Machine Learning for Optical Motion Capture-Driven Musculoskeletal Modelling from Inertial Motion Capture Data.

Area of Science:

• Biomechanics
• Machine Learning
• Motion Capture Technology

Background:

• Marker-based Optical Motion Capture (OMC) provides detailed in vivo joint loading data but is lab-bound and costly.
• Inertial Motion Capture (IMC) offers a portable, cost-effective alternative but with lower accuracy.
• Musculoskeletal (MSK) modeling, crucial for analyzing motion capture data, is computationally intensive.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) approach for predicting high-fidelity MSK model outputs from easily acquired IMC data.
• To bridge the accuracy gap between IMC and OMC by leveraging ML for MSK analysis.
• To demonstrate the feasibility of translating advanced MSK modeling from laboratory settings to real-world applications.

Main Methods:

• Simultaneously collected OMC and IMC data from subjects were used to train various neural network (NN) architectures.
• Feed-Forward Neural Networks (FFNNs) and Recurrent Neural Networks (RNNs) (including LSTM and GRU) were investigated.
• A comprehensive hyperparameter search was conducted in both subject-exposed (SE) and subject-naive (SN) settings.

Main Results:

• Both FFNN and RNN models demonstrated comparable performance in predicting OMC-driven MSK outputs from IMC data.
• High agreement was observed between predicted and gold-standard OMC-driven MSK estimates (e.g., ravg,SE,FFNN=0.90±0.19).
• Models showed robust performance across both SE and SN testing scenarios.

Conclusions:

• ML models effectively map IMC data to accurate MSK outputs, mimicking gold-standard OMC results.
• This ML-driven approach facilitates the transition of MSK modeling from controlled lab environments to field-based applications.
• The study validates the potential of using ML with IMC for more accessible and widespread biomechanical analysis.