Adaptive refinement and prompt-guided conditioning for clinically realistic LLM-generated synthetic pediatric data

The development of reliable healthcare systems requires high-quality and representative datasets. Current issues, such as limited data access, bias, and the lack of diversity in Electronic Health Record (EHR) data, limit the model's generalizability, reliability, and robustness. Specifically, generating synthetic pediatric intensive care unit (PICU) data is challenging due to high age-related variability. Subsequently, the prediction task in this domain poses additional challenges due to complex feature dependencies. Most current approaches, such as GAN-based models, do not distinguish between features and the outcome variable, so they are not designed to improve prediction. Leveraging the power of LLMs for medical data generation required careful consideration of the prompt design to preserve the underlying data distribution. Besides, it is essential to consider dependency-aware guides for the models to generate high-quality synthetic data that not only mitigates existing dataset limitations but also improves prediction performance.

This work proposes an LLM-based synthetic data generation pipeline integrating a rule-guided prompting method and advanced post-processing strategies. We validate the proposed framework on the task of predicting pediatric intensive care length of stay. We introduce a structure-prompt design that includes per-class guides for the most influential features, integrating key feature-feature interactions and highly correlated feature pairs. Subsequently, the rules were formulated based on the extracted statistics and frequency distributions derived from realistic clinical data. We introduce a density-aware refinement method to ensure diversity, improve fidelity, and preserve utility. Furthermore, we employed two LLMs, which are ChatGPT and Claude, and conducted a comprehensive evaluation.

Both the Claude and Fine-tuned ChatGPT models demonstrated results comparable to baselines (CTGAN and TVAE) in terms of Macro-F1, PR-AUC, and MCC. Despite the reduction in sample size caused by the density-based filtering stage, the models achieved slight improvements over the raw synthetic data in utility evaluation. This indicates that the proposed post-processing strategy improved sample quality without compromising downstream performance.

The findings demonstrate the effectiveness of integrating prompt-guided statistical modeling with density-aware adaptive refinement in addressing the data scarcity and generating realistic synthetic data in such risk-sensitive domains. Additionally, it highlights the success in supporting robust predictive modeling in data-constrained settings.