Exploring the Potential of Generative Adversarial Networks for Synthetic Medical Data Generation
Luka Jovanović, Aleksandar Petrović, Tamara Živković, Miloš Antonijević, Nebojša Bačanin, Miodrag Živković
Abstract
Proper artificial intelligence (AI) and machine learning (ML) model implementation and training are highly reliant on quality input data. However, the high costs of clinical studies and increasing privacy concerns limit the availability of medical data to researchers. This work explores the potential of applying generative adversarial networks (GAN) for generating synthetic medical data based on publicly available real-world datasets. Synthetic data is generated based on real-world observations and used to train several contemporary ML and AI models. These models are subjected to comparison with models trained on real data. Both trained models are evaluated under identical conditions on real-world testing data. These well-known, high-quality datasets are explored in this study covering diabetes and heart disease. While models trained on synthetic data show a slight decrease in objective performance, they still demonstrate viably accurate outcomes. This suggests that the introduced approach presents a potential stepping stone towards improvising data availability for researchers while respecting patients’ best interests and patient privacy.