Generating unseen diseases patient data using ontology enhanced generative adversarial networks

Area of Science:

• Health Informatics
• Artificial Intelligence
• Bioinformatics

Background:

• Synthetic health data generation is crucial for research, AI development, and data privacy.
• Generative Adversarial Networks (GANs) are limited by training data, struggling with rare or unseen diseases.

Purpose of the Study:

• To propose Onto-CGAN, a novel generative framework integrating disease ontologies with GANs.
• To generate realistic synthetic health data for unseen diseases.
• To address data scarcity for rare disease research.

Main Methods:

• Developed Onto-CGAN, a framework combining disease ontologies and GANs.
• Evaluated generated data quality using variable distributions, correlation coefficients, and machine learning model performance.

Main Results:

• Onto-CGAN successfully generates unseen diseases with statistical properties similar to real data.
• The framework significantly improves the training of machine learning models.
• Demonstrated comparable statistical characteristics to real data for generated unseen diseases.

Conclusions:

• Onto-CGAN effectively generates synthetic data for rare and unseen diseases.
• This approach enhances data augmentation, hypothesis generation, and preclinical model validation.
• Addresses limitations of traditional GANs in health data synthesis.