Briefs · Frontier Science
Your medical history just became a weather forecast
Delphi-2M does not predict your future. It predicts your weather.
Published in Nature on 17 September, Delphi-2M is a generative transformer — the same architecture family as ChatGPT — trained not on words, but rather on medical histories. It learned from 400,000 UK Biobank participants, and was validated, without changing a single parameter, on 1.9 million Danish patients.
The result is a model that predicts the rates of more than 1,000 diseases, conditional on each person’s past disease history, with accuracy comparable to existing single-disease models. Its generative nature also lets it sample synthetic future health trajectories, estimating disease burden for up to 20 years.
Firstly, this is not a crystal ball. A forecast is a probability, not a destiny, and the authors are explicit that Delphi-2M is not a clinical tool yet. Secondly, it is a shift in kind: one model, a thousand diseases, instead of a thousand models, one disease each.
Conclusion: medicine is not fortune-telling, but rather forecasting — disease, decoded, twenty years out.
Read the original at Nature / PubMed
Summary in our own words. The original belongs to its publisher.