Lower Limb Joint Torque Prediction Using Long Short-Term Memory Network and Gaussian Process Regression.

Area of Science:

• Biomechanics and Movement Science
• Machine Learning in Healthcare
• Computational Neuroscience

Background:

• Accurate joint torque prediction is crucial for biomechanical analysis and applications.
• Traditional methods like inverse dynamics and EMG-driven neuromusculoskeletal models have limitations.
• Machine learning offers a promising alternative for joint torque prediction using sEMG and kinematic data.

Purpose of the Study:

• To predict lower limb joint torque in the sagittal plane during walking.
• To evaluate the effectiveness of Long Short-Term Memory (LSTM) and Gaussian Process Regression (GPR) models.
• To utilize surface electromyography (sEMG) signal features and joint angles as inputs.

Main Methods:

• Extracted seven features from sEMG signals of five muscles.
• Used three joint angles as input parameters.
• Implemented and compared LSTM and GPR machine learning models.
• Validated model performance using Normalized Root Mean Squared Error (NRMSE) and Pearson correlation coefficient (R).

Main Results:

• Both LSTM and GPR models achieved high prediction accuracy for joint torque.
• Majority of NRMSE values were below 15% for both models.
• Most Pearson correlation coefficient (R) values exceeded 0.85, indicating strong linear relationships.
• Most coefficient of determination (R2) values surpassed 0.75, demonstrating good model fit.

Conclusions:

• Machine learning models, specifically LSTM and GPR, can accurately predict lower limb joint torque during walking.
• sEMG signal features and joint angles are effective inputs for these predictive models.
• The findings support the feasibility of using machine learning for non-invasive joint torque estimation.