Scholars Train AI to Predict Future Health

Novel Model Uses UK Data to Give Forecasting a Try

Doctors keep meticulous details on their patients’ health. iStock

Artificial intelligence (AI) is expected to help doctors understand disease risk, progression, and treatments. In a September study in Nature, researchers report using an AI model called Delphi-2M to identify patterns of progression in over 1,000 diseases, based on hundreds of thousands of medical records. The AI model was found to reliably track and predict outcomes both in the short term and 20 years out. Study highlights include:

• Delphi-2M was trained on data from about 402,800 individuals in the UK Biobank, validated internally on nearly 100,600 UK individuals, and tested externally on around 1.93 million Danish individuals.  

• The model predicted rates for more than 1,000 diseases and included death as an outcome.

• In the UK internal validation, AI predictions were above average, at about  0.76 in a statistical measure known as “area under curve,” or AUC.

• AUC dropped to roughly 0.70 when examining prediction horizons of 10 years.

• External validation on Danish data showed somewhat lower but correlated performance (average AUC ≈ 0.67).

• Delphi-2M can sample entire future health trajectories based on past health up to a given age (e.g., age 60), and these synthetic trajectories show disease incidence patterns that come close to observed real-world data for ages 70–75. 

• Using SHAP (Shapley Additive Explanations)—a method that aids in understanding how different health factors influenced the final prediction—the study showed how disease diagnoses cluster. For example, cancers raise long-term mortality risk.

• Modeling was limited for older age groups (especially over 80 years) because of lack of data.

• Researchers suggested that advancements in this field could support precision medicine by tailoring screening or diagnostic interventions based on an individual's predicted trajectory. 

Caution: Predictions are probabilistic, not deterministic; multiple future health trajectories are possible for any given individual. Clusters or associations in predictions do not imply causation. 

Source:

• Learning the natural history of human disease with generative transformers. Nature. 17 September 2025.