How Prompt Design Affects Synthetic Healthcare Data Generated by LLMs

Large Language Models (LLMs) show potential for generating synthetic clinical data, addressing privacy and data scarcity challenges. This study investigates how prompt design influences the statistical realism of such data, comparing structured versus narrative prompts across different model architectures and sizes. Evaluating six LLMs with Jensen-Shannon divergence and Kolmogorov-Smirnov metrics, the results reveal that performance depends on the interaction between prompt style and model type: proprietary models favor structured prompts, while large open-weight models perform better with narrative ones. Model size alone does not guarantee higher data quality. These findings highlight the need to adapt prompt design to model characteristics to improve the realism of synthetic clinical data.