What it is
Delphi-2M adapts the GPT architecture to model how human diseases progress and compete over a lifetime. Trained on records from 400,000 UK Biobank participants and validated without retraining on 1.9 million Danish individuals, it predicts the rates of more than 1,000 diseases from each person's history at an accuracy comparable to dedicated single-disease models. Because it is generative, it can also sample synthetic health trajectories up to 20 years ahead, producing data usable for training other models.
Why it matters
This is a foundation-model approach to whole-body health rather than one predictor per disease, letting a single system reason about co-morbidities and their timing across more than 1,000 conditions. Holding at useful accuracy on an external 1.9 million-person cohort with no parameter changes is strong evidence the patterns generalize beyond the training population. The synthetic trajectories also offer a privacy-preserving substitute for real records, since models trained purely on generated data still learned.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed undergenerative AI, disease prediction, foundation models, UK Biobank, synthetic data
Watch
A short explainer of this result.