A generative transformer modelling the course of more than 1,000 diseases (Delphi-2M)
A GPT-style model trained on 0.4 million UK Biobank participants predicts each person's rates of more than 1,000 diseases from their past diagnoses, with an average age-stratified AUC of about 0.76 and accuracy comparable to single-disease models, and held up on 1.9 million Danish registry records with no retraining (AUC 0.67 against 0.69 on UK longitudinal testing).
- Model
- Delphi-2M (modified GPT-2 architecture)
- Field
- Medicine
- Date
- 2025-09-17
- Human collaborators
- Artem Shmatko, Ewan Birney, Tom Fitzgerald, Moritz Gerstung
What was found
Each health history is a sequence of top-level ICD-10 diagnoses plus sex, body mass, smoking, alcohol and death, each at the age it was recorded, and the architecture is changed from GPT-2 to model continuous time and competing risks. Death is the most confidently predicted event, with an AUC of 0.97, and accuracy falls from 0.76 to 0.70 for predictions ten years ahead. Because it is generative, the model can sample synthetic future health trajectories up to 20 years ahead, which the authors use to estimate disease burden and to train other models that never see real records. Explainability analyses recover clusters of co-occurring diseases and their effects over time, and also show biases inherited from UK Biobank. Code is public. A correction was published in November 2025.
Novelty check
Europe PMC, titles (transformer OR generative OR foundation model) AND (health records OR medical events OR disease trajectories OR patient timelines), to 16 September 2025: transformer models of health records predate it, including BEHRT (2020), Foresight (Lancet Digital Health, 2024) and ETHOS (npj Digital Medicine, 2024), and Epic's much larger CoMET preprint appeared a month earlier. Delphi-2M's distinct parts are joint modelling of more than 1,000 diseases with competing risks in continuous time and a no-retraining test on a national registry.
Caveats and known objections
Retrospective, and not used in care. UK Biobank volunteers are older, healthier and less diverse than the UK population, and the authors show the model inherits those biases. The Danish test was run by co-authors at the University of Copenhagen and Statistics Denmark, so it is external data but not an independent check. Accuracy comparable to single-disease models is an average over many diseases, some of which are predicted poorly.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2025). A generative transformer modelling the course of more than 1,000 diseases (Delphi-2M). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2025-09-17-delphi-2m-disease-trajectoriesBibTeX
@misc{whataifound-germancancerresearchcenter-2025-trajectories,
title = {A generative transformer modelling the course of more than 1,000 diseases (Delphi-2M)},
author = {{whataifound.org}},
year = {2025},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by German Cancer Research Center (DKFZ) / EMBL-EBI / University of Copenhagen. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2025-09-17-delphi-2m-disease-trajectories}
}