SDAC: assessing synthetic health data
SEARCH is a European project that builds the means to do health research with synthetic data, so that patient records do not have to move between institutions. I work on the part about trust: does a synthetic dataset protect the people behind the real data, and is it still good enough to use?
- Project
- SEARCH, the Synthetic Healthcare Data Governance Hub
- Funding
- Innovative Health Initiative Joint Undertaking, grant agreement No. 101172997
- Role
- Senior AI engineer
- Size
- A consortium of 26 institutions
The problem
The questions are not what holds health research back. Access is. Records are sensitive, they are spread over many countries, and each institution keeps its own. So the default answer to a request is no, and research tends to use the few datasets that are already open.
Synthetic data is the obvious solution. You generate records that have the statistical structure of the real ones and belong to nobody.
But “synthetic” is a claim, not a guarantee. There are two ways to fail:
- The generator memorises. Its records look synthetic, and some of them are real patients.
- The generator is too careful. Its records are safe, and useless for analysis.
From the outside, both look correct. A dataset can pass on privacy and fail on usefulness, or the opposite.
So the first question for SEARCH is not how to generate synthetic health data. There are many methods. The first question is how a person who receives a dataset can know whether to rely on it, and how much checking is in proportion to its use.
What I work on
SDAC is the assessment side of that question: the criteria and the method that are used to evaluate a synthetic dataset before anyone trusts it, for each type of data in the project.
The project documents on this subject are not public. This page describes the problem, not the framework.
How it connects to my own research
This work has a pair. My model for relational health records generates synthetic patients. SDAC asks how you can decide whether to trust records like those.