Overview

• The Nature-published study from EMBL, DKFZ and the University of Copenhagen trained on roughly 400,000–454,000 UK Biobank records and validated on about two million Danish patient histories.
• The system estimates probabilities and onset timing for more than 1,000 conditions, performing best for clear progression diseases such as certain cancers, heart attacks and septicemia, with AUCs above 0.85 in some evaluations.
• Forecasts are probabilistic and time-bound, offering calibrated, weather-style risk estimates across horizons up to roughly 20 years with diminishing accuracy at longer ranges and for variable conditions like mental health or pregnancy complications.
• Researchers stress the model is not ready for clinical decisions due to demographic skew in training data and the need for prospective validation, population-specific testing and formal oversight.
• Delphi-2M can simulate synthetic patient trajectories to support research and health-system planning while aiming to preserve privacy.