Do you really Want to Know
Would you want to know as much as 20 years ahead whether you’d develop something like pancreatic cancer? We may soon be confronted with this question. Researchers have built an AI that predicts a person’s risks for more than 1,000 diseases up to two decades before symptoms appear. It could help people avoid or ease the effect of certain illnesses. But that kind of information can also be stressful.
The AI we’re talking about is Delphi-2M, a large language model AI like ChatGPT. In this case, it models health outcomes using past medical events, sex, body mass, lifestyle factors such as smoking and alcohol consumption and how they interact to predict future health issues.
The AI was trained using 400,000 UK Biobank participants, and then tested using 2 million health records from Denmark.
Delphi-2M reads the information and predicts the illnesses and when they’ll happen if the most likely sets of events occur. It can generate a complete health trajectory, but the accuracy drops to 60 percent at about 20 years out. It’s also not as good at predicting rare diseases and conditions that result from being in certain environments. Plus, it’s trained on the UK Biobank, which is biased toward mostly white, educated and relatively healthy people.
Once the model is improved with more diverse data, one day, doctors could use it along with patients’ data to make longer-term health plans and even help them live longer lives.
More Information
New AI System Predicts Risk of 1,000 Diseases Years in Advance
Delphi-2M reads your medical history like a language model reads text—and forecasts health problems 20 years out with surprising accuracy.
Learning the natural history of human disease with generative transformers
Decision-making in healthcare relies on understanding patients’ past and current health states to predict and, ultimately, change their future course1,2,3. Artificial intelligence (AI) methods promise to aid this task by learning patterns of disease progression from large corpora of health records4,5. However, their potential has not been fully investigated at scale. Here we modify the GPT6 (generative pretrained transformer) architecture to model the progression and competing nature of human diseases.