Learning the natural history of human disease with generative transformers.
Level 3 - non-randomized controlled study
Retrospective cohort prognostic model development and external validation study
PubMed 40963019 · doi:10.1038/s41586-025-09529-3
What was done
Investigators adapted a generative pretrained transformer architecture into Delphi-2M to model human disease progression and competing health risks. The model was trained on health records from 0.4 million UK Biobank participants and externally validated on 1.9 million Danish individuals without parameter adjustment. The authors evaluated its ability to predict rates across more than 1,000 diseases, generate synthetic multi-decade health trajectories, and identify temporal co-morbidity clusters using explainable AI methods.
What was found
Delphi-2M predicted the rates of more than 1,000 diseases conditional on prior disease history with accuracy reported as comparable to existing single-disease models. It generated synthetic future health trajectories covering up to 20 years, which supported training secondary AI models. The abstract provides no specific numerical values, error rates, or discrimination metrics.
Why it matters
This study shows that a single generative AI model can capture population-scale, multi-disease trajectories and generalize across different national healthcare systems without site-specific parameter fine-tuning.
Limits
The abstract reports no numerical performance metrics (such as C-statistics, AUC, or calibration measures). Explainable AI methods demonstrated that the model learned biases present in the training datasets. The study was conducted in UK and Danish cohorts, so performance in other demographics or healthcare systems was not assessed, and prospective clinical utility was not tested.
Cited by
- supports A machine learning study utilizing UK Biobank data from 500,000 participants was able to predict approximately 1,000 diseases before clinical onset.