SynthCraft: An AI partner for synthetic data generation to support data access and augmentation in healthcare

Access to high-quality data provides the foundation for biomedical research. But data access is often limited or challenging due to privacy constraints, whilst the data themselves may be unrepresentative or sparse. Synthetic data can support both privacy-preserving data access and advanced analytical workflows, including data augmentation or the development of digital twins. However, the use of synthetic data remains limited due to the complexity of the methods themselves, their use, and their evaluation. To address this, we developed SynthCraft, an AI tool to support the principled, transparent, application of state-of-the-art synthetic data generation methods. SynthCraft couples a reinforcement learning-based reasoning engine with large language models (LLMs) to orchestrate the workflow necessary for the generation of synthetic data based on dynamic interaction with the user through natural language. We demonstrate the capability of SynthCraft with both tabular and genomic datasets: the National Health and Nutrition Examination Survey (NHANES) and the Cancer Genome Atlas (TCGA). Using SynthCraft, we analysed the privacy, statistical fidelity, and downstream utility of four different synthetic data generators both with and without explicit privacy-preserving designs when applied to both the NHANES and TCGA datasets. We show that how different generators perform differently - and that no single method was optimal - across varying use-cases and datasets. Furthermore, we demonstrate how SynthCraft can be used for data augmentation as part of a workflow to attempt to mitigate imbalances in the proportion of individuals from different ethnic backgrounds. In conclusion, a human-in-the-loop AI partner using LLMs can support the generation of synthetic datasets. Such tools could improve the quality, reproducibility, and transparency of research methods, whilst increasing their accessibility. Research into their use across different methodological areas is warranted.