Real-Time Prediction of Joint Forces by Motion Capture and Machine Learning

Area of Science:

• Biomechanics
• Machine Learning
• Human Motion Analysis

Background:

• Conventional biomechanical modeling is computationally intensive due to large equation systems.
• Improving computational speed and stability is crucial for real-time applications in human motion analysis.

Purpose of the Study:

• To apply machine learning to train surrogate models for predicting knee contact forces (KCFs) in near real-time.
• To enhance the computational efficiency of biomechanical modeling for human gait analysis.

Main Methods:

• Utilized artificial neural networks (ANNs) and support vector regression (SVR) to predict KCFs.
• Fused optical motion capture and musculoskeletal modeling data, including kinematic and force variables.
• Investigated training schemes with data from all subjects or subsets, with and without ground reaction forces (GRFs).

Main Results:

• ANNs demonstrated superior performance in predicting KCFs compared to SVR.
• High prediction accuracy (Pearson R: 0.89-0.98) was achieved with ANNs, even without ground reaction forces (GRFs).
• The models showed robust predictive power across different training and data inclusion scenarios.

Conclusions:

• Artificial neural networks are highly effective for predicting multi-component KCFs during walking at various speeds.
• The developed models offer a computationally efficient solution for real-time knee loading condition assessment.
• This approach has significant potential for guiding patient treatment and rehabilitation in clinical settings.