AI beat doctors on hard diagnosis cases in a new Science study
An AI that acts like a panel of doctors got far more tough cases right than the doctors it was tested against. The authors say it was a controlled test, not a hospital.
A new study in the journal Science tested how well AI can work out what is wrong with a patient. The result is hard to ignore.
The team built an AI system that acts like a panel of doctors. It does not see the whole case at once. It starts with a short note about the patient. Then it asks questions and picks tests, one step at a time, the way a real doctor does. It keeps updating its list of possible illnesses as new facts come in. It also tries to choose tests that do not cost too much.
The researchers first tried it on hard cases from the New England Journal of Medicine. These are puzzle cases that doctors study to sharpen their skills. The AI got 83.2 percent of them right. Generalist doctors in the study got 36.1 percent right.
They also used emergency room admissions from a research data set. There, the AI scored 80.3 percent. The doctors scored 46.8 percent.
The paper says the work drew on many teams at Microsoft.
The authors are careful about what this means. They say the results come from a controlled test setting. Still, they say it shows promise for diagnosis that is more accurate and costs less.
For now, this is research. It is not an app that checks your symptoms. If you feel sick, a doctor is still the right person to call.
Open it and check the claim yourself. That is the point.