New AI Tool Predicts Which of 1k Diseases Someone May Develop in 20 Years (www.scientificamerican.com)

Researchers announced Delphi-2M, a modified large language model (a generative pre‑trained transformer) that predicts an individual’s likelihood of developing 1,258 diseases up to 20 years ahead using prior medical history plus age, sex, BMI and lifestyle factors (e.g., smoking, alcohol). Trained on 400,000 participants from the UK Biobank and tested on 1.9 million Danish registry records, Delphi-2M matched or exceeded the accuracy of many existing single‑disease risk models and outperformed a biomarker‑based machine‑learning approach for multi‑disease prediction. It performs best for conditions with predictable progression (some cancers) and can produce whole health trajectories rather than isolated risk scores. The work is significant because it demonstrates that transformer architectures can be repurposed to model longitudinal, multi‑morbidity risk at scale—potentially enabling clinicians and health systems to identify high‑risk patients and prioritize preventive interventions across many conditions simultaneously. Important caveats remain: training was limited to datasets that capture only first disease events, raising questions about repeat episodes, generalizability, and bias; cross‑country validation and integration with clinical workflows are still needed. If those hurdles are addressed, Delphi‑2M could shift preventive care from single‑disease screening toward coordinated, multi‑condition risk management.