Marika NiihoriA new paper from the van der Schaar Lab introduces SynthCraft – an AI partner that allows clinicians and healthcare researchers to generate and evaluate synthetic data using natural language, with no machine learning expertise required.
The Data Problem at the Heart of Healthcare Research
Access to high-quality clinical data underpins virtually every advance in modern healthcare research. Yet for decades, legitimate privacy concerns, regulatory constraints, and the fragmentation of health systems have made such access difficult. Even when data can be shared, it is often incomplete, biased, or doesn’t represent the full diversity of patients. Researchers are frequently forced to work with a narrow or skewed window onto the real world. One that may miss entire populations, rare disease presentations, or the complex interplay between demographics factors and health outcomes. The consequences are significant. AI models trained on such data inevitably inherit these blind spots, reinforcing existing biases and perpetuating the cycle.
Synthetic data has emerged as one of the most promising responses to this challenge. Instead of using real patient records, synthetic datasets are artificially generated to mirror the statistical patterns and relationships present in real clinical data. This preserves their analytical value while avoiding the direct exposure of patient information and addressing key privacy constraints. It opens the door to data augmentation: enriching datasets that are sparse or imbalanced, supporting the development of digital twins and enabling robust evaluation of how models perform when real-world conditions shift.
Despite its promises, the use of synthetic data in healthcare remains limited in practice. Generating high-quality synthetic data is technically complex, the methodological landscape is evolving rapidly, and there is currently no agreed standard for how synthetic data should be evaluated across datasets and use cases. Beyond these technical challenges, most existing tools still require advanced programming skills and familiarity with complex data pipelines. Many clinicians and healthcare researchers simply may not have this expertise or are likely to not have the time to apply effectively. Guidelines and training can help bridge the gap, but their reach is limited and there is no mechanism to ensure they are consistently followed in practice.
Developing synthetic data has been a central focus of the van der Schaar Lab for several years. The group has developed methods for generating, evaluating, and applying synthetic healthcare data across a range of use cases – from privacy-preserving data access to augmentation and fairness. At the same time, the lab has recognised that the practical adoption of these methods depends not only on technical advances, but also on making them accessible to the clinicians and healthcare researchers who need them. An important first step in this direction was development of SynthCity: an open-source library that provides a standardised, comprehensive interface for generating and benchmarking synthetic datasets across tabular, genomic, and other data types. SynthCity is widely used by machine learning researchers and data scientists. The platform provides a flexible framework for developing and evaluating synthetic data generators, supporting a wide range of modelling approaches. However, SynthCity, as powerful as it is, still requires significant programming expertise. While the tools exist, the technical expertise required to use them does not always match the skillsets of those conducting clinical trials and healthcare research.
Introducing SynthCraft: An AI Partner for Synthetic Data
The recently accepted paper, “SynthCraft: an AI partner for synthetic data generation to support data access and augmentation in healthcare” in PLOS Digital Health addresses this challenge by introducing a framework designed to make synthetic data generation more accessible to clinicians and healthcare researchers (Figure 1). Rather than requiring the users to write code or configure complex pipelines, SynthCraft operates as a human-in-the-loop AI partner: a large language model (LLM)-based framework that guides researchers through the entire synthetic data workflow using natural language. Users describe their data, their goals, and their constraints – SynthCraft handles the rest, orchestrating the underlying SynthCity tools, discussing trade-offs, and iterating based on feedback.
Figure 1 SynthCity and SynthCraft serve different users across the knowledge spectrum, while sharing the same underlying synthetic data engine.
The system uses a reinforcement learning-based reasoning engine to adapt its workflow dynamically to the user’s needs and the characteristics of their dataset. At each stage, SynthCraft can select appropriate actions, receive feedback, and course-correct: for example, returning to try alternative generators if the quality of synthetic data falls short at the evaluation stage. On completion, it produces a structured report of every step taken, decision made, any other quality metrics and piece of code run, supporting the transparency and reproducibility that rigorous research demands.
Figure 2 The SynthCraft workflow. The researcher inputs their dataset and research question in natural language; SynthCraft guides the full pipeline from data preparation through to a reproducible report.
Crucially, the underlying LLMs never have direct access to the real or synthetic data at any point. All analyses run on the user’s own device or a compliant cloud environment, maintaining the privacy safeguards that are non-negotiable in healthcare settings.
The paper demonstrates SynthCraft’s capabilities across two datasets: the National Health and Nutrition Examination Survey (NHANES), a national survey that measures the health and nutrition of adults and children in the United States, and The Cancer Genome Atlas (TCGA), a high-dimensional genomic dataset covering more than 11,000 cancer patients.
Using SynthCraft, the team analysed the privacy, statistical fidelity, and downstream utility of four different synthetic data generators – PATE-GAN, DDPM, ADS-GAN, and CT-GAN – with and without explicit privacy-preserving design. In keeping with previous work, we found that no single generator was optimal across all use cases and datasets.
The paper also explores data augmentation to address ethnic imbalances common in research cohorts. Applied to NHANES, SynthCraft guides an ethnicity-specific augmentation pipeline to improve representation across demographic groups. Based on this case study, the augmentation did not uniformly improve predictive performance. In itself, this is an important result: demographic balance in a dataset does not automatically remove bias in downstream models. Iterative testing across generators and use cases is essential – SynthCraft is designed to support exactly that.
In the genomic setting, SynthCraft generated synthetic versions of TCGA data that enabled regression models to predict tumour purity with accuracy comparable to models trained on the real dataset, demonstrating the tool’s capability beyond tabular clinical data.
Dr Thomas Callender, the first author of the paper states:
Biomedical research has, for too long, had problems of accessibility, reproducibility and transparency. Tools like SynthCraft are a first step into a future of human-AI partnership in research methods. They can empower researchers, giving them access to state-of-the-art tools that can support their research. And, they herald a new approach to improving standards in research.
To further support this statement Dr Anders Boyd comments that,
We have to provide the data underlying our analysis in biomedical research—this is an absolute must for reproducibility. However, the highly sensitive nature of our data precludes us from doing so. Providing synthetic data is a natural next step, yet the options are numerous and the programming for some arduous. SynthCraft is meant to be the perfect companion for this task—allowing to select, produce and test synthetic data with the goal of optimizing reproducibility.
SynthCraft is part of a broader vision for how AI partners enable the next generation of improved quality, transparency and accessibility of scientific research. Rather than replacing researchers, SynthCraft ensures that state-of-the-art methods are not confined to those with the technical background to use them, but accessible to researchers, clinicians and institutions that need them most.
If you are interested in incorporating these AI partners, please see below:
- SynthCraft is available at: https://github.com/vanderschaarlab/SynthCraft
- SynthCity is available at: https://github.com/vanderschaarlab/synthcity
- The paper is now available at PLOS Digital Health: https://doi.org/10.1371/journal.pdig.0001290
