Large language models as synthetic electronic health record data generators

Electronic health record (EHR) data consists of a wealth of information that can be used for driving clinical research and improving patient care. However, due to the complex and sensitive nature of EHR data, there are strict data regulations and privacy concerns around data sharing. Generating adequately validated synthetic EHR data from scratch, such that it is representative of real data, is a viable and attractive solution to address such data-sharing bottlenecks. In this work, we investigate the adoption and implementation of large language models (LLMs) as a sustainable and scalable deep learning approach for generating high-fidelity EHR data. The findings of this study demonstrate that LLMs outperform commonly used generative modeling frameworks, such as variational autoencoders and generative adversarial networks, and recently introduced diffusion models.