Synthetic Healthcare GAN API
Design synthetic patient-record workflows around purpose, utility, and controlled release.
Choose a purpose-limited representation
An interface test may need appointment categories and synthetic identifiers, while a research experiment may need relationships among coded observations. Define the intended task before choosing fields. Minimize the representation rather than copying every variable from a source system.
Document units, missing-value meanings, and relationships that matter for the permitted use. Structured coded records, longitudinal notes, medical images, and time series are different modeling problems. The linked validation guide discusses medGAN research without treating that research as a deployment certification.
Separate generation from release
A proposed job should reference an approved schema and model revision, request a bounded cohort, and return validation results with the synthetic records. Keep direct identifiers out of illustrative payloads and generate surrogate keys outside the model when that is appropriate for the task.
Evaluate privacy and utility on separate tracks. Passing a structural check or producing realistic-looking records does not establish anonymity, clinical suitability, or compliance. Match the release decision to the actual recipient, source data, evaluation evidence, and authorized use.
Make the limitations travel with the dataset
Keep a manifest containing schema version, relevant constraints, accepted row count, generator configuration, and known exclusions. Preserve the synthetic designation in exports so a receiving team does not treat generated records as newly observed patients.
Use domain review for meaningful clinical constraints and obtain appropriate review for consequential use. A dataset intended for a controlled prototype should not silently inherit permission for patient-care decisions or unrestricted publication.
Purpose
A named use, audience, and excluded use cases.
Validation
Meaningful schema and domain checks.
Release
Separate privacy evidence and documented approval.
Questions for this workflow
Is synthetic patient data automatically anonymous?
Do not assume so. Evaluate disclosure risk for the actual source, generator, recipient, and release context, separately from usefulness.
Does a successful dataset test establish clinical efficacy?
No. A software or research test does not establish a system’s suitability for patient care. That use requires its own evidence and qualified review.
Synthetic Healthcare Data: Validate Utility and Privacy Separately
A purpose-limited workflow for synthetic patient records, covering schema design, utility testing, privacy review, and release decisions.
Read the complete guide