New AI spots disease risks years before symptoms show

Researchers have developed a groundbreaking AI system called Delphi-2M that can predict the risk of over 1,000 diseases, up to a decade in advance. Trained on data from over 400,000 UK Biobank participants and validated on 1.9 million Danish medical records, Delphi-2M accurately forecasts the progression of conditions like type 2 diabetes, heart attacks, and sepsis.

With its ability to identify high-risk individuals, this AI tool has the potential to revolutionize preventive healthcare, enabling doctors to intervene early and help patients avoid serious illnesses through targeted treatments and lifestyle modifications. The study was a collaboration between the European Molecular Biology Laboratory, the German Cancer Research Centre (DKFZ), and the University of Copenhagen.

How Delphi-2M works

Delphi-2M harnesses the power of generative language models, much like ChatGPT. By recognizing predictable patterns in medical events, the AI model can forecast health outcomes based on these patterns. The system analyzes a patient's medical history, including diagnoses and factors such as age, sex, smoking habits, and obesity, to assess the likelihood of developing diseases like cancer, diabetes, heart conditions, or respiratory illnesses.

While AI models have been used in healthcare to calculate the risk of specific diseases, Delphi-2M stands out by predicting the risk for over 1,000 diseases simultaneously. Trained using anonymized data from 400,000 individuals in the UK, its reliability was further validated using medical data from nearly 2 million Danish citizens. The results showed comparable accuracy across both populations.

'In the past, sewerage was a tipping point for health, now AI'

Just as sewage systems improved health at the time, AI will fulfill that role now.

Read More

Accuracy and predictive power

Scientists report that Delphi-2M's accuracy is on par with existing models that focus on single diseases. The AI model achieves an average accuracy score of 0.76, where 1.0 represents a perfect prediction. For longer-term predictions, extending beyond 10 years, the average score remains at 0.70. In some cases, the researchers suggest that the system can predict diseases up to 20 years in advance, particularly for conditions with a predictable disease progression, such as heart attacks and various cancers.

The insights provided by Delphi-2M could empower patients to make specific lifestyle changes to reduce their risk of developing certain diseases. Beyond individual health, this technology could also assist hospitals in anticipating future healthcare demands. Professor Ewan Birney, interim executive director at the European Molecular Biology Laboratory, noted the model's reliability, telling the BBC, "If our model says it's a one-in-10 risk for the next year, it really does seem like it turns out to be one in 10."

The researchers plan to conduct further testing of Delphi-2M across diverse populations and countries. While the AI model, described in the journal Nature, requires refinement and testing before clinical application, scientists anticipate that it will be ready for use within a few years.