Synthetic data in rheumatology: a systematic literature review.

OBJECTIVE

To systematically review the methods, validation frameworks, applications and regulatory landscape of synthetic data generation in healthcare, with a specific focus on rheumatology.

METHODS

We searched PubMed/MEDLINE, Embase and arXiv (January 2016-January 2026). From 1099 database records and 7 expert-identified studies, 701 remained after removing 405 duplicates. Title and abstract screening excluded 353 records; full-text assessment of 348 articles excluded 120 (E1: n = 62; E2: n = 1; E3: n = 23; E4: n = 34), yielding 228 studies. Quality assessment of 13 rheumatology-specific studies using a 10-item checklist revealed 1 high-quality, 8 moderate-quality, and 4 low-quality studies, with systematic gaps in privacy evaluation (2/13) and reproducibility (1/13).

RESULTS

The 228 included studies encompassed six generation-method categories (generative adversarial networks, diffusion models, large language models, statistical methods, rule-based simulators and hybrid approaches) and five thematic categories (digital twins, privacy-preserving methods, reviews, validation frameworks and healthcare applications), detailed in the supplementary methods. Validation remained inconsistent, with only a minority evaluating fidelity, utility and privacy simultaneously. In rheumatology, 13 studies spanned rheumatoid arthritis (n = 3), osteoarthritis (n = 3), Sjogren's syndrome (n = 2), systemic sclerosis (n = 2), systemic lupus erythematosus (n = 2) and ankylosing spondylitis (n = 1). Only 2 of 13 formally evaluated privacy and only 1 shared code for reproducibility.

CONCLUSION

Synthetic data hold substantial promise for rheumatology, particularly for augmenting small cohorts in low-prevalence diseases, enabling privacy-preserving registry sharing and simulating clinical trial populations. Realizing this potential requires standardized validation, greater attention to privacy and reproducibility and regulatory clarity on synthetic health data.