By Chelsea Whyte
Diagnosing an illness requires taking in a lot of information and connecting the dots. Artificial intelligence may be well-suited to such a task and in recent tests one system could diagnose children’s illnesses better than some doctors.
Kang Zhang at the University of California in San Diego and his colleagues trained an AI on medical records from 1.3 million patient visits at a major medical centre in Guangzhou, China. The patients were all under 18 years old and visited their doctor between January 2016 and January 2017.
Their medical charts include text written by doctors and laboratory test results. To help the AI, Zhang and his team had human doctors annotate medical records to identify portions of text associated with the child’s complaint, their history of illness, and laboratory tests.
When tested on previously unseen cases, the AI could diagnoseglandular fever (also known as mononucleosis), roseola, influenza, chicken pox and hand-foot-mouth disease with between 90 and 97 per cent accuracy. It’s not perfect, but neither are human doctors, says Zhang.
“When you’re busy you can see 80 patients a day. And you can only grasp so much information. That’s where we potentially as human physicians might make mistakes. AI doesn’t have to sleep, it has a large memory and doesn’t lose energy,” he says.
The team compared the model’s accuracy to that of 20 paediatricians with varying years of experience. It outperformed the junior paediatricians, though the senior ones did better than the AI.
The AI could be used to triage patients in emergency departments. “Given sufficient data, AI should be able to tell if this is an urgent situation and needs referral or if it’s a cold,” says Zhang.
Chris Russell at the Alan Turing Institute in London says this won’t let people bypass doctors entirely when seeking medical treatment, because these medical records still need to be created by trained professionals, and their knowledge is key to the diagnosis. “Someone needs to be there discussing your symptoms and putting them into the machine. I don’t see how this technology could be used to take doctors out of the loop. It could be used to help them, but it’s a very long way from replacing medical professionals,” he says.
It’s possible that junior doctors who may rely on AI like this to make diagnoses could miss out on learning how to see patterns in patient complaints. And people may feel uncomfortable with this type of medical care. “If it’s deployed as an interface directly with the person where they type in their symptoms, I can see how people would be very uncomfortable with this. When you go see a doctor, you want to feel like there’s someone there who cares about you,” Russell says.
“But you don’t want to go to the emergency room and wait 5 hours because you have some pain in the abdomen that’s not appendicitis but just related to gastroenteritis or the food you ate. All those diseases have tell-tale signs, and just as we physicians ask a series of questions to drive a diagnosis, AI can do the same,” says Zhang.
He and his team are now training the AI to diagnose adult diseases, as well.
Journal reference: Nature Medicine, DOI: 10.1038/s41591-018/-0335-9
More on these topics: