Europe is building routes for hospitals to share evidence without simply pooling every patient record. Federated learning and the European Health Data Space put data quality, access and operational responsibility at the centre of clinical AI.
A medical image can be available to an algorithm without being useful for the question it is meant to answer. Researchers also need to know how the image was acquired, what happened to the patient and how the reference diagnosis was established. Those connections often require more work than moving the file. As hospitals expand artificial intelligence, the infrastructure challenge is increasingly about preserving clinical meaning across organisations.
Europe’s response includes the European Health Data Space and infrastructure such as Cancer Image Europe. These initiatives could widen access to evidence and make cross-border evaluation more practical. Their success will depend on the everyday work behind the data: permissions, consistent definitions, quality checks and reliable services that researchers and clinicians can actually use.
The distinction matters commercially. A company may have a sophisticated model but struggle to validate it in a second health system. A hospital may hold valuable records but lack the resources to prepare and maintain them for research. Investment in those capabilities can determine which innovations reach clinical evaluation, and which remain demonstrations on familiar datasets.
The European Health Data Space has a phased timetable
The European Health Data Space Regulation entered into force in March 2025, with major provisions applying in stages. The Commission identifies March 2029 for key requirements, including the first priority categories of cross-border primary-use exchange and most secondary-use rules. A second set of primary-use categories, including medical images, laboratory results and discharge reports, follows in March 2031, alongside additional secondary-use data categories. It is therefore misleading to describe the entire system as already operational.
The regulation distinguishes using information for a person’s care from secondary uses such as research. Its framework includes authorisation arrangements and secure processing environments for permitted secondary uses. It does not make identifiable health records an unrestricted resource for model developers. The applicable legal basis, purpose and safeguards still need to be established.
For hospitals, the preparatory question is concrete: can they describe the records they hold and explain their limitations? A catalogue entry becomes much more useful when it identifies the population, collection period, missing fields and outcome definitions. Without that information, researchers may discover only after lengthy access procedures that a dataset cannot support the proposed study.
Federated learning changes where training happens
Federated learning offers a different arrangement from collecting all training records in a central database. Participating sites train locally and exchange model updates. In the 2021 EXAM study, researchers used information from 20 institutions to predict oxygen requirements among symptomatic patients with COVID-19. The inputs combined chest X-rays with clinical data. The work demonstrated multi-institutional model development without transferring the underlying records to one repository; it was not a trial showing that use of the predictions improved patient outcomes.
A further study published in 2022 brought together data from 6,314 people with glioblastoma across 71 sites on six continents. Its task was identifying tumour boundaries on scans. This illustrates the potential to build a larger evidence base for a relatively uncommon cancer, while keeping the claim tied to image segmentation rather than assuming better surgical outcomes.
Neither example makes collaboration effortless. Each participating hospital still needs appropriate computing capacity, a workable local process and staff who understand the data. The central team must coordinate model versions and investigate failures. Federation redistributes the technical work; it does not remove it. A project budget that funds only the central algorithm risks leaving smaller partners unable to participate consistently.
Local data can still be difficult to compare
An international paediatric brain-tumour study published in 2024 tested a federated platform using MRI data from 19 centres. The researchers described substantial differences between sites, including uneven sample sizes, missing tumour classes, variation in imaging signals and differences in patient characteristics. These are properties of the clinical evidence, rather than problems solved merely by selecting a distributed computing architecture.
Those differences can alter what a model learns. A hospital treating a rare subtype may contribute few examples but supply information that a much larger centre lacks. Conversely, a dataset with many repeated observations may appear large without representing many independent patients. A useful collaboration needs to understand these contributions before deciding how to combine them.
Validation should also preserve the distinction between a new patient and a new setting. Splitting records within the same hospital tests something different from evaluating a model at a site that did not contribute to training. For developers seeking wider deployment, both questions matter. Publishing results by institution and clinically relevant subgroup can reveal where further work is needed without reducing an entire network to one reassuring average.
Interoperability must preserve the meaning of a record
HL7’s Fast Healthcare Interoperability Resources, known as FHIR, provides a standard for exchanging healthcare information electronically. Its structured resources can help applications work with information from different systems. Implementing an exchange standard, however, does not by itself establish that every participating organisation records a clinical event in the same way.
A laboratory result needs its unit and collection time. A medication entry may describe an order, a dispensing event or administration to a patient. An absent diagnosis may mean it was ruled out, never assessed or simply not recorded in that system. An AI pipeline that loses those distinctions can create a plausible-looking dataset with a different clinical meaning from the original records.
The work is therefore partly clinical and partly technical. Data specialists need access to the people who understand how a field is populated and when its interpretation changed. A new laboratory instrument, coding policy or electronic record can affect continuity. Documenting such changes would help investigators distinguish a shift in patient health from a shift in measurement, and help hospitals assess when a model needs re-evaluation.
Shared infrastructure needs a security model
Keeping records at their original institution can reduce some exposure, but it is not a complete privacy guarantee. Research on federated health systems treats federated learning, secure aggregation and differential privacy as distinct techniques that can be combined. Model updates, access arrangements and the final model require their own assessment. The appropriate protections depend on the data and the threat being addressed.
Operational resilience matters as well. ENISA’s health-sector analysis identifies ransomware and data breaches among the threats affecting healthcare. Its incident evidence concerns healthcare infrastructure more broadly; it does not establish that AI caused those attacks. It does show why a clinical AI service must be considered within the hospital’s wider dependencies and recovery arrangements.
A practical evaluation could test what happens when a data connection fails, an update is delayed or a partner withdraws. Staff need to recognise stale information and know whether the service remains usable. Recovery should include checking that the restored model and data mappings match the approved configuration. These are ordinary operational questions, but their consequences become clinical when an algorithm influences care.
Measure the research that the infrastructure enables
The Commission describes Cancer Image Europe as an infrastructure moving towards large-scale, multi-centre validation of AI tools. That direction gives shared data a practical purpose beyond increasing the number of images in a catalogue. A resource becomes more valuable when it enables independent teams to answer clinically relevant questions across settings.
For funders, useful measures would include how long a suitable study takes to start, whether participating sites can reproduce its analysis and how often a validation changes a deployment decision. Hospitals should also be able to see the resources required to remain in the network. Shared infrastructure succeeds when local participation is sustainable, not just when a central platform launches.
This creates an industrial opportunity for services that maintain reliable clinical data, support secure collaboration and make validation reproducible. Their contribution may be less visible than a new model’s headline performance, but it determines whether that performance can be investigated beyond its original setting. Europe’s health data investment will be most consequential when it makes those investigations routine.
