Overview

• The model was trained on anonymized records from 400,000 UK Biobank participants and evaluated on about 1.9 million patients in Denmark’s national registry.
• It estimates time-based probabilities for future conditions across 1,258 diseases, performing best for illnesses with consistent progression such as certain cancers, heart attacks and septicemia.
• Accuracy matched or exceeded many single-disease tools and in some comparisons surpassed biomarker-based algorithms, with shorter-range forecasts proving more reliable than longer-term ones.
• Its generative design can simulate synthetic future health trajectories for research and planning, though experts warn of data biases and re-identification risks that require secure processing environments.
• Authors and independent commentators say the system is not ready for clinics and call for diversified validations, explainability studies and randomized trials, with potential deployment estimated in roughly 5–10 years.