Addressing the Challenges in Using Synthetic Data for Health Research: Application to Cardiology.

Abstract

Unlabelled: Synthetic data offer significant potential for cardiology research by enabling data sharing, preserving privacy, and supporting machine learning model development. By generating artificial patient records that reflect real-world distributions, synthetic data can accelerate clinical research, improve model performance for rare cardiovascular conditions, and facilitate transnational collaborations that would otherwise be restricted by data-sharing barriers. Despite these advantages, the increasing use of synthetic data raises important ethical, regulatory, and methodological concerns that remain insufficiently addressed. Key challenges include assessing the validity and generalizability of synthetic datasets, understanding their limitations in representing complex and heterogeneous patient populations, and preventing the amplification of existing biases in cardiovascular care. Current regulatory frameworks, including the General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA), do not fully address emerging risks such as reidentification and data leakage, and there is no harmonized guidance to govern the use of synthetic data as stand-alone evidence for medical device evaluation or therapeutic research. In this viewpoint, we argue that responsible integration of synthetic data in cardiology requires, first, clear differentiation between synthetic data as a privacy-preserving distributional substitute and synthetic data as a counterfactual simulation tool, and, second, fit-for-purpose governance frameworks that pair rigorous utility and fidelity testing with explicit, adversary-aware privacy evaluation before synthetic cohorts are accepted as evidence in research or product evaluation. A prerequisite for that governance is conceptual clarity about what synthetic data are being used for. Synthetic data in health care serve 2 fundamentally distinct roles that carry entirely different validity requirements, failure modes, and regulatory implications, yet they are routinely conflated. The first role is as a privacy-preserving distributional substitute: the goal is statistical fidelity to the real data distribution, so that analyses of the synthetic dataset yield results equivalent to those of the original. The second role is as a tool for counterfactual simulation: the goal is to generate data that could not have been observed, such as rare conditions, hypothetical interventions, or extrapolations to new populations. These 2 roles are methodologically distinct. A dataset that accurately reflects real-world distributions may be inadequate for extrapolating findings to underrepresented subgroups. Conversely, a simulator optimized for novel scenario generation may systematically diverge from real-world distributions. This distinction informs every subsequent discussion of validity, bias, and regulation in this viewpoint and our proposed 4 concrete actions for the cardiology research community, including mandatory 3-layer (fidelity, utility, and privacy) validation, systematic subgroup reporting, explicit intended-use scoping, and domain-specific acceptability thresholds for synthetic data-based evidence.