Particle.news
Download on the App Store

AI Model Predicts Risk for 1,000+ Diseases Decades Ahead After Cross-National Test

A Nature study details explainable, multi-disease forecasting with cautions on bias, privacy controls, and deployment readiness.

Overview

  • Delphi-2M was trained on 402,799 UK Biobank records and validated on about 1.9–2.0 million Danish records without retraining.
  • The model reported roughly 76% accuracy for near-term predictions and about 70% at 10 years, outperforming many single-disease tools on several tasks.
  • It represents age-stamped ICD diagnoses as sequential tokens in a modified transformer that forecasts both likely next conditions and timing.
  • It can generate synthetic health trajectories and datasets that preserved key relationships, with a secondary model trained on synthetic data reaching 74% accuracy.
  • Researchers highlight limitations including declining long-horizon performance, weaker results for conditions like Type 2 diabetes, dataset bias in UK Biobank, and restricted access to trained weights despite public training code.