Overview

• Trained on roughly 402,000 UK Biobank records and tested on about 2 million Danish records, the model reported aggregate performance around 0.7 AUC.
• Delphi-2M predicts a person’s next diagnosis and estimated timing up to 20 years, using prior disease history plus factors such as sex, BMI, smoking and alcohol use.
• The system can simulate synthetic patient trajectories that preserve statistical patterns, and the team released code and data assets with relatively modest compute requirements.
• Researchers and reporters emphasize this is a research demonstration, noting limits from UK Biobank’s demographic skew, transferability concerns across health systems, and privacy risk management needs.
• Potential applications discussed include targeted prevention and screening, clinical decision support, trial design, insurance modeling, and health-system planning, contingent on further validation and governance.