Scientists announced Wednesday that they have developed an advanced artificial intelligence model capable of predicting the likelihood of more than 1,000 medical conditions years before diagnosis, a breakthrough built on the same technology that powers consumer chatbots like ChatGPT.
Known as Delphi-2M, the AI system uses a patient’s medical history to forecast long-term health outcomes, according to a paper published in the journal Nature by a team of researchers from institutions in the UK, Denmark, Germany, and Switzerland.
Trained on data from the UK Biobank, a biomedical database containing health information from around half a million participants, Delphi-2M uses a neural network based on the transformer architecture, the same core technology behind large language models like ChatGPT.
“Understanding a sequence of medical diagnoses is a bit like learning grammar in a text,” said Moritz Gerstung, an AI expert at the German Cancer Research Center. “Delphi-2M learns the patterns in healthcare data — the preceding diagnoses, the combinations they occur in, and the order — allowing for highly meaningful and health-relevant predictions.”
Charts presented by Gerstung showed that Delphi-2M was able to identify individuals at significantly higher or lower risk of experiencing events such as heart attacks — sometimes with greater accuracy than traditional predictors like age or lifestyle factors.
The model’s predictive power was validated using anonymised data from nearly two million individuals in Denmark’s public health system.
However, Gerstung and his colleagues cautioned that the system is not yet ready for clinical use and requires further validation. The datasets used — particularly those from the UK and Denmark — are not fully representative, with biases related to age, ethnicity, and healthcare access.
“This is still a long way from improved healthcare,” noted Peter Bannister, a health technology researcher and fellow at the UK’s Institution of Engineering and Technology. He highlighted that systemic biases in the training data could limit the model’s generalisability.
Despite these limitations, the researchers are optimistic about the long-term potential of Delphi-2M.
“This could eventually guide more personalised monitoring and earlier interventions, effectively moving us toward preventative medicine,” Gerstung said.
Tom Fitzgerald of the European Molecular Biology Laboratory, a co-author of the study, said that at scale, tools like Delphi-2M could also help optimise the allocation of resources in overburdened healthcare systems.
While many doctors already use software to assess disease risk — such as QRISK3, commonly used by UK general practitioners to estimate the likelihood of heart attacks or strokes — Delphi-2M’s strength lies in its breadth and timeline.
“Unlike traditional tools, Delphi-2M can assess risk across a wide range of diseases and over an extended time horizon,” said co-author Ewan Birney.
Experts in the field praised the work as a step forward in medical AI. Gustavo Sudre, a professor at King’s College London specialising in AI for healthcare, described the project as “a significant step toward scalable, interpretable, and — most importantly — ethically responsible predictive modelling.”
The issue of “explainable AI”, ensuring that machine learning systems can be understood and trusted by humans, remains a central challenge in the development of medical AI. Many large models function as “black boxes,” with even their creators unable to fully explain their internal decision-making processes.
Still, Delphi-2M represents a promising advance at the intersection of machine learning and medicine, raising hopes for more proactive and precise healthcare in the years to come.
AFP
