Epidemiol Prev. 2026 May-Jun;50(3):279-289. doi: 10.19191/EP26.3.A992.059.
ABSTRACT
BACKGROUND: the use of real-world (RW) data has become an integral part of pharmacoepidemiological research and regulatory assessment. However, access to individual-level data, especially for purposes other than the original ones (secondary use), is often restricted by legal and ethical constraints related to privacy protection. In this context, generating synthetic data, defined as artificially generated data that reproduce the statistical properties and relationships of real data without containing information traceable to real individuals, offers an innovative opportunity to balance privacy protection with the need to generate RW evidence to support regulatory decision-making.
OBJECTIVES: to compare two methods of generating tabular synthetic data: synthpop, based on transparent and interpretable inferential methodologies, and Conditional Tabular-Generative Adversarial Networks (CT-GANs), which leverage deep learning approaches to reproduce complex multivariate distributions. The focus is on evaluating the ability of both approaches to preserve the statistical structure of real data and reduce the risk of disclosure/re-identification.
DESIGN: comparative study of two synthetic data generation methods.
SETTING AND PARTICIPANTS: a large anonymized RW dataset including 42,926 patients hospitalized for COVID-19 in Italy during the early phase of the pandemic. The dataset was obtained through record linkage of administrative databases and the COVID-19 registry using the TheShinISS tool.
MAIN OUTCOME MEASURES: synthetic versions of the original dataset were generated using both synthpop and CT-GAN methods. Real and synthetic data were compared using three type of measures: 1. general utility measures (univariate, bivariate, and global), including comparison of variable distributions using plots, descriptive statistics, standardized mean differences (SMD), and Propensity Score Mean Squared Error; 2. specific utility measures, assessing the similarity of associations estimates from univariate and multivariable Cox models by evaluating the overlap of 95% confidence intervals (CI) of hazard ratios (HR), and comparison of Kaplan-Meier curves using the log-rank test; 3. disclosure (re-identification) measures, estimating the risk of identity and attribute disclosure using dedicated metrics (Unique in Original – UiO, replicated Uniques – repU, Disclosive in Original – Dorig, Disclosive in Synthetic Correct Original – DiSCO).
RESULTS: in terms of general utility, results confirmed the superiority of synthpop over CT-GAN, with more than half of the variables exceeding the acceptable SMD threshold. Moreover, synthpop showed higher specific utility than CT-GAN, with a median overlap of 95%CI of HR of 75% (interquartile range, IQR: 66%-95%) compared with 0% (IQR: 0%-6%) for CT-GAN. Concerning disclosure measures, although the original dataset already presented a negligible risk of identity disclosure (UiO=0.23%), making synthesis largely redundant, both methods further reduced this risk (synthpop: repU=0.06%; GAN: repU=0.05%).
CONCLUSIONS: in this study, considering the three evaluated aspects, synthpop performed better in balancing statistical accuracy and privacy protection. It also offered greater methodological transparency based on explicit statistical models and required lower computational time. These findings contribute to the ongoing debate on the potential use of synthetic data in research and regulatory assessments supporting their integration into RW data analysis workflows. The disclosure measures adopted may serve as a practical starting point, however, the definition of shared standards for these measures, as well as acceptable disclosure risk thresholds, would be desirable and should be developed collaboratively by the scientific community and data protection authorities. Future studies should focus on generating synthetic data where non-explicit relationships exist in real data (i.e., relationships that are not directly attributable to additive or multiplicative structures).
PMID:42444462 | DOI:10.19191/EP26.3.A992.059
