Scientists announced on Wednesday that they have developed a new AI model that can predict medical diagnoses years before they happen. The system is built on the same kind of technology that powers chatbots like ChatGPT.
The AI, named Delphi-2M, can forecast the likelihood of more than 1,000 diseases by analyzing a patient’s medical history. The research team from the UK, Denmark, Germany, and Switzerland published their findings in the journal Nature.
To build the model, researchers used data from the UK Biobank, a massive biomedical database with information on about half a million people.
Delphi-2M is based on “transformer” neural networks — the same architecture behind ChatGPT. These models are usually used for language tasks, but scientists found that medical data works in a similar way.
Moritz Gerstung, an AI expert at the German Cancer Research Center, explained that reading patterns in medical diagnoses is “like learning grammar in a text.” The AI learns how health conditions appear in sequence and combination, making its predictions more accurate.
Charts presented by Gerstung showed that Delphi-2M could identify patients with much higher or lower risks of heart attacks than what age or other basic factors would normally suggest.
To test its accuracy, the team ran Delphi-2M on medical records from nearly two million people in Denmark’s national health database.
Still, Gerstung and his colleagues stressed that the tool is not ready for hospitals yet. The datasets used in the study had limitations, such as being skewed by age, ethnicity, and healthcare access.
Peter Bannister, a health technology expert from Britain’s Institution of Engineering and Technology, noted that more testing is needed. However, he said the system could eventually help doctors monitor patients earlier and deliver preventive care.
Tom Fitzgerald, a co-author from the European Molecular Biology Laboratory, added that AI like Delphi-2M could also help manage scarce healthcare resources more effectively.
Doctors already use computer tools to measure disease risk, such as the QRISK3 program in the UK, which estimates the chance of heart attack or stroke. But unlike those, Delphi-2M can look at many diseases at once and over longer periods.
Ewan Birney, another co-author, said this makes Delphi-2M a big step forward.
Gustavo Sudre, a professor of medical AI at King’s College London, called the research “an important step towards predictive models that are scalable, clear to interpret, and ethically responsible.”
He stressed the importance of “explainable AI,” since many large AI models are still like black boxes — even their creators don’t fully understand how they work.
