Why look at many diseases at once?
Doctors already use risk scores, for example for heart disease. But each score looks at a single condition. As people get older, many live with several illnesses at the same time, and one condition can lead to another.
Mapping these chains is hard: there are thousands of possible diagnoses, and the order in which they appear matters. So the researchers wanted one tool that could follow a person's whole health story.
What did the researchers do?
The team reused the design of large language models, the AI systems that guess the next word in a sentence. Here, the "words" are diagnoses coded with ICD-10 (the international list of diseases), plus sex, body mass, smoking, alcohol and death.
The model, called Delphi-2M, learned from the records of 400,000 volunteers in UK Biobank, a large British research cohort. It was then tested, with no changes, on health registries covering 1.9 million people in Denmark. Accuracy was measured with the AUC, a score that runs from 0.5 (no better than chance) to 1 (perfect).
What were the results?
On the UK validation data, Delphi-2M reached an average score of about 0.76 across all diseases, with accuracy comparable to tools built for a single disease. It performed close to the risk scores used in clinics for cardiovascular disease and dementia. For diabetes, however, a single blood marker (HbA1c) did better.
On the Danish data, the average score fell to 0.67, without any retraining. The model can also generate made-up health journeys over 20 years, which are useful for research, for instance to train other AI systems without using real people's data.
Key figures
- more than 1,000diseases whose risk Delphi-2M estimates from each person's health history
- 400,000UK participants used for training, then 1.9 million people in Denmark for testing
- 0.76average accuracy score (AUC) on the UK validation data
- 0.67average score (AUC) on the Danish data, with no retuning
Your health story, kept clear with Mima
This study makes one point clearly: the order in which health problems appear matters. It is not a test you will be offered tomorrow, but your doctor already relies on your history to decide what to keep an eye on.
With Mima, you can talk through what you are going through day by day and keep track of your appointments and prescriptions, so you arrive with a clear story. A question to ask: "Given my history, is there anything we should watch more closely?"
What could this mean for patients?
For now, nothing direct: Delphi-2M is not used in GP surgeries or hospitals. In future, tools like this might help identify people who need closer follow-up, or help health services plan for the needs of a population.
The authors see such models as additions to current tools, not replacements. The judgment of a doctor who knows the patient remains essential.
What are the limits?
UK Biobank volunteers are healthier than average, and the model reproduces that bias. It also picked up quirks linked to how these data were collected. Its estimates differ by ethnic background and level of deprivation, which calls for care to avoid widening inequalities.
Finally, the authors warn that a link in time between two diseases does not prove that one causes the other. More studies, especially in real clinical settings, are needed before any use in care.
Frequently asked questions
Can Delphi-2M tell me which diseases I will get?
No. It estimates risks across groups of people, with imperfect accuracy. It is not a test available to patients.
What data was Delphi-2M trained on?
On the records of 400,000 UK Biobank participants. It was then tested on 1.9 million people in Denmark.
Why does the AI work less well in Denmark?
Health data are collected differently from one country to another. The average score drops from 0.76 in the UK to 0.67 in Denmark.
Does this AI replace doctors?
No. The authors see it as an addition to existing tools, and the doctor is still the one who interprets and decides.
The original study
Shmatko A, Jung AW, Gaurav K, et al. “Learning the natural history of human disease with generative transformers” Nature, September 17, 2025. DOI : 10.1038/s41586-025-09529-3
Read the full study in NatureThis article summarizes a scientific study and is not medical advice.