Overview

• Published in Nature, the work from EMBL, DKFZ and the University of Copenhagen reports cross-country validation of the model in mid-September 2025.
• The transformer-based system learns sequences of diagnoses, risk factors and timing between events to estimate both disease risk and likely onset windows.
• Training drew on roughly 400,000–454,000 UK Biobank records, with retrospective validation on about 1.9 million patients from Denmark.
• Performance was strongest for conditions with consistent progression—such as some cancers, heart attacks and septicemia—and weaker for variable conditions like mental-health disorders or pregnancy complications.
• The model outputs calibrated probabilities, can simulate health trajectories up to around 20 years and generate synthetic patient data, but it shows demographic biases that must be addressed before any clinical use.