Scientists detail Delphi-2M, a generative AI model trained on large-scale health records that can predict susceptibility to 1,000+ diseases decades from now
meet Delphi-2M Markus Kasanmascheff / WinBuzzer : New AI Predicts Over 1,000 Diseases Decades in Advance Using Health Records James Gallagher / BBC : AI can forecast your future health - just like the weather Andrew Gregory / The Guardian : New AI tool can predict a person's risk of more than 1,000 diseases, say experts Vicky Hatch / EMBL : AI model forecasts disease risk decades in advance Eric Topol / Ground Truths : Dawn of A New Era of Primary Prevention in Medicine Nature : This AI tool predicts your risk of 1,000 diseases — by looking at your medical records Thomas Moore / Sky News : New AI model could predict risk of disease a decade before symptoms Nature Podcast on YouTube : This AI tool predicts your risk of 1,000 diseases — by looking at your medical records Gemma Conroy / Nature : Which diseases will you have in 20 years? This AI makes accurate predictions Daniel Mediavilla / El País : New AI model can predict what diseases you'll have in 20 years X: Guy Venables / @godfreychurch : My illustration for Thursday's @MetroUK on a story about a new AI tool that can estimate the risk of more than 1,000 diseases for people in one go, providing key information about their future health. Or not. @MetroPicDesk @NHSuk [image] Eric Topol / @erictopol : The first large language health model that predicts >1,000 diseases and the time they will occur in a person was just published @Nature, by @MoritzGerstung @artem_shmatko and colleagues https://www.nature.com/... This, along with other recent groundbreaking reports, represents the [image] Joe Pickrell / @joe_pickrell : Fun language model approach to modeling health, with application to the UK Biobank. “Learning the natural history of human disease with generative transformers” https://www.nature.com/... [image] @embl : Imagine a future where your medical history could help forecast what health conditions you might face. Introducing a new AI model trained on anonymised medical records that estimates long-term risk for over 1,000 diseases over a decade in advance. https://www.embl.org/... [image] LinkedIn: Mireille Hildebrandt : “The predictions across more than 1,000 diseases generally matched the accuracy of existing tools that have a far narrower focus, such as the QRisk score for heart conditions. … Bluesky: Magdalena Skipper / @magdalenaskipper : The shape of things to come? — Modified GPT6 architecture, trained on UK Biobank & validated in a 2nd population, allows for modelling natural history of human disease — #MedSky 🧪 — @nature.com — www.nature.com/articles/s41...
Context & Ripple Effects
Delphi-2M extends a progression from single-condition or modality-specific prediction toward broader models of disease trajectories. Earlier work included RETFound's risk predictions from retinal images and Foresight, a model trained on anonymized NHS records at national scale to forecast disease and hospital outcomes.
Its reported external validation matters because the model is being presented not merely as a health-record-trained system, but as one tested beyond its initial population. The result also underscores the strategic value of large, longitudinal datasets such as UK Biobank for medical AI research.
First-order effects
- Researchers now have a generative model reported to estimate susceptibility and likely timing across more than 1,000 diseases, expanding the scope of health-record-based risk modeling beyond narrower prediction tools.
- Organizations holding longitudinal medical-record datasets gain a clearer use case for training and validating models that represent disease histories over long time horizons.
Second-order effects
- Developers of clinical prediction tools will face pressure to show whether narrower models retain advantages in accuracy, interpretability, or validation against broad multi-disease systems.
- External validation raises the bar for comparable health AI research, while making access to diverse, well-governed patient data a more important competitive and research constraint.
Third-order effects
- If such results reproduce across populations and settings, preventive medicine could increasingly be organized around continuously updated, multi-condition risk estimates rather than separate disease-specific scores.
- The likely durable bottleneck is not model architecture alone but trusted access to representative longitudinal health data, concentrating influence among institutions able to govern, link, and validate those records.
The trend: Health AI is moving from predicting isolated outcomes toward foundation-style models that learn long-term disease trajectories from large longitudinal datasets.