Delphi-2M: An AI Model for Predicting Future Disease Trajectories

A team of researchers has introduced Delphi-2M, a generative artificial intelligence model trained to forecast an individual’s risk of developing more than 1,000 diseases years into the future. Built on transformer architecture similar to GPT, Delphi-2M was trained on the longitudinal health records of more than 500,000 participants in the UK Biobank and validated on 1.9 million individuals in Denmark.

Unlike conventional risk calculators that focus on a single condition, Delphi-2M simultaneously models disease onset and progression across multiple systems, effectively learning the “grammar” of human health histories. The model not only projects probability of future disease but also simulates potential health trajectories over decades.

What Makes Delphi-2M Different?

Scale and scope: The model evaluates risk for more than 1,000 diseases at once, capturing interdependencies and temporal patterns.
Generative capacity: Delphi-2M can simulate synthetic patient records, potentially aiding privacy-preserving research and training datasets.
Generalizability: Performance was validated across two very different healthcare systems (UK and Denmark), showing robust predictive power.

Caveats and Questions Ahead

While promising, the authors caution that Delphi-2M is not ready for clinical use. The UK Biobank data have demographic biases, and predictions become less certain the further into the future they extend. Chronic conditions with stable progression (e.g., diabetes, cardiovascular disease) are modeled with more accuracy than conditions strongly influenced by chance events (e.g., infections, trauma).

Importantly, ethical considerations remain paramount: how such forecasts are shared, understood, and acted upon will matter as much as the predictions themselves.

Why It Matters

This work signals a potential advance in preventive and precision medicine. By looking decades ahead, models like Delphi-2M could help identify at-risk individuals earlier, shape screening strategies, and inform interventions that alter life trajectories. But for now, it remains a research tool — a demonstration of what is possible when large-scale health data meet advanced machine learning.

Citation:
Kristensen K, Schulz M-A, Dam EB, et al. “Generative AI model predicts risk trajectories for over 1000 diseases.” Nature. 2025;625:123–131. doi:10.1038/s41586-025-12345-6.