Overview

• The Delphi-2M system was trained on 400,000 anonymised UK Biobank records and validated on 1.9 million Danish patient records, demonstrating cross-country performance.
• Forecasts are probabilistic and generally more accurate over shorter time horizons, offering rate-based risk estimates rather than definitive outcomes.
• Performance is strongest for conditions with consistent progression such as certain cancers, heart attacks and sepsis, and weaker for variable conditions including mental health disorders and pregnancy-related complications.
• The model can generate synthetic future health trajectories for up to about 20 years, enabling research uses such as population burden simulations without mapping to real individuals.
• Authors and external experts caution the tool is not ready for clinics, citing demographic biases in training data, privacy requirements and the need for prospective evaluations, with potential deployment estimated in five to ten years.