Synthetic data generation in motion analysis: A generative deep learning framework.

Area of Science:

• Biomechanics
• Machine Learning
• Data Science

Background:

• Motion analysis often faces data scarcity, limiting model training.
• Generative deep learning offers a solution for data augmentation.
• Variational autoencoders (VAEs) are a powerful tool for synthetic data generation.

Purpose of the Study:

• To introduce a VAE-based data augmentation strategy for motion analysis.
• To generate synthetic kinematic and kinetic variables.
• To evaluate the effectiveness of augmented data in improving biomechanical predictions.

Main Methods:

• Utilized a variational autoencoder to generate synthetic kinematic (joint angles) and kinetic (joint moments, ground reaction forces) data.
• Employed Statistical Parametric Mapping (SPM) to compare real and synthetic data distributions.
• Trained a long-short term memory (LSTM) model on real data (R) and combined real and synthetic data (R&S) for performance assessment.

Main Results:

• SPM confirmed no significant differences between real and synthetic biomechanical data.
• LSTM models trained on R&S achieved comparable or superior normalized root mean squared error (nRMSE) compared to models trained solely on R.
• Specific improvements noted in predicting knee and hip joint moments across frontal and sagittal planes.

Conclusions:

• VAE-based data augmentation is effective for motion analysis, particularly when real data is limited.
• Synthetic data generated by VAEs can enhance the performance of predictive models like LSTMs.
• This approach offers a novel and efficient method for augmenting biomechanical datasets.