How probable is it that someone will suffer a heart attack, develop diabetes or mental health problems in the next 20 years? A new AI model, developed by the European Molecular Biology Laboratory (EMBL) and the German Cancer Research Centre (DKFZ), among others, promises predictions for over 1,000 diseases – based on real-world data.
From individual diseases to the big picture
Previous models have mostly focused on a single disease – such as cardiovascular risk or genetic predispositions for cancer. Delphi-2M thinks bigger: it has been trained with data sets from more than two million people, combines genetic, clinical and demographic information, and uses this to calculate probabilities for a wide variety of clinical pictures. At its core, it is a so-called Generative Pre-trained Transformer (GPT) model – an algorithm that recognises patterns and makes predictions based on them.
Delphi-2M performed remarkably well in tests: for diseases with clear progression patterns, such as certain types of cancer or heart attacks, the model achieved higher accuracy than specialised individual models in some cases. It was more difficult with psychiatric disorders, pregnancy complications or rare conditions, where the progression is more complex and the data available is more scarce.
Opportunities and pitfalls
Experts are impressed – but also cautious. Prof. Robert Ranisch (University of Potsdam) sees Delphi-2M as ‘an impressive example of the potential of generative AI in health research’. But he warns that bias, discrimination and the responsible handling of sensitive health data remain key challenges.
The ethical dimension is also under discussion: no one should be confronted with a personal risk analysis without having given their consent. The right not to know must be preserved, as must data protection.
But it’s not just about predictions. As Carsten Marr from the Helmholtz Centre in Munich points out: ‘What is particularly exciting is the discovery of previously unknown correlations between diseases – connections that we might otherwise never have recognised.’
AI oracle or tool?
The average AUC (area under the curve, a measure of the performance of a classification model) of 0.76 shows that For individual patients, Delphi-2M is not yet a precise oracle. But as a research tool and basis for prevention strategies, it has enormous potential.
Ultimately, Delphi-2M remains what its name promises – a tool that allows us to look into the future, but not a machine of destiny. How we deal with these insights is less a question of technology than one of ethics.
Or, as Ranisch puts it: ‘Such predictions are not verdicts of fate – but valuable pointers for prevention and therapy.’