Overview
- Researchers adapted large language model methods to medical diagnosis sequences to learn patterns that predict future health events.
- Delphi-2M estimated both risk and likely timing for more than 1,000 conditions, including cancers, diabetes, and cardiac and respiratory diseases.
- Training used records from nearly 500,000 UK Biobank participants with verification across about 1.9 million Danish health records dating back to 1978.
- The study reported performance of AUC 0.76 on UK data and 0.67 on Danish data, indicating promise but far from perfect prediction.
- Authors and external experts say deployment awaits wider testing to address demographic and outcome biases, improve interpretability, and set ethical governance, with the model differing from single-disease tools like QRISK3 by making multi-disease, long-horizon forecasts.