 12, 9842. https://doi.org/10.1038/s41598-022-13386-9.