Scientists from Britain, Denmark, Germany and Switzerland have developed an artificial intelligence model capable of predicting medical diagnoses years ahead, using the same “transformer” technology behind chatbots like ChatGPT.
The model, called Delphi-2M, can forecast the likelihood of more than 1,000 diseases based on a patient’s medical history, according to findings published Wednesday in Nature.
Researchers trained Delphi-2M on the UK Biobank, which contains detailed biomedical data from around half a million people. They then tested its accuracy on records from nearly two million patients in Denmark’s public health database.
“Understanding a sequence of medical diagnoses is a bit like learning grammar in a text,” said Moritz Gerstung of the German Cancer Research Center. The AI, he explained, identifies how conditions emerge and combine, allowing for “very meaningful and health-relevant predictions”.
Charts presented by the team showed Delphi-2M could identify individuals at much higher or lower risk of a heart attack than standard assessments based solely on age and other risk factors.
However, the researchers cautioned that the system is not yet ready for clinical use. Both the British and Danish datasets are limited by demographic and healthcare biases.
“This is still a long way from improved healthcare,” said Peter Bannister, a health technology fellow at Britain’s Institution of Engineering and Technology.
If validated further, Delphi-2M could guide earlier monitoring and preventative treatments, while also helping strained health systems allocate resources more effectively, said co-author Tom Fitzgerald of the European Molecular Biology Laboratory.
Unlike current tools such as QRISK3, which predicts risk for specific conditions like heart attack or stroke, Delphi-2M can analyse risk for hundreds of diseases simultaneously over long timeframes.
King’s College London professor Gustavo Sudre called the research “a significant step towards scalable, interpretable and — most importantly — ethically responsible predictive modelling”.