Healthcare & Bio AI
A transformer model predicts rates of more than 1,000 diseases about as accurately as single-disease models
In a 2025 paper published in Nature, a peer-reviewed journal, researchers describe Delphi-2M, a model based on the GPT architecture, a transformer design that learns by predicting the next item in a sequence. The model predicts disease rates from a person's past disease history. The team trained it on records from 0.4 million UK Biobank participants and validated it on records from 1.9 million Danish individuals. Delphi-2M predicted the rates of more than 1,000 diseases with accuracy comparable to existing single-disease models, and the Danish test required no change to its parameters. The model could also sample synthetic future health paths and estimate possible disease burden up to 20 years ahead.
What did the researchers do?
Researchers from the European Bioinformatics Institute and other institutions modified the GPT architecture to model how diseases progress and compete within a person's health history. They trained Delphi-2M on data from 0.4 million UK Biobank participants. They then validated it on external data from 1.9 million Danish individuals, with no change in parameters.
FAQ
How well did the model predict disease?
The abstract reports that Delphi-2M predicted the rates of more than 1,000 diseases from each person's past disease history with accuracy comparable to existing single-disease models. Separately, the authors validated the model on data from 1.9 million Danish individuals without changing its parameters.
What can the generative model do besides prediction?
Because Delphi-2M is generative, it can sample synthetic future health trajectories. The authors report that these samples provide meaningful estimates of possible disease burden for up to 20 years and can help train other AI models that have never seen actual patient data.
What did the explainability analysis show?
The authors used explainable AI methods to examine the model's predictions. They found clusters of co-morbidities, conditions that occur together, within and across disease chapters, the broad disease categories the authors used. They also found time-dependent consequences for future health. The same methods showed biases the model learned from its training data.
Limits and what's next
The abstract describes one model trained on 0.4 million UK Biobank participants and validated on 1.9 million Danish individuals, so the evidence comes from health records in those two settings. The work is a predictive modeling study of disease histories using health records. The explainability analysis found biases learned from the training data. The authors frame the work as a step toward better understanding personalized health risks and informing precision medicine approaches.
Source: Artem Shmatko et al., “Learning the natural history of human disease with generative transformers,” Nature, 2025.