Paper to Practice

Learning the natural history of human disease with generative transformers

Delphi-2M adapts the GPT architecture to health records: trained on 402,799 UK Biobank participants, it reads a person's diagnosis history, predicts the rates and timing of 1,256 diseases plus death, and can simulate future health trajectories. It reached an average AUC of 0.76 internally and 0.67 on 1.93 million Danish registry records without retraining.