Synthetic health data is artificially generated patient data that preserves the statistical properties of real clinical records without containing any actual patient's information. It retains the distributions, correlations and relational structures that analysts care about. A synthetic dataset looks and behaves like real clinical data: it has patients, diagnoses, medications, visits, and lab values in realistic proportions. But no row corresponds to a real person, because every record was generated by a model rather than collected from a human being.
That distinction, generated, not collected, is why synthetic data has become a central idea in health-data privacy. Real patient data is legally radioactive: it can rarely leave the institution that holds it, and every copy multiplies breach risk. Synthetic data, done properly, carries the statistical value across the boundary while the real records stay where they are.
How is synthetic health data generated?
A generative model is trained on (or fitted to statistics of) real clinical data, such as electronic health records, insurance claims and registries, and learns its statistical structure: how age relates to diagnoses, how diagnoses co-occur, how utilisation is distributed. The model then generates entirely new records that follow those learned patterns. Inputs commonly arrive in healthcare-native formats such as HL7 FHIR R4 or the OMOP Common Data Model, and outputs are produced in the same schemas so downstream tools work unchanged.
The critical caveat: generation alone is not a privacy guarantee. A generative model can memorise unusual records and reproduce something close to them. This is why the serious end of the field pairs synthesis with differential privacy, a mathematical ceiling on how much any single patient can influence the output. Without a formal guarantee, "synthetic" is a description of how the data was made, not of how private it is.
What synthetic data is not
- It is not masked or de-identified data. De-identification takes real records and removes or blurs identifiers; the records are still real people's data, and re-identification research keeps showing how recoverable they are. Synthetic records were never real to begin with. (Full comparison: Synthetic Data vs. De-Identification.)
- It is not "fake data" in the QA sense. Faker-style generators produce random plausible-looking values with no statistical relationship to any real population. Synthetic health data is fitted to a real cohort and is statistically faithful to it.
- It is not automatically anonymous in law. Regulators, including the EDPB in its 2026 draft anonymisation guidelines, treat synthetic data as anonymous only where identification of individuals is effectively precluded, which depends on how the synthesis was done. (Details: Is Synthetic Health Data Anonymous Under GDPR?)
What is synthetic health data good for?
Honesty about fit matters more in this field than in most, because the failure mode, trusting a synthetic dataset for a task it cannot support, happens silently. Based on studies comparing analyses on real versus synthetic cohorts, such as El Emam and colleagues' 2024 replicability study, and on reviews such as Giuffrè and Shung (2023):
| Use case | Suitability |
|---|---|
| Software testing, development and QA environments, demo data | Excellent. Realistic, linkable, zero PHI in lower environments |
| Clinical trial feasibility and cohort sizing | Good to excellent. Strongest for common eligibility criteria, weaker when eligibility depends on rare codes or extreme lab values |
| Machine-learning pre-training and AI benchmarking | Good. Fine-tune on real data where possible |
| Population epidemiology | Moderate. Hypothesis generation, not confirmation |
| Pharmacovigilance / safety-signal detection | Poor. Rare signals live in distribution tails that privacy protection deliberately smooths |
| Rare-disease cohort analysis | Unsuitable. Small cohorts and strong privacy are mathematically in tension |
| Regulatory submissions as primary evidence | Restricted. Secondary or supporting evidence only |
A vendor who claims synthetic data is good for everything is telling you something about the vendor. At Synthetix Health we print contraindications on every dataset certificate for exactly this reason.
What makes a synthetic dataset trustworthy?
Three properties, each of which should be evidenced rather than asserted:
- A formal privacy guarantee. ε-differential privacy provides a provable bound on what the output reveals about any individual, one that holds against any attacker with any auxiliary data. The parameters (ε, δ) should be stated on the dataset, not hidden.
- Measured fidelity. Distribution comparisons, correlation preservation and, for relational data like claims, referential integrity checks (does every claim resolve to a patient and a visit?), reported against the source.
- Provenance and auditability. A record of how the dataset was produced: source characteristics, parameters, privacy budget consumed, and a tamper-evident audit trail a compliance team can file.
Why synthetic health data matters now
Two forces converged. First, healthcare AI needs vastly more training and testing data than privacy law lets institutions share, and demand for realistic clinical data is growing faster than any consent process can. Second, the legal ground under de-identification, the traditional alternative, is eroding, as the EDPB's 2026 draft anonymisation guidelines set identification-risk tests that policy-based masking struggles to meet. The combination makes provably private synthesis the natural successor. The value of the data moves, and the risk does not.
Frequently asked questions
Does synthetic health data contain real patient information?
No. Every record is generated by a model rather than taken from a person. But the model was trained on real data, so without a formal privacy guarantee such as differential privacy, a model can memorise and partially reproduce unusual real records. Trustworthy synthetic data states its privacy guarantee explicitly.
Is synthetic health data legal to share?
Generally yes, when generated with a demonstrable privacy guarantee: data that cannot be linked to identifiable individuals falls, where identification is effectively precluded, outside the GDPR's definition of personal data and outside HIPAA's definition of PHI. The legal analysis depends on how the data was generated, which is why certification and documented parameters matter.
Can synthetic data be re-identified?
Records generated under ε-differential privacy carry a mathematical bound on what any output can reveal about any individual in the source data, a guarantee that holds regardless of attacker resources or auxiliary datasets. Synthetic data generated without such a guarantee has no such protection, and memorisation-based leakage is a documented risk.
What's the difference between synthetic data and anonymised data?
Anonymised (de-identified) data is real patient data with identifiers removed or generalised. The records still belong to real people. Synthetic data consists of new, artificial records fitted to the statistics of the real cohort. De-identification subtracts information and hopes enough was removed; formally private synthesis provides a provable ceiling on disclosure.
What formats does synthetic health data come in?
The same formats as real clinical data: HL7 FHIR R4 resources, OMOP Common Data Model tables, flat CSV/Parquet extracts, or claims-shaped relational tables, so existing analytics pipelines, ETL, and tools run without modification.
- EDPB Draft Guidelines 02/2026 on Anonymisation (July 2026), public consultation open to 30 Oct 2026
- CMS DE-SynPUF, synthetic Medicare claims public use files
- HHS guidance on HIPAA de-identification (Safe Harbor and Expert Determination)
- Synthetix Health published patent application 202641026065 · PCT/IN2026/051623
- El Emam K, et al. An evaluation of the replicability of analyses using synthetic health data. Sci Rep. 2024;14:6978. doi:10.1038/s41598-024-57207-7
- Giuffrè M, Shung DL. Harnessing the power of synthetic data in healthcare. npj Digit Med. 2023;6:186. doi:10.1038/s41746-023-00927-3
Continue reading in Research:
Go deeper
For how to judge a synthetic dataset against a specific task, read Fit for which purpose?, and for the release criteria a data access body could apply, read What makes a release adequate under the EHDS?
Read the analysis