Share AI models not patient data for research

The concept is called federated learning. Rather than hospitals and research institutes sending raw health data to a central location, the data stays local. An AI model is sent to the data, learns from it, and sends back only the updated model parameters. No patient records leave the institution.

The proposal in Science targets the European Health Data Space, an EU initiative to make health data more available for research. The authors argue that the current legal and technical barriers to data sharing are bypassed when models are shared instead of data.

Relevance for longevity research

Research on ageing requires large datasets. Longevity studies follow people for decades, across multiple countries and institutions. The current fragmentation of European health data significantly slows this type of research.

With federated learning, researchers can train AI models on data from thousands of patients in multiple countries, without that data ever leaving a hospital. This is not only safer for privacy; it makes research possible that is currently practically unachievable.

What still needs to be solved

Shared models are not completely watertight. Under certain conditions, they can leak information about the data they were trained on. The authors acknowledge this and call for additional technical safeguards.

The approach also requires technical standardisation. If each hospital structures its data differently, models struggle to learn from the combination. Interoperability is therefore an infrastructure problem that better modelling alone cannot fix.

For the longevity field, this is nonetheless a relevant policy direction. Better access to European health data would accelerate both basic and clinical ageing research.

Read the original article