AI to detect common illness rivals performance of experienced physicians

Researchers in the U.S. and China have built a system using artificial intelligence that automatically and successfully diagnoses common childhood medical conditions after processing clinical data, according to The New York Times.

The researchers published a paper on the system in Nature Medicine. They said the system was highly accurate and one day may help physicians diagnose complex conditions. The system relies on a neural network that can learn tasks by analyzing large amounts of data.

The researchers looked at the records of about 600,000 Chinese patients who received treatment at a pediatric hospital during an 18-month period. The system learned to link common medical conditions to specific patient information that physicians, nurses and other technicians collected.

A group of physicians added labels to hospital records that identified information on certain medical conditions, and then the system analyzed the labeled data.

Next, the neural network received patients' symptoms as determined during a physical exam and made connections on its own between the written records and observed symptoms.

When the system was tested on unlabeled data, it rivaled the performance of experienced physicians. It was more than 90 percent accurate at diagnosing asthma. The accuracy of physicians in the study ranged from 80 percent to 94 percent.

The system was 87 percent accurate in diagnosing gastrointestinal disease, compared to physicians' accuracy of 82 percent  to 90 percent.

Experts don't understand why the neural networks make the decisions they do and how they teach themselves, and they said extensive clinical trials are needed given the challenges of interpreting decisions the networks make.

More articles on clinical leadership & infection control:
Why hospital mergers may harm care quality
Sutter Health to offer free symptom-checking chatbot service
Johns Hopkins All Children's submits corrective plan to CMS amid safety lapse probe

© Copyright ASC COMMUNICATIONS 2021. Interested in LINKING to or REPRINTING this content? View our policies by clicking here.


Featured Whitepapers

Featured Webinars