For decades, public health has been reactive. Patients develop symptoms, doctors investigate, hospitals intervene. But what if we could move decisively upstream, identifying risks before illness takes hold and redesigning healthcare around prevention? That is the promise of Delphi-2M, a new artificial intelligence model described this week in Nature.
Developed by scientists at the European Molecular Biology Laboratory and tested in Denmark and the UK, Delphi-2M scans patterns in anonymised health records and estimates an individual’s risk of developing more than 1,000 diseases. It does not predict the exact day someone will have a heart attack or the moment an infection will strike. Instead, it works like a weather forecast. Just as the Met Office can predict a 70% chance of rain, this system can estimate a 70% chance of type 2 diabetes or sepsis within the coming years.
The scale is unprecedented. Traditional risk models focus narrowly: cholesterol levels for heart disease, family history for cancer, blood pressure for stroke. Delphi-2M can make simultaneous assessments across a wide spectrum of conditions, giving a more holistic picture of health and disease progression. Researchers report the strongest accuracy in conditions with clear trajectories, such as cardiovascular disease and diabetes.
What makes this research compelling is not only its technical sophistication, but its potential applications. With early warnings, doctors could recommend lifestyle interventions tailored to individual risk, such as alcohol reduction for those prone to liver disease. Public health teams could design smarter screening programmes. Hospitals could forecast demand years in advance, aligning staff, equipment and budgets with predicted needs. Imagine knowing today how many heart attacks a city like Norwich will face in 2030, and planning accordingly.
Yet the promise comes with caveats. The model was built on UK Biobank data, which largely represents people aged 40 to 70 and is not fully reflective of society. Biases in training data risk amplifying inequalities if not carefully addressed. The system is also far from ready for clinical use. As Professor Ewan Birney, who led the research, has cautioned, this is research, not a finished tool. Regulation, refinement and rigorous testing must come before rollout.
There are also profound ethical questions. How should individuals be told they are at high risk of disease that may never materialise? How do we prevent insurers or employers from misusing such data? The lessons of genomics are instructive. It took a decade for genetic research to move from scientific confidence to safe and routine healthcare practice. Predictive AI will demand the same careful stewardship.
Still, the trajectory is clear. As Professor Moritz Gerstung of the German Cancer Research Centre notes, generative models like Delphi-2M open the door to personalised, anticipatory care at scale. For a health service under pressure, the prospect of shifting resources from crisis response to prevention is transformative.
If used responsibly, this technology could change not only how we treat disease, but how we understand health itself. The future of medicine may look less like the emergency room and more like tomorrow’s weather report.
