Imagine checking a health forecast the way you check the weather: sunny today, 70% chance of diabetes in ten years. That’s essentially what a new artificial intelligence model, dubbed Delphi-2M, is aiming to deliver. Built by European scientists and trained on vast troves of anonymized medical records, the system can estimate a person’s likelihood of developing more than a thousand different conditions well into the future.
Delphi-2M works a lot like the tech behind popular chatbots. Where chatbots learn the patterns of language to predict the next word, Delphi learns the patterns in clinical histories—hospital admissions, GP notes, diagnoses, medications, even lifestyle factors like smoking—to predict what tends to happen next in a person’s medical journey and roughly when. It doesn’t spit out a calendar date for a heart attack, but it will tell you the probability that one occurs in the next year or five.
“Just like weather, where we could have a 70% chance of rain, we can do that for healthcare,” said Prof Ewan Birney, interim executive director of the European Molecular Biology Laboratory. “And we can do that not just for one disease, but all diseases at the same time.”
The team started by training Delphi-2M on data from more than 400,000 volunteers in UK Biobank, the long-running research project that links medical records with lifestyle information. Then they stress-tested it on fresh ground, comparing its forecasts to outcomes in another set of Biobank participants and, crucially, in 1.9 million real-world patient histories from Denmark. The model held up.
“If our model says it’s a one-in-ten risk for the next year, it really does seem like it turns out to be one in ten,” Birney said.
It proved strongest on conditions that follow well-trodden biological paths—think type 2 diabetes, heart attacks, and sepsis—rather than more random, externally driven events like many infections or rare congenital disorders.
The implications run in two directions at once. For individuals, this kind of long-range risk map could flag who might benefit from early interventions—medication akin to how statins are prescribed off a calculated heart-risk score, or targeted lifestyle changes such as cutting back alcohol for those trending toward liver disease. For health systems, it opens a new planning horizon. If you can credibly estimate how many heart attacks a city like Norwich will have in 2030, you can staff cath labs, stock critical drugs, and shape screening programs years before demand hits.
None of this is ready for your GP’s desktop tomorrow. The study, published in Nature, is still research. The model will need refining, external validation in more populations, and careful regulation before it guides care. The training data skew heavily to people aged 40 to 70, reflecting who joins UK Biobank, so biases are a live concern; the team is already upgrading Delphi to ingest richer inputs like imaging, genetics and blood biomarkers to firm up its predictions and broaden its reach. Prof Moritz Gerstung, who heads the AI in oncology division at Germany’s DKFZ and co-led the work, argued that even in its current form the system has utility at scale: it can help forecast collective healthcare needs across regions, not just risks for a single patient. King’s College London researcher Prof Gustavo Sudre, independent of the project, called the work a significant step toward predictive models that are scalable, interpretable, and ethically responsible.
What makes Delphi-2M striking isn’t only its accuracy; it’s the way familiar clinical breadcrumbs suddenly add up to a map of the road ahead. Birney said he was “pleasantly surprised” by how far the model got using ordinary records alone, noting that its performance matched or beat some tools that already layer in genomics and proteomics. That’s partly why the researchers liken this moment to the early days of genetics in medicine: a decade can pass between scientific confidence and routine clinical use, but the direction of travel is clear.
The crossover with everyday AI is impossible to miss. Delphi’s architecture borrows from the transformer tech that powers modern chatbots, which are built to capture long-range patterns and relationships. Swapped into the clinic, that pattern-spotting talent translates into timelines of risk rather than paragraphs of prose. Because the model is generative, it can simulate plausible futures from the sequence of your past healthcare events, much like a weather model spins up many possible storms to see which path looks most likely.
There are caveats worth repeating. A forecast is not fate; a 30% chance of a condition means seven in ten people like you will not get it. Communication will matter as much as calibration, or else risk numbers could spook people or, worse, entrench inequities if access to prevention isn’t fair. Privacy and consent are paramount, even with anonymized data. And while the system handled the leap from the UK to Denmark, it will need to be tested in more diverse populations before anyone can claim it’s universal.
Still, the promise is hard to ignore. If clinicians can see around corners, they can act sooner. If hospitals can predict the coming tide, they can staff and stock accordingly. And if public health can identify the neighborhoods most likely to shoulder tomorrow’s disease burden, they can push resources upstream.
“This is the beginning of a new way to understand human health and disease progression,” Gerstung said.
Birney’s view is equally pragmatic: the technology to make these predictions has arrived, now the painstaking work of testing and regulation begins.
Delphi-2M emerged from a collaboration between the European Molecular Biology Laboratory, the German Cancer Research Centre, and the University of Copenhagen. The team holds patents on some of the model’s key ideas and is exploring commercialization with their institutions. But Birney’s takeaway is simpler. Don’t underestimate what’s already in the clinical file. In his words, “the straightforward medical record” turns out to be a surprisingly powerful crystal ball.
With input from the Guardian, the Financial Times, and BBC.
The latest news in your social feeds
Subscribe to our social media platforms to stay tuned