Generating Patient’s Electronic Health Records with Unseen Diseases Using Ontology-enhanced Generative Adversarial Networks

Abstract The generation of realistic synthetic personal health data such as those found in electronic health records may be useful for fundamental research, applied AI models, and to examine safeguards for data privacy. Generative Adversarial Networks (GANs) have been employed for this purpose, but their ability to generate data is largely constrained by the limitations of the data they are trained on. Consequently, these models fail to generate realistic synthetic data relevant to rare diseases or complex conditions. In this paper, we propose a novel generative framework (Onto-CGAN) that integrates knowledge drawn from disease ontologies, with conditional GAN models to to generate realistic synthetic health data for diseases not present in the training data. We examine the quality of the generated data through 1) distributions of single variables, 2) correlation coefficients of variable pairs, and 3) predictive model performance using four machine learn- ing models. Our results demonstrate that Onto-CGAN can generate unseen data with statistical characteristics comparable to real data, and significantly improve the training of machine learning models in classifying diseases absent from the training set. This work represents a first step towards generating complex synthetic health data with characteristics that closely mimic those of real data such that they can be used in situations where there is a paucity of available data, and may find applications for data augmentation, hypothesis generation, and pre-clinical validation of clinical models.