In the current medical environment, the word "big illness" is like a bomb hanging around the waist. It has the power to destroy a family and can be detonated at any time.
Nowadays, artificial intelligence technology is widely used in various industries, and the medical and health field is one of the important application scenarios. According to statistics, by 2025, the total value of the world artificial intelligence market will reach 127 billion US dollars, of which the medical industry will account for 1/5 of the market size.
The industry believes that artificial intelligence will be a good medicine in terms of the improvement of medical standards and the sinking of medical resources.
Under such a background, what kind of problem can be solved by "artificial intelligence", how to use it, when it can be used, and why it can not be used now becomes a topic worth exploring.
We consulted a number of industry professionals. In summary, the current medical artificial intelligence is in a state of "promising prospects and difficult progress", and some applications are already coming.
In this article, we have compiled the opinions of experts, hoping to summarize the current face of the medical artificial intelligence industry to some extent. The following is the full text, Enjoy it.
What kind of problem can the "artificial intelligence" drug solve, how to use it, when can it be used, and why can't it be used now?
In general, the application of artificial intelligence in the medical field can be divided into six sub-areas - virtual assistant, medical record and literature analysis, medical image-assisted diagnosis, diagnosis and treatment prediction, drug development, and gene sequencing. In this salon, the experts' discussion is mainly concentrated in the first four areas.
Virtual assistant - question and answer can not talk, can only do multiple choice questions
In general, the virtual assistants in the medical field and the virtual assistants in the general sense are the same in terms of mission objectives - solving some problems through dialogue between people and machines. However, to be careful, it is also different.
The official definition of the medical virtual assistant is to use the speech recognition and natural language processing technology to compare the patient's description of the disease with the standard medical knowledge base, so as to complete the information system of the patient's self-diagnosis, consultation, consultation and other services.
Unlike general-purpose virtual assistants such as Siri and Cortana, when users talk to a universal virtual assistant, they can express themselves freely, and the virtual assistant understands the user's intentions (of course, the understanding needs to be strengthened); but when the user talks with the medical virtual assistant Because the patient's description is basically not a standard medical term, it is difficult to compare it with a standard medical knowledge base to draw conclusions.
"At present, the common practice in the medical industry is to communicate with users in a multiple-choice way to understand problems and to be divided." Zhao Guangguang from China Information and Communication Research Institute said, "Some of the products of the University of Science and Technology are currently in certain The hospital has actually been used."
Zhao Sunshine is the leader of the Artificial Intelligence Working Group of the Internet Medical Alliance of China's ICT, and is also the leader of the recently published White Paper on Medical Artificial Intelligence Technology and Applications.
Shanghai Senyi Medical Technology Co., Ltd. focuses on the combination of artificial intelligence and medical treatment. CEO Zhang Shaodian introduces Senyi's medical virtual assistant products. "I don't really want to call our product a chat bot. It's actually a search engine. The people who do the technology know the level of the chat bot." Zhang Shaodian said.
Senyi has cooperated with Shanghai First Maternal and Child Health Hospital and Shanghai Children's Medical Heart Center to try artificial intelligence virtual assistants. The solution is to push the content from the expert knowledge base to the patient after identifying the patient's problem and give the source of the answer.
"Where is the use of this thing?" Zhang Shaodian said, "When a patient has a problem, there is a general situation where you don't believe in Baidu but can't find an expert. With such a virtual assistant, it gives you the answer. The literature written by experts can play a certain role."
Medical records and literature analysis - helping doctors improve efficiency
When it comes to the combination of artificial intelligence and medical care, the most common is the number of doctors who input electronic medical records by voice. The voice input technology for medical scenes has become the land of artificial intelligence companies such as Keda Xunfei and Yunzhisheng.
"Voice input technology frees the doctor's hands, which is especially important for dentists." Zhao Guangyang said, "The dentist is often a person on the operating table, both hands are occupied, no hands to write medical records. The way of speech recognition can make some basic records of the patient's basic information and surgery, and improve the efficiency of doctors."
While liberating doctors' hands, electronic medical records have also served as a cornerstone of the development of medical artificial intelligence. Under the level of speech recognition, how to use natural language processing technology to transform unstructured natural language into structured data for subsequent data mining is an important issue.
Zhang Shaodian said that the use of natural language processing technology to transform unstructured data on medical records into structured data is mainly divided into the following steps.
First, identify the named entities in the sentence. Simply put, which words are diseases, which words are drugs, which words are symptoms, and which words are surgical names, that is, the classification of various word categories.
Then, you need to find the association between the semantics, that is, who has modified who, who is bound, who who has denied, etc., that is, defines the linear relationship between words and words.
"Why is semantic relevance particularly important in the medical field?" Zhang Shaodian said, "For example, you know that this person is not hurt enough. You also need to know the location, severity, time, acute and chronic information of the pain. This information is important. ."
In the natural language processing technology in the medical field, it is often necessary to face the situation where the input is not standard. Every doctor has his own medical record writing habits, such as myocardial infarction, some doctors will write myocardial infarction, some doctors will write myocardial infarction, myocardial infarction, and even write English MI (Myocardial Infarction).
