News · 27 March 2025
The European Health Data Space: what secondary use means for AI developers
The European Health Data Space (EHDS) Regulation entered into force on 26 March 2025. It creates a common framework for how electronic health data is accessed and shared across the EU, and most of its provisions apply in stages over the following years.
Two purposes
The Regulation separates primary use, which is the use of data to deliver care to the person it concerns, from secondary use, which means using data for research, innovation, policymaking, regulatory activity and, importantly for us, the training and testing of AI systems.
For primary use, individuals gain stronger rights to access and share their own electronic health records across borders and to restrict some access. For secondary use, designated health data access bodies in each member state will process requests, grant permits and provide data in secure processing environments.
What changes for AI teams
Today, access to clinical data for model development is negotiated hospital by hospital, project by project. The EHDS aims to make that more consistent: a defined application route, conditions of use and standardised technical environments. A few implications follow.
- Data will often be analysed where it lives. Secure processing environments mean code goes to the data. Pipelines must be portable, reproducible and documented.
- Metadata matters. Datasets will be described in catalogues. Quality labels and documentation of provenance will help researchers decide whether data is fit for purpose.
- Purpose limitation is explicit. Permits will specify what you may do; unrelated reuse will not be allowed.
- Opt-outs and exclusions exist. Representativeness, which is already a challenge for clinical AI, will need conscious attention.
Why this is good news, with a caveat
Fragmented data access is a major reason medical AI is validated on narrow populations. A consistent framework can improve generalisability, and it favours teams who already document datasets and validation carefully.
The caveat is time: the secondary-use machinery will phase in over several years, and national implementation will differ. Teams should not wait. Building dataset documentation, privacy-preserving workflows and external validation habits now will make future permit applications simpler.
Written in March 2025; details reflect the situation at that time.