Tom Fitzgerald, Senior Research Staff Scientist at European Bioinformatics Institute presents how generative transformer models can be used to read longitudinal health records and learn how human diseases emerge, interact, and evolve over time.
This work shows how GPT‑style AI can move beyond language to simulate future health trajectories, offering new tools for population health research and precision medicine.
Abstract:
Modern medicine generates enormous amounts of data, but making sense of how illnesses unfold over time remains one of healthcare’s hardest challenges. Patients rarely experience diseases in isolation—conditions emerge, interact, and compete over years or even decades. Understanding these timelines is crucial for predicting risk, planning care, and improving population health, yet most predictive tools focus on single diseases and short time horizons.
In this talk, we explore a new approach that borrows ideas from the same generative AI systems behind large language models. Instead of predicting words in a sentence, researchers adapted transformer models to “read” a person’s medical history and learn the natural progression of human disease across a lifetime. Trained on health records from hundreds of thousands of individuals and validated across different countries, this model can estimate the likelihood of more than a thousand diseases based on earlier health events.
Because the model is generative, it can also simulate plausible future health trajectories—offering a way to explore long‑term disease burden and to train other AI systems without exposing real patient data. Importantly, explainable AI techniques reveal patterns the model learns: common clusters of co‑occurring conditions, how early diagnoses shape future risk, and where societal or data‑driven biases may influence predictions.
This work illustrates how general‑purpose AI architectures can be repurposed to model complex, real‑world systems beyond language. For a GenAI audience, it raises broader questions about trust, bias, simulation, and the role of generative models in decision‑making—hinting at a future where AI helps clinicians and researchers understand not just diseases, but the stories of health that unfold over time.
Free
Tom Fitzgerald is a Senior Research Staff Scientist and member of faculty at the European Molecular Biology Laboratory–European Bioinformatics Institute (EMBL‑EBI), based at the Wellcome Genome Campus in the UK, where he works on large‑scale analysis of population and genomic data. He is a co‑author of the Nature paper Learning the natural history of human disease with generative transformers, which introduces DELPHI‑2M, a transformer‑based model trained on national‑scale health records to learn and simulate patterns of disease progression over time. Over his career at EMBL‑EBI and previously the Wellcome Sanger Institute, Tom has contributed to high‑impact research at the intersection of genomics, population health, and data‑driven modeling, with a focus on developing scalable analytical frameworks that turn complex biomedical data into insights about human disease.