Scientifically Speaking | What if your health could be forecast like weather?

A new study in Nature describes an artificial intelligence system that helps forecast disease like systems currently predict the weather

Imagine your doctor handing you a printout that reads like a weather forecast: 70% chance of diabetes by 2034, scattered heart problems in your late fifties, with a high-pressure front of hypertension moving in around age 45.

The system, called Delphi-2M, represents the most ambitious attempt yet to use generative AI to map the arc of human disease. (Representational image)

The genomics revolution promised personalised medicine. Artificial intelligence brings us closer to it. A new study in Nature describes an artificial intelligence system that helps forecast disease like systems currently predict the weather. But instead of telling you there’s a 70% chance of rain on Saturday, it can tell you there’s a 70% chance of developing diabetes, cancer, or heart disease in the next decade.

The system, called Delphi-2M, represents the most ambitious attempt yet to use generative AI to map the arc of human disease. Trained on medical records from more than 400,000 people in the UK and tested on nearly two million in Denmark, it can predict the risk and timing of over 1,000 different diseases simultaneously.

Delphi works much like ChatGPT, using the same transformer architecture that powers conversational AI. But instead of predicting the next word in a sentence, it predicts events in a patient’s life.

Feed it a patient’s medical history of asthma, high blood pressure, and smoking, and it doesn’t just predict what disease strikes next, but sets a countdown timer. It learned that diabetes often travels with eye disease and nerve damage. It discovered that cancers cast long shadows over mortality risk, while the danger from sepsis fades quickly. It found clusters of mental health conditions that persist for years.

The numbers are jaw-dropping. Delphi can predict the next disease diagnosis with accuracy comparable to specialised single-disease risk calculators. But unlike those narrow tools, it sees the whole picture. The model works best for diseases with consistent patterns of progression like cardiovascular disease, type 2 diabetes, and blood poisoning. It is less informative for rare or highly unpredictable conditions. Its first real use may be at the population level, helping health systems prepare, long before it is ready to guide individual patients in the clinic.

Even more stunning, it can generate “synthetic patients” out of data, complete with life histories of disease that never happened to real people but follow the statistical patterns of human illness. This feature could allow new medical AI models to be trained without exposing sensitive personal records. When researchers tested the model on Danish health records, accuracy was almost the same, suggesting that many of the patterns it learned do generalise across populations.

In India for example, instead of waiting for chest pain or blurred vision to appear, doctors could identify high-risk patients years in advance. Hospitals could prepare for the number of heart attacks in a city in 2030 the way meteorologists track storms. With our shortage of specialists and strained healthcare infrastructure, that kind of foresight could mean the difference between prevention and crisis management.

But before we embrace this vision of predictive medicine, we should remember that we’ve been here before.

The genomics revolution made strikingly similar promises two decades ago. The Human Genome Project was completed with fanfare about ushering in an era of personalised medicine. Companies like 23andMe sold the dream of unlocking genetic destiny. It didn’t happen. Most genetic risk scores turned out to have somewhat limited clinical utility, though researchers are finding new uses today. Genes influence health, but they don’t determine it.

Now AI researchers are making similar promises with similar confidence. But there’s a difference and AI might have an upper hand. While genes are largely fixed at birth, medical histories evolve throughout life, accumulating information that becomes increasingly predictive over time.

The Delphi study, to its credit, is refreshingly honest about limitations. The model inherited significant biases from its training data. UK Biobank participants were mostly white, affluent, and between 40 and 70 years old. Early deaths, childhood illnesses, and much of the diversity of real populations were missing from the dataset.

For India, these limitations are particularly concerning. Our health records are scattered across public and private systems, rural clinics, and informal care networks. Lifestyles and genetic backgrounds differ dramatically from the Western populations that dominate current AI training datasets. A model trained primarily on middle-aged Europeans might miss crucial patterns in Indian childhood diseases, maternal health, or elderly care needs. So, we need our own models.

The model also revealed troubling artifacts in how it learned. Patients with any hospital record were predicted to develop other hospital-diagnosed conditions at much higher rates. This was clearly a data quirk. The AI had learned that some patients simply have better access to hospital care, not that they’re sicker. Such biases could perpetuate existing healthcare inequalities if deployed without careful oversight.

Tech companies and insurers are circling like vultures. The dystopian question looms: will AI fortune-telling mean you pay penalty premiums for your predicted future?

There are also deeper questions about what these predictions actually mean. Accuracy at the population level doesn’t guarantee usefulness for any individual. A 70% disease risk is valuable information, but only if acting on it improves outcomes.

Yet despite these caveats, we’re witnessing a seismic shift in how medicine could work. We have conjured crystal balls for centuries. But still we are stuck in endless cycles of crisis management, treating heart attacks, managing strokes, starting insulin after diabetes has already taken hold. Delphi hints at a fundamental shift toward forecasting as a routine part of medical care.

This change will not happen overnight. The researchers themselves acknowledge that individual clinical applications may take five to ten years of additional development and testing. That span covers not just refining the model, but also conducting trials to prove benefit, navigating regulatory approval, and preparing for safe use. Population-level uses, such as helping governments predict cancer rates or dialysis needs, could arrive much sooner.

This distinction between population and individual predictions is crucial. Weather forecasts work because meteorologists can be wrong about rain at your house while being right about regional precipitation patterns. Medical AI might follow a similar path, proving useful for health system planning before it’s ready to guide individual treatment decisions.

(Anirban Mahapatra is a scientist and author, most recently of the popular science book, When The Drugs Don’t Work: The Hidden Pandemic That Could End Medicine. The views expressed are personal.)