AI Model Turns Health Data Into Disease Forecasts

By Bio-IT World Staff 

October 23, 2025 | Researchers have developed a generative transformer model capable of forecasting an individual’s risk of developing over 1,200 diseases a decade in advance. The model, called Delphi-2M, was validated using data on about 2 million individuals from Denmark’s National Patient Registry without retraining (Nature, DOI: 10.1038/s41586-025-09529-3). It was initially trained on anonymized health, genetic, and imaging data from 400,000 volunteers in the UK Biobank. 

While predictive algorithms exist, Delphi-2M forecasts the full spectrum of human diseases and was taught on the entire International Classification of Diseases, Tenth Revision (ICD-10). The model can map out a patient’s healthcare journey across time or even predict a risk at a certain age for a certain condition. By quantifying the rate at which new cases occur, it provides both individual and population-level risk trajectories, like weather forecasts for health. 

According to Tomas Fitzgerald, senior scientist at the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL-EBI), the model’s predictive accuracy is “similar to if not better” than other currently implemented disease risk scores. Its strength lies in detecting subtle, multivariate relationships between competing morbidities and risks related to smoking and alcohol use, as well as sex and body mass index. 

Delphi-2M also uses “attention maps,” which helps researchers visualize which clinical features the model weighs most heavily in its predictions. For instance, it pays strong attention to whether a patient is male or female, underscoring well-known physiological differences in disease progression between men and women. 

Though Delphi-2M performs well for conditions with consistent progression patterns, such as certain cancers and heart attacks, it must be mentioned that it is less reliable for variable conditions, such as mental health disorders or pregnancy complications. 

Fitzgerald envisions multiple applications for Delphi-2M: early disease detection in clinical settings, population health forecasting, and even resource planning for healthcare systems. Currently, he and his colleagues are collaborating with many other groups interested in the model. 

The team is now working to enhance Delphi-2M by integrating biomarker data, genotypes, and prescription records. “In the models where we’ve integrated biomarker data, overall, they provide a marginal increase [in performance] across most diseases, in particular metabolic conditions,” said Fitzgerald.  

To read the full story written by Deborah Borfitz, please head over to Diagnostics World News.