A groundbreaking artificial intelligence model named Delphi-2M has demonstrated the remarkable ability to forecast a person's risk of developing more than 1,000 different diseases throughout their lifetime. While not yet ready for clinical deployment, this innovation represents a significant leap toward predictive medicine that could transform how doctors approach patient care.
How Delphi-2M Works: Learning from Medical Histories
Developed by research teams at the European Molecular Biology Laboratory (EMBL) in Cambridge and the German Cancer Research Centre in Heidelberg, Delphi-2M takes inspiration from the same technology that powers advanced language models like ChatGPT. The research findings were officially published in the prestigious journal Nature on September 17th.
The model applies principles similar to large language models but adapts them for medical prediction. Instead of analyzing text patterns to predict the next word in a sentence, Delphi-2M analyzes sequences of medical diagnoses to forecast which conditions a patient might develop next. The crucial adaptation involved teaching the AI to account for the time between medical events, replacing the positional encoding used in language models with age-based encoding.
Training and Testing on Massive Health Datasets
The researchers trained Delphi-2M using comprehensive health data from 400,000 people in the UK Biobank, one of the world's most complete biological datasets. The model learned to recognize patterns in the timing and sequence of ICD-10 codes - the international medical shorthand doctors use to record diagnoses.
After initial training, the model underwent validation using data from an additional 100,000 people in the UK Biobank. To ensure robustness, researchers further tested Delphi-2M on Danish health records, utilizing data from 1.9 million Danes dating back to 1978. This provided a more diverse and representative sample than the UK data alone.
Impressive Predictive Performance with Room for Improvement
Researchers measured the model's performance using AUC (area under the curve) metrics, where a score of 1 represents perfect prediction and 0.5 indicates random guessing. Delphi-2M achieved an average score of 0.76 for predicting diagnoses within five years using British data, with a slight decrease to 0.67 for Danish data.
The model proved particularly accurate at predicting conditions that typically follow specific previous diagnoses, such as death following sepsis. However, it struggled more with events caused by random external factors like viral infections. Predictably, accuracy diminished over longer timeframes, with the model scoring 0.7 when forecasting ten years into the future.
Future Applications and Ethical Considerations
Beyond helping identify high-risk patients, Delphi-2M could assist health authorities in allocating budgets for disease areas that may require additional funding in coming years. The model also offers valuable insights for biologists by revealing which conditions cluster together, potentially suggesting previously unexplored relationships between diseases.
However, real-world clinical applications remain years away. Delphi-2M must undergo rigorous trials to determine whether it actually leads to better patient outcomes. The development team is already working on enhancements, including the ability to process more sophisticated data like medical images and genome sequences available in the UK Biobank.
Delphi-2M isn't the only AI health forecasting model in development. Competing systems include Foresight from King's College London and ETHOS from Harvard University, though the Foresight project faced temporary pauses due to data access concerns.
As geneticist Ewan Birney from EMBL enthusiastically noted about the potential of such AI models, "I'm like a kid in a candy shop." While patients won't immediately benefit from Delphi-2M in their doctor's offices, the technology represents an exciting frontier in preventive medicine that could ultimately transform healthcare delivery worldwide.
