Recently, a research team led by Xia Huimin, a professor at Guangzhou Women and Children Medical Center, Zhang Kang, a professor at the University of California, San Diego, and Yi Tu Technology, an artificial intelligence company, designed a disease diagnosis system based on AI, and added medical knowledge maps to it, so that AI can "diagnose diseases" according to electronic medical records read by human doctors.
The results are also quite optimistic: through the detection of 55 common pediatric diseases and some critical diseases included in the system, the diagnostic level of AI can reach the professional level of pediatric attending doctors.
At present, the research result of "using artificial intelligence to evaluate and accurately diagnose pediatric diseases" was published online in the journal Nature-Medicine in mid-February.
Combining deep learning technology with professional medical knowledge map is the biggest feature of artificial intelligence aided diagnosis platform. Ni Haohao, president of Yi Tu Medical, said in an interview with the author that it is "very feasible" to learn clinical data in the future and provide doctors with more auxiliary diagnosis capabilities (diseases).
In order to make the AI-aided diagnosis platform have professional pediatric medical knowledge, the research team asked it to learn diagnostic logic in 1.36 million electronic medical records of 567,000 children. These electronic medical records from Guangzhou Women and Children Medical Center from June 20 16 to July 20 17 cover1.01.60 billion data points, and the initial diagnosis includes 55 common pediatric diseases.
In addition to integrating medical knowledge, the research team also used Yi Tu Science and Technology's natural language processing (NLP) technology to build a natural language processing model to annotate these electronic medical records. The model can roughly classify clinical information without "training" by standardizing medical records.
"Rough classification refers to taking the whole electronic medical record as input and expert diagnosis results as output, so as to realize rough classification. But this does not really understand the disease itself, and it is difficult to explain why such a diagnosis is made. " Ni Hao told me that although NLP model has broken through the barrier between medical record text language and computer language, knowledge map is the key for AI diagnostic platform to acquire expert ability.
This is also their next important work: an expert team composed of more than 30 senior pediatricians and 10 informatics researchers manually marked, continuously tested and iterated 6 183 charts on electronic medical records to ensure the accuracy of diagnosis.
After "training optimization verification" on the AI diagnostic platform through charts marked by senior medical experts, researchers found that NLP model after deep learning can well mark electronic medical records, achieving the highest sensitivity and accuracy in marking physical examination and chief complaint items respectively. In other words, NLP model of deep learning can accurately read the information recorded in electronic medical records, and can accurately make annotations that meet clinical standards. And this is also the most critical part of the whole research.
"NLP model has the ability to understand electronic medical records by introducing knowledge maps to deeply deconstruct the electronic medical records of each disease. For example, which features are closely related to hand, foot and mouth disease, and what are the most relevant features of Kawasaki disease, so that the model has better medical interpretability on the basis of accurate diagnosis. " Ni Hao explained, "With the help of knowledge map and deep learning technology to deconstruct electronic medical records, we can really understand clinical data. Based on this, algorithms such as machine learning classification are useful, otherwise it is impossible to use electronic medical records as' black boxes' to build high-precision interpretable models. "
Using deep learning technology and medical knowledge map to deconstruct electronic medical record data, researchers have built a high-quality intelligent disease database, which will make it easier to build various diagnostic models in the future.
Building a multi-level diagnosis model is the second step for researchers to build the AI diagnosis platform into a pediatrician. Ni Haohao said that this diagnostic model based on logistic regression classifier will be divided into several major systems, such as respiratory diseases, gastrointestinal diseases and systemic diseases, and then subdivided under each category-this is to let AI simulate the diagnosis and treatment path of human doctors and judge the data of target children step by step.
The results show that based on the data accurately read by NLP model, AI diagnostic model can accurately diagnose pediatric diseases: the average accuracy rate is 90%, and the diagnostic accuracy rate of neuropsychiatric disorders is as high as 98%.
The diagnosis model also performs well in the classification and diagnosis of corresponding pediatric diseases. The diagnostic accuracy of the system for upper respiratory diseases and lower respiratory diseases is 89% and 87% respectively. At the same time, the system also has high diagnostic accuracy for common systemic diseases and high-risk diseases, such as infectious mononucleosis, chickenpox, rosacea, influenza, hand, foot and mouth disease, bacterial meningitis and so on.
This reveals that the diagnosis system can judge common pediatric diseases with high accuracy according to the clinical data information marked by NLP system.
The researchers then used 1 1926 clinical cases to compare the diagnostic level of pediatric diseases between the AI diagnostic system and five clinical treatment groups, among which the treatment groups participating in the study gradually increased their clinical working hours and qualifications. The results show that the average score of the AI diagnostic system F 1 reflecting the comprehensive performance of the model is higher than that of the treatment group composed of two young doctors, but slightly lower than that of the treatment group composed of three senior doctors.
The paper believes that this shows that AI diagnosis system can assist young treatment teams to diagnose diseases and improve the team's diagnosis and treatment level.
The system was put into clinical application in Guangzhou Women and Children Medical Center from June 5th to/kloc-0 to October 6th this year. From 65438+ 10 month 1 to 65438+ 10 month 2 1 in just 20 days, doctors in this hospital actually called it for auxiliary diagnosis for 30276 times, and the coincidence rate between diagnosis and clinic reached 87.4%. Sun Xin, director of the medical department of Guangzhou Women and Children Center, said after experiencing the system that the system is "more scientific" in disease grouping and classification.
After the publication of the above paper, The New York Times commented on this research, saying that "after the pediatric hospital visited the data of hundreds of thousands of children in China for 18 months, such a huge amount of data can be used for research, which is also China's advantage in global artificial intelligence and competition."
"Data is indeed one of the core keys of our research results." Ni Hao said, "However, high-quality standard data comes from a strong joint team. We have specially developed a data standard system and marked a lot of data. "
Xia Huimin, one of the authors of the paper and a professor at Guangzhou Women and Children Medical Center, said that the inspiration of this article is that "AI will diagnose more diseases by systematically learning text medical records". However, he warned that there is still a lot of basic work to be done, such as the integration of high-quality data is a long-term process.
The author understands that in recent three years, the hospital has paid attention to the standardization and structuring of data, and realized the mutual communication and interconnection of more than 50 diagnostic data subsystems, laying the foundation for the application of the system.
"In addition, after I learned a lot of data, the accuracy of its diagnosis results needs a wider range of data to verify and compare." Xia Huimin said.
Among the four elements of AI technology, the scene is also very important. Zhang Kang, another reporter of the newspaper, believes that it is of great significance to study pediatric diseases.
"The diagnosis of pediatric diseases is a major pain point in medical care. Some pediatric diseases are threatening and need to be treated as soon as possible, and children are just not good at expressing their illness, so it is very necessary to diagnose pediatric diseases quickly and accurately. " Zhang Kang said that the demand for pediatricians is in short supply at present, and the AI diagnostic system constructed in this paper will greatly assist the seriously insufficient medical resources.
Related paper information: doi:10.1038/s 41591-0/8-0335-9.