Scientists train AI to forecast over 1,000 diseases years in advance

What if your lifetime medical chart is just a sequence of words, and your future illnesses are simply the next sentences in a paragraph?

That is the premise behind Delphi-2M, a new AI system developed by researchers from Britain, Denmark, Germany, and Switzerland. Detailed in the journal Nature, the model can forecast diagnosis rates for over 1,000 diseases years in advance.

Its technical foundation will sound familiar: it relies on the transformer architecture, the same core engine behind ChatGPT.

Reading patient charts like grammar

Instead of reading text prompts, Delphi-2M reads medical timelines.

German Cancer Research Center AI expert Moritz Gerstung compared the process to learning grammar. The system learns the patterns in healthcare data, looking at what diagnoses came first, which conditions appear together, and in what exact succession.

By tracking those underlying rules, the model can make meaningful long-range forecasts. Gerstung showed that it could flag individuals at a significantly higher or lower risk of a heart attack than standard factors like age alone would suggest.

The team trained Delphi-2M on UK Biobank records from roughly 500,000 participants, then verified its predictive power against health data from nearly two million people in Denmark.

From one-off calculators to full-body forecasts

Doctors already rely on predictive software. British general practitioners, for instance, routinely use the QRISK3 tool to evaluate heart attack and stroke risks.

The big difference here is breadth. As co-author Ewan Birney pointed out, Delphi-2M can evaluate all diseases simultaneously across extended timelines, rather than tackling a single condition at a time.

Tom Fitzgerald, a co-author from the European Molecular Biology Laboratory, said this approach could eventually help allocate resources across strained healthcare systems.

Still, nobody is putting this into hospitals tomorrow. The researchers stressed that Delphi-2M is not ready for clinical care.

Peter Bannister, a fellow at Britain's Institution of Engineering and Technology, pointed out that both the British and Danish training cohorts carry inherent biases in age, ethnicity, and historical outcomes.

King's College London professor Gustavo Sudre took a more encouraging view, describing the work as a meaningful step toward interpretable and ethically responsible predictive AI.

Why it matters

We think the real breakthrough here is conceptual. Instead of building narrow, custom statistical models for heart disease, diabetes, or cancer individually, the team treated an entire human health trajectory as a continuous language problem.

For everyday patients, this is not something you can use right now, nor should you look for a way to test it. A long-range risk probability without a concrete clinical intervention is more likely to cause panic than improve your health.

For the health tech industry, this puts legacy diagnostic calculators on notice. If a single foundational model can parse raw medical records and generate an entire predictive risk profile, single-purpose risk calculators will start looking like relics.

The clear risk, however, is generalizability. Because Delphi-2M was trained entirely on European populations in Britain and Denmark, we expect its accuracy to drop significantly in diverse demographic and socioeconomic contexts. Proving that it works safely outside those two databases is the real test.