Overview

• Delphi-2M was trained on anonymised records from 400,000 UK Biobank participants and evaluated on 1.9 million patients in Denmark, demonstrating cross‑dataset performance.
• The generative model estimates when diseases may occur and can sample synthetic long‑term health trajectories with meaningful projections up to 20 years.
• Accuracy was stronger for conditions with clearer progression such as some cancers and heart attacks, with weaker reliability for variable disorders including mental health issues and pregnancy complications.
• Researchers describe potential use by GPs within five to ten years, contingent on extensive validation, regulatory approval, and integration into clinical workflows.
• The system learns from sequences of medical events plus lifestyle factors like obesity, smoking, and alcohol use, with experts cautioning about ethics, privacy, bias, and possible psychological harms such as fatalism.