Particle.news
Download on the App Store

Delphi-2M AI Model Published in Nature Forecasts Risk for 1,000+ Diseases

Experts say the probabilistic system needs extensive validation before any clinical use.

Overview

  • Developed by EMBL with DKFZ and the University of Copenhagen, the transformer-based model (~2 million parameters) models health trajectories up to one or two decades ahead.
  • Training used records from 400,000 UK Biobank participants and external validation drew on about 1.9 million Danish patient records.
  • Reported performance matched or exceeded specialist models for some conditions, with a 5‑year C‑Index around 0.85 and an average AUC near 0.76.
  • Accuracy was stronger for diseases with clear progression such as heart attacks, certain cancers and mortality, but weaker for infections, mental health, pregnancy complications and rare diseases.
  • The model provides probabilistic risk estimates and raises ethical concerns over bias and potential misuse by insurers or employers, with researchers pointing to population-level planning as a nearer-term application and to added data sources like blood tests and wearables for improvement.