Using Deep Learning Models to Predict Prosthetic Ankle Torque.

Area of Science:

• Biomechanics
• Machine Learning
• Prosthetics

Background:

• Inverse dynamics via motion capture is standard for biomechanical data but is lab-bound and marker-intensive.
• A practical alternative is needed for real-time prosthetic control systems.
• Predictive controllers require rapid, accurate biomechanical information.

Purpose of the Study:

• To develop deep learning models for estimating and predicting prosthetic ankle torque.
• To utilize minimal input signals for accurate dynamical system modeling.
• To enable predictive control in high-bandwidth prosthesis systems.

Main Methods:

• Applied deep learning to create dynamical system models.
• Used a hyperparameter optimization protocol for automatic model selection.
• Trained deep neural networks on six input signals to predict ankle torque.

Main Results:

• Deep neural networks predicted future ankle torques with high accuracy (2.9 ± 1.6% RMSE).
• Compared to analytical models (26.6 ± 40.9% RMSE), deep learning showed superior performance.
• Future predictions (half gait cycle) showed minimal performance degradation (1.7% RMSE increase).

Conclusions:

• Deep learning offers a viable method for accurate prosthetic ankle torque prediction.
• This approach reduces reliance on complex motion capture setups.
• Enables development of advanced predictive control for powered prosthetic limbs.