At 00:14 on February 12, Beijing time, the internationally renowned medical research journal Nature Medicine published an article online titled "Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence". This is the first time that a top medical journal in the world has published research results on the use of natural language processing (NLP) technology to make clinical intelligent diagnosis based on Chinese text-based electronic medical records (EMR).

If this result is used for diagnosis, it is expected to reduce omissions and help early treatment. The research team strives to put it into practical use as a device to help doctors make judgments.

Think like a doctor and read medical records accurately

This article is a major scientific research result of using artificial intelligence technology to diagnose pediatric diseases by Professor Xia Huimin, Professor Zhang Kang (University of California, San Diego), Dr. Liang Huiying, Director of the Data Center, Director Sun Xin of the Medical Department, Director He Liya of the Pediatric Clinic, and the research team of Yitu Medical Ni Hao, Kangrui Intelligent Technology and other research teams and Guangdong Provincial Key Laboratory of Regenerative Medicine. This is also another important milestone achieved by the team in the application of AI technology in medical treatment in less than a year after publishing a paper on AI image diagnosis on the cover of Cell magazine. It marks the arrival of the era of AI simulating human doctors to diagnose diseases.

In recent years, AI has performed well in diagnostic tools based on medical images, but it is generally still limited to relatively standardized static image data. In this latest scientific research achievement, AI, based on image recognition, has gradually acquired a certain ability to analyze and reason about the condition by automatically learning the diagnostic logic in medical record text data (doctors' knowledge and language), and can further understand and analyze complex medical records, which means that AI may be able to "think" like doctors.

Researchers trained AI to understand clinical feature data in massive electronic medical records, including patient complaints, symptoms, personal history, physical examinations, laboratory test results, imaging test results, medication information and other data. The research team used YITU Medical's NLP technology to establish a medical record intelligent analysis system to deeply mine and analyze the information of medical texts, and transform medical record data in the form of unstructured text into standardized, standardized and structured data so that AI can "understand" medical records accurately and completely. To this end, doctors, scientists and technicians worked together, and an expert team consisting of more than 30 senior pediatricians and more than 10 informatics researchers manually annotated more than 6,000 charts in electronic medical records, and continued to test and iterate the model.

The research team also developed an intelligent recommendation system for diagnostic results, which simulates the diagnosis and treatment path of human doctors and judges the target children step by step. Sun Xin, director of the Medical Department of Guangzhou Women and Children's Center, believes that "the high-quality medical prior knowledge input of professional pediatricians has become the key advantage of this system."

For some dangerous and potentially life-threatening diseases (such as acute asthma attacks, bacterial meningitis, etc.), the algorithm also showed strong diagnostic performance. He Liya, director of the Pediatric Department of Guangzhou Women and Children's Center, believes: "This is of great significance in clinical applications, because with the assistance of AI rapid triage, the limited resources of medical services can be used for patients who need help most."

Can be applied to the diagnosis of a variety of common pediatric diseases

By automatically learning the diagnostic logic from 1.36 million high-quality electronic text medical records from 567,000 children, the AI can be applied to the diagnosis of a variety of common pediatric diseases with an accuracy comparable to that of experienced pediatricians. The researchers randomly selected 12,000 medical records of children and divided the 20 participating pediatricians into five groups according to their seniority and clinical experience to see which group of doctors the AI's performance was close to. The results showed that the average score of the AI model was higher than that of the two groups of junior doctors and close to that of the three groups of senior doctors.

According to the researchers, the AI system can obtain oral texts from patients or parents through human-computer interaction, including information such as the chief complaint, symptoms, medical history, and medication history, make a rough diagnosis, and give a possible range of diseases; through face-to-face consultations with doctors or remote consultations over the Internet, detailed medical conditions and differential diagnostic features are obtained, and the model recalculates based on this to give a specific and accurate diagnosis; if there is laboratory test or imaging examination data, the AI model can further confirm its diagnosis results. More importantly, it has the function of incremental learning. In practice, it will enhance the memory of the adopted results, and for the unadopted results, it will continue to learn to improve its ability after verification. "Dr. Liang Huiying, director of the data center of Guangzhou Women and Children's Medical Center (first author of this article), revealed that after three months of improvement and iteration after the launch, the system has been called more than 30,000 times in the first quarter of 2019. He emphasized that the data of these calls is a compass for evaluating the practical performance of "Assisted Diagnosis Bear" and improving its targeted capabilities.

In the future, AI may be able to diagnose more diseases

According to the research team, the artificial intelligence-assisted diagnosis system will be applied to clinical practice in a variety of ways. First, it can be used as a triage program. For example, when a patient comes to the emergency department, the nurse can obtain their vital signs, basic medical history and physical examination data and input them into the model, allowing the algorithm to generate predictive diagnoses to help physicians screen which patients to prioritize; another potential application is to help physicians diagnose complex or rare diseases. In this way, physicians can use AI-generated diagnoses to help broaden differential diagnoses and think about diagnostic possibilities that may not be immediately apparent.

Regarding the future of AI-assisted diagnosis systems, Professor Xia Huimin, director of the Guangzhou Women and Children's Medical Center, said that this research will become an important milestone in the implementation of AI technology in medical care. Its greatest contribution is that AI can not only "see pictures", but also "read words", and can understand the disease information contained in the text like humans. Through systematic learning of text medical records, AI may be able to diagnose more diseases. Experts pointed out that a lot of basic work still needs to be improved, such as the integration of high-quality data, which is a long process, because the collection and analysis of big data requires the cooperation of multiple experts including algorithm engineers, clinicians, epidemiologists, etc. In addition, after AI has learned massive amounts of data, the accuracy of its diagnostic results still needs to be verified and compared with a larger range of data. ”