Synthetic health data are artificial records generated by statistical, machine-learning or rule-based methods to reproduce the properties of real patient data. They are increasingly used for research, artificial intelligence development, education and data sharing when access to real data is restricted. However, published studies that generate or evaluate synthetic health data describe their methods and results inconsistently. A systematic review of 134 studies evaluating synthetic tabular health data found no consensus on evaluation methods, inconsistent application of metrics, inadequate reporting of dataset characteristics and limited reproducibility (Nafis et al., 2025). As a result, readers, editors and data-access bodies cannot judge the fidelity, utility and privacy of a synthetic dataset, nor compare results across studies.
No existing reporting guideline fills this gap. A search of the EQUATOR library and of the register of guidelines under development for “synthetic data” (06-October-2026) returned no matching guideline. Existing AI guidelines such as TRIPOD+AI address prediction models rather than the data themselves. Related initiatives differ in nature and scope: a quality assessment framework for synthetic tabular health data (Vallevik et al., 2024) and a scorecard for synthetic medical data proposed by FDA scientists (Zamzmi et al., 2025) are single-group proposals developed without a formal consensus process, while an expert consensus on privacy metrics for synthetic data (Pilgram et al., 2025) covers a single evaluation domain and recommends metrics rather than reporting items. These will inform the initial list of candidate items for DOSSIER.
DOSSIER will comprise a checklist and an accompanying explanation and elaboration document. It will specify the minimum information that a research article generating or evaluating a synthetic health dataset should report: the source data, the generation method, the intended use, and the evidence produced on fidelity, utility and privacy, including the methods, comparators and results. DOSSIER will require this evidence to be reported, but will neither set numerical thresholds nor prescribe specific metrics, because published guidance has not yet established acceptable levels. It will apply to health data of any modality.
Predictive models trained on synthetic data are out of scope and remain covered by TRIPOD+AI. Studies that only reuse a synthetic dataset produced elsewhere are expected to cite the report of that dataset and to report their own analysis with the guideline that fits their design. Where a study both generates and uses synthetic data, DOSSIER will apply to the generation and evaluation components.
The primary intended users are authors, journal editors and peer reviewers; secondary users include data-access bodies and data-protection officers who rely on such reports.
Registered: 11 November 2026
Last updated:
Contact: Miguel Rujas Email: [email protected]