An artificial intelligence system called Delphi-2M is pushing the boundaries of predictive medicine, estimating a person’s risk of more than 1,000 diseases up to two decades before symptoms appear. By analyzing patterns in medical records, Delphi-2M forecasts which conditions are most likely to emerge and when.
“If our model says it’s a one-in-10 risk for the next year, it really does seem like it turns out to be one in 10,” said Prof. Ewan Birney, interim director of the European Molecular Biology Laboratory, told the BBC.
Training the AI model on massive health data
Developed by researchers at the EMBL, the German Cancer Research Centre, and the University of Copenhagen, Delphi-2M was trained on anonymised health data from approximately 400,000 UK residents, which included hospital admissions, GP visits, and lifestyle factors such as smoking and alcohol consumption.
It was then tested on 1.9 million people in Denmark, showing results that matched or outperformed existing risk models for conditions such as type-2 diabetes, heart attacks, and sepsis.
“Only a small performance drop was observed when applied to data from Danish disease registries, demonstrating that models are—even without additional finetuning—largely applicable across national healthcare systems,” the researchers wrote in Nature.
Delphi-2M is about 76% accurate in predicting a person’s next likely health problem, and it maintains around 70% accuracy even when looking 10 years into the future. The research also found that, when trained on synthetic data, the AI model worked well enough that it could be viable for privacy-sensitive applications.
Clinicians could spot high-risk patients early enough to intervene with targeted treatment or lifestyle advice. Public health services could forecast demand, planning resources for projected increases in specific conditions years ahead. Screening programs could also be reshaped, with AI guiding who is most likely to benefit.
What this AI model could mean for medicine
Delphi-2M does have limitations. The AI model performs best on diseases with clear progression and struggles with more random events. Bias is another concern: UK Biobank data largely reflects people aged 40 to 70, not the whole population.
Other sources include healthy volunteer bias, differences in how diseases are recorded (for example, a GP diagnosis versus a self-report), and performance dips in certain demographic subgroups. Researchers are now working to broaden the model by including genetics, blood analysis, and imaging data, according to the BBC.
AI’s potential in healthcare has proven to be one of the most exciting aspects of the technology’s boom. Last year, one study found that ChatGPT outperformed doctors when assessing medical case histories, even when they were given access to AI tools.
Just this month, it was reported that half of all stroke patients in the UK are now expected to make a full recovery since rolling out an AI tool across the country that helps doctors decide the best course of treatment.
Microsoft’s MAI-DxO medical AI achieves over four times the accuracy of human doctors.