.jpg)
A recent test by a company called Kangfuzi said that their symptom diagnosis system has hit more than 90% of the typical symptoms of common diseases, more than 10% of the symptom checker of a top international medical clinic. At the same time, according to relevant data, the average misdiagnosis rate of various hospitals in China is around 30%.
Based on this, we contacted Mr. Zhang Chao, the founder of Kangfuzi. According to reports, from the technical point of view, its intelligent diagnosis is trained in three major steps.
Knowledge extraction: similar to doctors' memory medical knowledge. The primary task of medical AI is to build a medical knowledge map to support a range of applications. Knowledge maps have always been a must for major AI companies. Compared with semi-structured sites such as Wikipedia and Encyclopedia, Google and Baidu have built knowledge maps. Kangfuzi is based on unstructured text extraction to build knowledge maps. . Because of the complexity of medicine, its knowledge still exists in unstructured texts such as textbooks, essays, and popular science articles. According to reports, Kang Fuzi has internationally leading original technology in knowledge extraction: 1.) Their system can automatically learn from the massive literature to write the "law" of a certain knowledge and then carry out two large-scale automatic extraction; A set of high performance computing frameworks was designed to mitigate the complex calculations in 1..
Knowledge indicates that similar doctors are accumulating medical experience. After getting structured medical knowledge, Kang Fuzi wants to make this knowledge capable of reasoning. Here are divided into two aspects, 1.) Knowledge vectorization representation, based on deep learning technology, vectorize the disease and symptoms, and then achieve some reasoning behavior, such as the patient said "stomach is uncomfortable", the system will interact, is "disgusting" ", "stomach acid" or "stomach bloating"? 2.) The weight of the knowledge relationship is expressed. Many traditional probabilistic statistical models are based on independent hypotheses, but in reality this is unreasonable. For example, when deriving a possible disease based on a set of symptoms, the evolutionary logic between symptoms must be considered.
Logic application: Similar doctors are doing consultation services. Restricted and medical complexity and knowledge barriers are not just patients, sometimes even doctors can't think about it. At this time, the system needs to be able to interact intelligently, analyze the patient's condition, and make intelligent questions to obtain more patient characteristics.
When talking about the difference between the effects of traditional giants, Zhang Chao explained that the traditional diagnostic thinking quickly reached the bottleneck in effect, mainly because:
The scale of knowledge: Most of the traditional diagnostic thinking uses the manually edited knowledge base. The knowledge base is small. Kangfuzi uses the international advanced information automatic extraction technology, and the number of diagnostic knowledge has exceeded 5 million;
Knowledge update: Here is an example. The typical symptom of lobar pneumonia is “rust rustâ€. This knowledge was discovered decades ago and is widely used in various clinical knowledge bases, but in today's clinical cases, large Leaf pneumonia often does not have the symptoms of “rust rust†because a significant proportion of patients take antibiotics (such as amoxicillin) early, causing rust stains to be inconspicuous or not present; traditional diagnostic tools are in the information The update is lagging behind, resulting in a decline in diagnostic results.
Knowledge reasoning; traditional diagnostic thinking mostly adopts the structure of decision trees. For example, patients complain of symptoms of “toothache, cough, headacheâ€. When traditional diagnostics see “toothacheâ€, they basically go to the diagnosis of periodontal disease. And Kang Fuzi intelligent diagnosis can be based on the knowledge base to reason that this "toothache" is caused by a cold, and then give a correct diagnosis.
Knowledge indicates: If the clinical tools need to be "grounded", then we must work hard on natural language processing, so that the machine can understand the diversity of the user's description, which is also a problem that traditional diagnostic tools have to overcome.
According to reports, the current Kang Fuzi has learned nearly 10,000 medical books, 20 million medical papers, and built a knowledge map for nearly 10,000 kinds of diseases, thousands of symptoms, laboratory indicators, medication response and other knowledge.
Lipstick,Red Lipstick,Loreal Lipstick,Lipstick Set
Guangzhou Lingxue Cosmetics Co., Ltd , https://www.gzlxgj188.com