Weather-Style AI Forecasts 1,000+ Disease Risks Years Ahead
Delphi-2M forecasts risk for 1,000+ diseases like a weather report, reading health records to assign probabilities. Trained on UK and Danish data, it could guide care and planning.
AI That Forecasts Health Risk Like a Weather Report
An AI model called Delphi-2M can forecast a person's risk of more than 1,000 diseases over the next decade. It reads patterns in medical records and outputs probabilities, much like a 70% chance of rain. The goal: spot high-risk patients early and help hospitals plan demand years ahead.
Built by teams at the European Molecular Biology Laboratory, the German Cancer Research Centre (DKFZ), and the University of Copenhagen, the model was trained on anonymized data from the UK Biobank and validated on 1.9 million health records in Denmark. According to Prof Ewan Birney, if the model says there's a one-in-10 risk next year, it aligns with observed outcomes at that rate.
How Delphi-2M Works
Delphi-2M uses technology similar to large language models. Instead of predicting the next word, it predicts the next medical event and when it might happen, across 1,231 disease categories. It does not provide exact dates; it assigns calibrated probabilities.
Training data included hospital admissions, GP records, and lifestyle factors such as smoking from more than 400,000 UK Biobank participants. Performance is strongest for conditions with clearer progression-type 2 diabetes, heart attacks, and sepsis-than for more random events like infections.
What You Can Do With It
- Early risk stratification: Similar to using cardiovascular risk scores for statins, Delphi-2M can flag high-risk patients early. That opens the door to targeted advice or therapies-e.g., cutting back alcohol for those at elevated risk of specific liver disorders.
- Screening strategy: Inform who to screen, when, and for what, based on projected risk trajectories rather than broad age cut-offs alone.
- Service planning: Forecast local demand years ahead, such as expected heart attack volumes in a specific city by 2030, to guide staffing, beds, and equipment.
- Research: Map disease pathways and multimorbidity at scale to prioritize interventions and trials.
What the Researchers Say
"Just like weather, where we could have a 70% chance of rain, we can do that for healthcare," said Prof Ewan Birney. "And we can do that not just for one disease, but all diseases at the same time."
Prof Moritz Gerstung (DKFZ) noted that generative models like this could eventually support personalized care and system-wide planning. Prof Gustavo Sudre (King's College London) called the work a step toward scalable, interpretable, and ethically responsible predictive modeling in medicine.
Status, Limits, and Next Steps
The model is research-grade and not ready for clinical deployment. The team stresses the need for rigorous testing, regulation, and thoughtful rollout-similar to how genomics moved from labs to clinics over roughly a decade.
Bias is a concern: UK Biobank skews to ages 40-70 and may not represent all populations. The next iteration will incorporate imaging, genetics, and blood tests to improve generalizability and granularity.
How to Judge Models Like This
- Calibration: Do predicted risks match observed outcomes across risk bands?
- Transportability: Does performance hold across countries, systems, and demographics?
- Use-case utility: Does it improve clinical choices, resource allocation, or screening yield?
- Safety and governance: Audit for bias, monitor post-deployment drift, and ensure clear accountability.
Who Built It
- European Molecular Biology Laboratory (EMBL-EBI)
- German Cancer Research Centre (DKFZ)
- University of Copenhagen
Read More
Upskill on AI in Practice
If you're building or deploying predictive models in healthcare or population science, see the latest applied courses at Complete AI Training.
