Multi-Action Knee Contact Force Prediction by Domain Adaptation.

Area of Science:

• Biomechanics
• Machine Learning
• Musculoskeletal Dynamics

Background:

• Current musculoskeletal dynamics estimation methods often fail to generalize across different tasks, being limited to predefined actions like gait.
• Accurate estimation of internal biomechanical forces is crucial for understanding joint health and developing effective rehabilitation strategies.

Purpose of the Study:

• To develop a generalized deep learning model capable of estimating internal biomechanical forces for multiple actions simultaneously.
• To improve the adaptability and accuracy of musculoskeletal dynamics estimation by incorporating unsupervised domain adaptation techniques.

Main Methods:

• A Bidirectional Long Short-Term Memory network was developed for knee contact force prediction.
• Correlation alignment layers were integrated to minimize domain shift between kinematic data from different actions.
• A Neural State Machine (NSM) simulation platform was utilized for real-time testing and visualization across varied 3D scene geometries.

Main Results:

• The proposed deep learning architecture with domain adaptation demonstrated superior performance compared to benchmark models, evidenced by lower Normalized RMSE (NRMSE) and significant t-test results.
• The model successfully predicted knee contact forces for multiple action classes using a single architecture.
• Alignment across action classes and real-to-synthetic data showed effective domain adaptation.

Conclusions:

• The developed method enables accurate prediction of knee contact forces for diverse actions, advancing the estimation of internal forces for intermediate movements.
• This approach holds potential for personalized rehabilitation by leveraging knowledge of hidden motion states.
• The model's seamless integration into human motion simulation environments facilitates automated and computationally efficient biomechanical analysis.