Synthetic Data for Medical Software Testing
Holding
Synthetic data can strengthen medical software testing, but it cannot replace clinical validation and should not be used to hide weak provenance or weak performance evidence.
Authority
Medical device verification and validation expectations require evidence that software performs as intended. AI risk management requires attention to validity, reliability, privacy, and misuse. Synthetic data is a tool inside that framework, not a substitute for it.
Issue
The failure mode is synthetic confidence. Generated records may cover edge cases, but they may also encode assumptions that make the software look robust under artificial conditions while failing in real clinical variation.
Resolution
Use synthetic data for unit tests, integration tests, privacy-preserving demos, rare workflow simulation, and adversarial cases. Label it clearly, version the generator, document assumptions, and keep clinical validation separate from synthetic test coverage.
Evidence Package
The record should include generator specifications, seed and parameter records, coverage rationale, privacy review, test results, known limitations, and a mapping between synthetic tests and real validation evidence. Synthetic data should expand the test surface without inflating the clinical claim.