Trained on over 400k patient records, AI predicts health for up to 20 years (singularityhub.com)

Researchers at the German Cancer Research Center unveiled Delphi-2M, an LLM-style predictive model trained on >400,000 UK Biobank records (plus lifestyle covariates) that forecasts individual risk trajectories for 1,258 diseases up to 20 years out. Instead of natural language, Delphi treats standardized diagnostic codes as tokens and ingests additional inputs (blood tests, BMI, smoking/alcohol, sex) to predict multi-disease sequences. It produced explainable risk rationales, matched or outperformed conventional clinical risk scores and biomarker models for many conditions—especially cardiovascular disease and dementia—and handled population-level patterns (e.g., age and sex effects). The team also mapped nearly 2 million Danish patient records without retraining, indicating the model learned broadly applicable disease interactions, and it can generate synthetic records to mitigate privacy concerns. For the AI/ML and clinical communities, Delphi-2M demonstrates that sequence-based, transformer-like models can simultaneously model thousands of diseases and their timing, shifting tools from single-condition diagnostics toward long-range, multi-morbidity prevention and healthcare planning. Important caveats remain: predictions show associations not causation and inherit biases of the UK Biobank (skewed demographics, survivorship bias, sparse data for 80+). Delphi struggled with more lifestyle-dependent trajectories (e.g., Type 2 diabetes), and real-world deployment will require validation across diverse populations and careful handling of bias, consent, and clinical integration.