Researchers in China and the U.S. have designed a machine to diagnose sick people based on their doctors’ notes — with an accuracy that rivals its human counterparts.
They say the system, a type of neural network, could help physicians make differential diagnoses — or decide between two or more diseases with similar symptoms — as well as prioritize patients for treatment in emergency rooms.
It has learned to read electronic health records and extract a patient’s symptoms, history, and test results to make diagnoses, say the researchers, who trained it using data from half a million children that visited a Guangzhou pediatric hospital over 18 months to July 2017.
The research team, led by Kang Zhang of the University of California San Diego, published its work this week in Nature Medicine.
They pitted the system against 20 physicians on a random sample of 12,000 patient records from a second Guangzhou hospital. It rivaled doctors with many years of training in accuracy, but its performance depended on the disease.
It was better at spotting respiratory disease than were doctors with 25 years’ experience, and excellent at identifying potentially deadly bacterial meningitis, a brain infection, but poorer than human physicians at identifying encephalitis, or inflammation of the brain.
“This result suggests that this AI model may potentially assist junior physicians in diagnoses but may not necessarily outperform experienced physicians,” the researchers wrote.
It could also be used as a triaging tool at hospitals to ensure “physicians’ time is dedicated to the patients with the highest and/or most urgent needs,” as well as to reduce waiting times.
And it could help doctors make better differential diagnoses, because physicians tend to be biased towards diagnosing diseases they have seen regularly or recently, the researchers wrote.
Support independent journalism from China. Subscribe to Caixin Global starting at $0.99.