Privacy-preserving synthetic data from a national CIDP registry: validation against real-world evidence

Research in chronic inflammatory demyelinating polyradiculoneuropathy (CIDP) is constrained by low prevalence, limited data sharing, and privacy regulations. Disease registries mitigate some of these barriers, yet the reuse of individual-level data remains challenging. Artificial intelligence- and regression-based synthetic data (SD) generation may offer a privacy-preserving solution, provided analytical validity is maintained. We aimed to validate SD derived from the Italian CIDP Registry using a published case-control study as a reference framework.

We analyzed a matched case-control dataset including 195 CIDP cases and 195 controls to examine lifestyle and dietary factors associated with CIDP risk. SD were generated using sequential conditional models stratified by sex. Fidelity was assessed by comparing real and SD in terms of marginal distributions, data distinguishability (propensity score mean squared error), and preservation of associations with CIDP risk using univariable and multivariable logistic regression.

SD preserved the original case-control balance and sex distribution. Marginal distributions showed high concordance between real and SD, with minor smoothing in a small number of multi-level dietary variables. Distinguishability analyses indicated low separability between real and synthetic records. Univariable associations with CIDP risk were closely replicated, including the protective effects of higher fish and rice intake. Multivariable models demonstrated highly concordant effect estimates between datasets.

SD derived from the Italian CIDP Registry accurately reproduces key analytical properties of real-world data while preserving patient privacy. This approach enables reproducible and shareable CIDP research and may support future applications, including synthetic standard-of-care comparator cohorts and precision medicine-oriented investigations.