A Study On Synthesizing Electronic Health Record Data For Privacy Preservation

Abstract Predicting the risk of developing various diseases using machine learning models, deep learning, and neural networks trained on electronic health record (EHR) data has become crucial in today’s healthcare landscape. However, safeguarding personal information remains a significant concern due to the potential for sensitive information leakage from EHR databases, which often contain personal information that, if revealed, can lead to identity theft, discrimination, and other types of harm. This study aims to apply synthetic data as a replacement for real databases to protect personal information while ensuring the responsible use of data for research, AI applications, and collaborative analysis. The research was conducted on two different EHR databases using a statistical model (Gaussian Copula Synthesizer) and a deep learning model (Conditional Generative Adversarial Network Synthesizer). Specifically, the study focused on forecasting diabetes risks in the next year and predicting Intensive Care Unit outcomes in EHRs. The quality of these synthetic datasets was evaluated using statistical metrics and machine learning efficiency. Moreover, data security efficiency was assessed using statistical-based privacy metrics, including k-Anonymity, l-Diversity, and ε–Identifiability values. The results demonstrate the significant quality and privacy of synthetic data generated by both models, highlighting the potential of the proposed pipeline of synthetic data-generating models in future healthcare applications.