Optimizing Rare Disease Gait Classification through Data Balancing and Generative AI: Insights from Hereditary...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Rare Disease Research

Background:

• Gait analysis in rare diseases, such as primary hereditary cerebellar ataxia (pwCA), faces challenges due to small sample sizes and imbalanced datasets, limiting study interpretability.
• Accurate gait abnormality detection is crucial for diagnosing and managing pwCA, but data limitations hinder the development of robust diagnostic models.

Purpose of the Study:

• To evaluate the efficacy of data balancing techniques and generative artificial intelligence (AI) algorithms in creating synthetic gait data for pwCA.
• To assess if synthetic data can accurately reflect real gait abnormalities and improve diagnostic model performance and explainability.

Main Methods:

• Collected lumbar-level gait data using inertial measurement units from 30 pwCA and 100 healthy subjects.
• Applied various data balancing methods, including subsampling, oversampling, SMOTE, GANs, and specifically conditional tabular generative adversarial networks (ctGAN).
• Utilized a random forest classifier and calculated consistency and explainability metrics to evaluate the generated datasets.

Main Results:

• Conditional tabular generative adversarial networks (ctGAN) significantly enhanced classification performance compared to the original dataset and traditional augmentation methods.
• The synthetic data generated by ctGAN demonstrated consistency with known gait abnormalities in pwCA.
• ctGAN improved the explainability of the diagnostic models.

Conclusions:

• Conditional tabular generative adversarial networks (ctGAN) are effective for balancing tabular datasets in rare disease populations.
• ctGAN can improve the performance and explainability of diagnostic models for rare diseases like pwCA.
• This approach offers a promising solution for overcoming data limitations in rare disease research.