AI Model Forecasts Risk of More Than 1,000 Diseases Decades in Advance

Key Takeaways:

• Delphi-2M, a generative AI model, can predict susceptibility to over 1,000 diseases using anonymised medical records.
• The system has been tested successfully across large-scale UK and Danish datasets, showing remarkable accuracy and transferability.
• Experts say the model could transform population health forecasting within years and may eventually be adapted for personalised clinical use.
  
  
  

European scientists build AI model to predict long-term disease risk

European researchers have unveiled a powerful new artificial intelligence model that can predict a person’s susceptibility to more than 1,000 diseases decades before symptoms arise.

The system, known as Delphi-2M, was developed by scientists at the European Molecular Biology Laboratory (EMBL) in Cambridge. “Delphi uses a similar architecture to large language models but with key innovations to work with healthcare data,” explained Tom Fitzgerald of EMBL.

Delphi was trained on anonymised health records from 400,000 participants in the UK Biobank, a major long-term biomedical study. The researchers then validated its performance using records from 1.9 million patients in the Danish National Patient Registry.

Matching and exceeding existing prediction tools

The predictions made by Delphi-2M spanned more than 1,000 diseases and were generally comparable in accuracy to existing clinical tools that focus on specific conditions, such as the QRisk score used for cardiovascular disease risk. Results from the study were published in Nature on Wednesday.

“Our model is a proof of concept, showing that it’s possible for AI to learn many of our long-term health patterns and use this information to generate meaningful predictions,” said Ewan Birney, EMBL’s interim executive director. “We were surprised at how well the model transferred from the UK to Denmark though it had never seen a single bit of Danish data.”

Birney emphasised that turning Delphi into a clinically deployable forecasting tool could take five to ten years, but he noted that it could be used much sooner to inform public health strategies.

Population-level insights and healthcare planning

While Delphi generates predictions at the level of individual patients, its most immediate value may be in population health planning. “Although it makes predictions for each individual, it can be very useful at the population level to forecast collective healthcare needs, how many people will suffer from particular diseases such heart attacks, cancers or diabetes and what sort of treatment they need,” said Moritz Gerstung, head of AI at the German Cancer Research Center in Heidelberg and a member of the Delphi team.

The model performed best for diseases with well-understood and consistent progression patterns, such as cardiovascular disease, diabetes and sepsis (blood poisoning). It was less effective for conditions triggered by unpredictable environmental factors or for very rare congenital disorders.

Expanding to genomics and biological data

Researchers are now working to enhance Delphi by incorporating biological information, such as genomic and proteomic data. Despite this, Birney said they were “very pleasantly surprised” at how well the model performed using healthcare records alone, achieving results comparable to or better than some models that rely on genetic and protein-level data.

“I want to stress the power of the straightforward medical record,” Birney added.

The team has patented key aspects of Delphi’s approach to predicting disease risk and timing. “We are exploring whether there are commercialisation possibilities and how to do that with our respective institutions,” Birney confirmed.

Towards ethical and scalable predictive medicine

Independent experts have praised the work as an important step forward for responsible AI in medicine. “This research looks to be a significant step towards scalable, interpretable, and — most importantly — ethically responsible form of predictive modelling in medicine,” said Gustavo Sudre, professor of genomic neuroimaging and AI at King’s College London, who was not involved in the study.

He added that while the current model relies solely on anonymised health records, its architecture has been designed to handle richer data types in the future, including biomarkers, imaging and genomics.