Particle.news

Download on the App Store

EMBL’s Delphi-2M AI, Published in Nature, Forecasts Health Risks Across 1,000 Diseases

Trained on UK and Danish records, the model posts a five-year C-index near 0.85, with experts urging cautious, governed deployment.

Overview

  • The transformer model of about two million parameters was trained on 400,000 UK Biobank participants and validated on a 1.9 million-person Danish registry.
  • It models individual and population health trajectories and can project up to roughly 20 years, with short-term predictions for events like heart attack, certain cancers, and mortality matching specialized tools.
  • Performance drops for rare, complex or irregular conditions such as psychiatric disorders and pregnancy complications, and non-representative training data limit generalisability.
  • Ethicists emphasize that outputs are probabilistic rather than fate, underscore a right not to know, and distinguish lower-risk population analytics from higher-risk individual decisioning.
  • Researchers describe ongoing work to add inputs such as lab results or wearable data and experts warn of possible misuse by insurers or employers, while clinical adoption is expected to require years of further validation.