Synthetic Health Data

Sources and review notes

These links support definition-level research and do not establish the regulatory status, safety, or suitability of any product.

FDA's educational glossary defines synthetic data as data created artificially through statistical modeling or simulation to represent structures and relationships seen in actual patient data without containing real or specific information about individuals. FDA's CDRH regulatory science program is studying both the possibilities and limitations of supplementing representative medical patient datasets with synthetic data for AI development and assessment. NIST SP 800-226 provides guidance for evaluating differential privacy guarantees and includes privacy considerations for synthetic data. These sources do not establish that every synthetic dataset is de-identified, differentially private, representative, clinically valid, unbiased, or suitable as a substitute for independent real-world testing. Teams must document source-data authority and PHI handling, generation method, training and test separation, memorization and disclosure testing, privacy claims and threat models, intended-use utility metrics, subgroup representation, impossible-record detection, independent real-data validation, versioning, provenance, recipient restrictions, and monitoring for leakage, drift, and misuse.

• Source 1: www.fda.gov/science-research/artificial-intelligence-and-medical-products/fda-digital-health-and-artificial-intelligence-glossary-educational-resource
• Source 2: www.fda.gov/medical-devices/medical-device-regulatory-science-research-programs-conducted-osel/addressing-limitations-medical-data-ai
• Source 3: nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-226.pdf