Generating Synthetic Intensive Care Unit Patient Records Using Adversarially Filtered Large Language Models: Development and Evaluation Study (Preprint)

BACKGROUND Access to real-world electronic health records (EHRs) is limited by privacy regulations, governance constraints, data heterogeneity, and other factors that have been widely reported in the literature, collectively hindering the scalability of AI-driven clinical research. Synthetic health data have therefore emerged as a potential remedy, yet existing generative approaches often fail to ensure both statistical fidelity and clinical plausibility. OBJECTIVE This study aimed to develop and evaluate a novel framework for generating synthetic ICU patient records that balance statistical realism with medical consistency. We focus on distributional fidelity and clinical coherence as prerequisites for safe cohort augmentation, enabling privacy-preserving applications in healthcare AI. METHODS We designed a two-stage generation pipeline combining a prompt-based LLM with a post-hoc adversarial filtering mechanism. The LLM generated patient profiles in structured text format, which were parsed into tabular data before being filtered by a one-class discriminator (XGBoost) trained exclusively on real patient records. Records were retained based on a learned realism score. The framework was evaluated on the MIMIC-III dataset (n ≈ 40,000 patients, 20 selected clinical variables). Performance was compared with CTGAN and baseline LLM generation using (1) distributional similarity metrics [Kolmogorov–Smirnov (KS), total variation distance (TVD)], (2) inter-variable correlation preservation (Frobenius norm, mean correlation matrix distance), (3) rule-based medical consistency checks, and (4) downstream predictive tasks with Random Forest and LightGBM. RESULTS The adversarially filtered LLM demonstrated improved distributional alignment and preservation of inter-variable structure compared with CTGAN and baseline LLM generation, while eliminating medically implausible records identified in GAN-based outputs. In downstream predictive tasks, models trained on synthetic data generated by the proposed framework achieved performance comparable to models trained on real data when evaluated on an independent real test set. CONCLUSIONS We designed a two-stage generation pipeline combining a prompt-based LLM with a post-hoc adversarial filtering mechanism. The LLM generated patient profiles in structured text format, which were parsed into tabular data before being filtered by a one-class discriminator (XGBoost) trained exclusively on real patient records. Records were retained based on a learned realism score. The framework was evaluated on the MIMIC-III dataset (n ≈ 40,000 patients, 20 selected clinical variables). Performance was compared with CTGAN and baseline LLM generation using (1) distributional similarity metrics [Kolmogorov–Smirnov (KS), total variation distance (TVD)], (2) inter-variable correlation preservation (Frobenius norm, mean correlation matrix distance), (3) rule-based medical consistency checks, and (4) downstream predictive tasks with Random Forest and LightGBM.