Every year, millions of Americans leave the office of a doctor with a wrong diagnosis. Doctors try to be systematic when they identify a disease, but prejudices creep in. Alternatives are neglected.

Now, a group of researchers in the United States and China has tested a potential cure for far too human weaknesses: artificial intelligence.

In an article published Monday in Nature Medicine, scientists reported that they had built a system that automatically diagnose common childhood conditions – from influenza to meningitis – after treatment of symptoms, antecedents, laboratory results and other clinical data of the patient.

The researchers said the system was very accurate and that one day it could help doctors diagnose complex or rare conditions.

Drawing on the records of nearly 600,000 Chinese patients who visited a pediatric hospital over a period of 18 months, the extensive data collection used to form this new system highlights an advantage for the China in the global race to artificial intelligence.

With so many people and privacy standards imposing fewer restrictions on digital data sharing, it may be easier for Chinese companies and researchers to create and train digital systems. "In-depth learning" that rapidly changes the trajectory of health care. .

Monday, President Trump signed a decree designed to stimulate A.I.'s development in the United States, academia and universities. As part of this "American I.I. Initiative ", the administration will encourage federal agencies and universities to share data that can foster the development of automated systems.

Pooling health care data is a particularly difficult undertaking. While researchers went to a single Chinese hospital to get all the data needed to develop their artificial intelligence system, collecting such data from US facilities is rarely as simple.

"You have to go to several places," said Dr. George Shih, associate professor of clinical radiology at Weill Cornell Medical Center and co-founder of MD.ai, a company that helps researchers label A.I.'s service data. "The equipment is never the same. You must make sure that the data is anonymous. Even if you get an authorization, it's a huge amount of work. "

After redesigning Internet services, consumer devices and driverless cars early in the decade, deep learning is rapidly moving into a myriad of health care areas. Many organizations, including Googledevelop and test systems that analyze electronic health records to report medical conditions such as osteoporosis, diabetes, hypertension and heart failure.

Similar technologies are being developed to automatically detect signs of disease and disease during X-rays, MRIs and eye scans.

The new system is based on a neural network, a breed of artificial intelligence that accelerates the development of everything from health care to driverless cars at military applications. A neural network can learn a lot of tasks by itself by analyzing large amounts of data.

With the help of this technology, Dr. Kang Zhang, Head of the Department of Ophthalmic Genetics at the University of California at San Diego, has developed systems capable of analyzing the eye sweep of hemorrhages. , lesions and other signs of diabetic blindness. Ideally, such systems would be a first line of defense to screen patients and identify those who need more attention.

Now, Dr. Zhang and his colleagues have created a system that can diagnose an even wider range of conditions by recognizing patterns in the text, not just in medical images. It could increase what doctors can do on their own, he said.

"In some situations, doctors can not consider all the possibilities," he said. "This system can check on the spot and make sure the doctor has not forgotten anything."

The experimental system analyzed the electronic medical records of nearly 600,000 patients at the Guangzhou Medical Center for Women and Children in South China, and learned to associate common medical problems with specific information collected by doctors, nurses and other technicians.

First, a group of trained doctors annotated hospital records by adding labels identifying information about certain medical conditions. The system then analyzed the tagged data.

Then, the neural network has received new information, including the patient's symptoms as determined by a physical examination. Soon, he was able to make connections by himself between the written records and the observed symptoms.

When tested on untagged data, the software could compete with the performance of experienced physicians. The diagnosis of asthma was accurate to more than 90%; the accuracy of the physicians participating in the study ranged from 80 to 94%.

To diagnose gastrointestinal diseases, the accuracy of the system was 87%, compared to 82 to 90% for doctors.

Capable of recognizing patterns of data that humans could never identify alone, neural networks can be extremely powerful in the right situation. But even the experts have trouble understanding why such networks make particular decisions and how they teach themselves.

As a result, extensive testing is needed to reassure physicians and patients about the reliability of these systems.

The experts said Dr. Zhang's system required extensive clinical trials, given the difficulty of interpreting decisions made by neural networks.

"Medicine is a slowly evolving field," said Ben Shickel, a researcher at the University of Florida, specializing in the use of in-depth learning for health care. "Nobody will just deploy one of these techniques without rigorous tests that show exactly what's going on."

It could take years before deep learning systems are deployed in emergency rooms and clinics. But some are closer to reality: Google is currently conducting clinical trials of its eye scan system at two hospitals in southern India.

In-depth diagnostic tools are more likely to thrive in countries other than the United States, said Dr. Zhang. Automated screening systems can be particularly useful in places where doctors are scarce, especially in India and China.

The system developed by Dr. Zhang and his colleagues benefited from the wide range of data collected at the Guangzhou Hospital. Similar data sets from US hospitals are generally smaller, both because the average hospital is smaller and regulations make it difficult to share data from multiple facilities.

Dr. Zhang said he and his colleagues took care to protect the privacy of patients in the new study. But he acknowledged that researchers in China could have an advantage in collecting and analyzing this type of data.

"The size of the population – the size of the data – is a big difference," he said.