Enhancing biomechanical machine learning with limited data: generating realistic synthetic posture data using...

Area of Science:

• Biomechanics
• Machine Learning
• Artificial Intelligence

Background:

• Machine learning (ML) models, especially deep learning, require extensive data for optimal performance.
• Biomechanical datasets are often limited due to various data collection challenges.
• Methods for augmenting data in human posture ML models are scarce.

Purpose of the Study:

• To explore the feasibility of using generative AI to create realistic synthetic 3D human posture data.
• To assess the quality and distinguishability of AI-generated synthetic data compared to real data.
• To demonstrate the benefits of incorporating augmented posture data into ML model training.

Main Methods:

• Collected 3D posture data from 338 subjects using surface topography.
• Employed a Variational Autoencoder (VAE) architecture to generate synthetic posture data.
• Evaluated synthetic data using domain experts, ML classifiers, and Statistical Parametric Mapping (SPM).
• Utilized a deep autoencoder (AE) to showcase the impact of augmented data on feature representation.

Main Results:

• Synthetic data proved difficult to distinguish from real data by experts and ML classifiers, confirmed by SPM.
• Integrating synthetic data into AE training reduced reconstruction error compared to using real data alone.
• Achieved comparable reconstruction accuracy with reduced latent dimensions when using augmented data.

Conclusions:

• Generative AI offers a promising approach for data augmentation in biomechanics.
• Synthetic data can enhance the performance and efficiency of ML models in posture analysis.
• Harnessing generative AI can significantly advance ML applications within the biomechanics field.