For the machine, you must know that these words represent the same meaning when you store them, and the subsequent work can be done. "Otherwise, even the most simple search task can't be done because the keywords can't match." Zhang Shaodian said, "In addition, natural language processing technology can help doctors improve research efficiency. It is important to know that Chinese doctors are very strong. Just need."
Before doing research, a lot of literature search work is needed. Huang Hong, director of the Information Center of Huashan Hospital affiliated to Fudan University, believes that the first step in the application of artificial intelligence in the medical field can begin with doctors looking for documentation.
She said that because doctors are nervous, many times the job of finding documents is done by graduate students. Although there is now a database to look up, you don't have to go to the library to read paper materials, but the literature search is still a very heavy task.
Huang Hong cited such a case. When researchers conduct a study related to child disability, they need to read about 33,000 abstracts. The manual search is time-consuming and labor-intensive. After introducing machine learning technology, the efficiency is greatly improved.
"Now doctors do research, and most of the time is spent on data collection and structuring." Zhang Shaodian said, "That means you are looking for medical records, turning over medical records, and then grabbing the information you need from the case. Use Natural language processing technology can automate this process as much as possible."
"This matter may have little to do with clinical practice, but it is very important for doctors," Huang said.
Medical imaging-assisted diagnosis - reducing the rate of misdiagnosis and misdiagnosis
"The traditional medical industry has structural defects," said Zhao Guangyang.
He believes that the current allocation of medical resources is a 2-8 structure, that is to say, patients are concentrated in the top three hospitals, resulting in only 20% of the time in the top three hospitals to deal with incurable diseases. In fact, 80% of common diseases can go to primary hospitals.
Why do patients have to go to the top three hospitals regardless of illness? In essence, it is out of distrust of the primary hospital. That is to say, because the quality of the doctor's resources is difficult to sink, the patient will have to go to the top three hospitals if they do not have a bed and live in the corridor.
This is one of the typical application scenarios of medical artificial intelligence in Zhao Sunshine's view. "Retinopathy such as diabetes is very suitable for doing at the grassroots level," he said.
Ophthalmology equipment is more than 100,000 yuan, and millions of imports are needed. It is more difficult for grassroots hospitals to purchase so many professional equipment. At the same time, the fundus is also a special one among many organs. Doctors can directly see the blood vessels in the fundus, which provides a breakthrough for the application of artificial intelligence technology. Using artificial intelligence technology, grassroots hospitals can achieve some early screening work, which is now the hot "divisional medical care."
In addition to triage to primary hospitals, some experts believe that triage to patients is also a very promising research direction in the future. "It's not just hospitals that can diagnose, but self-diagnosis is also very important," Huang said.
Huang Hong cited the high incidence of acromegaly in the population as an example to illustrate the importance of self-diagnosis.
Acromegaly, as the name implies, the patient's symptoms are excessive growth of hands and feet, which is a disease caused by abnormal secretion of growth hormone. At present, many patients with acromegaly are in a certain stage when the symptoms accumulate, and they go to the hospital when there is obvious performance. But in fact, the early diagnosis of acromegaly can be done with only one APP that can perform face and limb scans.
"In the early stages of the patient, it is easy to do early screening by analyzing the data on his face, abdomen, buttocks, hands, etc." Huang Hong said.
In the early screening of cancer, artificial intelligence imaging technology can help doctors reduce the rate of misdiagnosis and missed diagnosis, and is very mature. Zhao Yangguang mentioned that the current misdiagnosis rate and missed diagnosis rate of radiology are up to 40%. This is why the diagnosis of cancer and malignant tumors requires the joint advice of multiple physicians.
"The radiologist has a lot of work pressure. Shooting a sequence of images will produce a lot of films. It is easy to miss the diagnosis by human eyes. The image recognition technology can provide a better supplement for the doctor's diagnosis." Zhao Sunguang said.
Early screening with image recognition is very meaningful.
Taking esophageal cancer as an example, the number of new esophageal cancers in China in 2015 was 477,000. For esophageal cancer, early treatment is critical. In the early five years of treatment of esophageal cancer, the survival rate of patients was 90%. In the next five years, the survival rate was less than 15%.
Zhao Sunguang said that the combination of artificial intelligence and medical imaging is as follows: First, the image is extracted from the radiology department; then the image segmentation technique is used to extract the meaningful region of the image; and some image recognition methods are used to preprocess the image, highlighting the image. Effective information; then extract the lesion area using an algorithm; finally pass the data to the model for training.
After training, the model is given a new picture, and the model can automatically mark the location of the lesion.
The ideal is full, but the reality is very skinny. "Everyone knows that the productization of the auxiliary treatment is very complicated." Zhang Shaodian said, "It involves the hospital's treatment process, the doctor's habits, the doctor's own acceptance, the acceptance of the medical industry, and Many issues related to ethics, law, etc."
Forecast of diagnosis and treatment results - early risk estimation
The auxiliary diagnosis of artificial intelligence is not only reflected in medical imaging, but also has been applied in the control of medical treatment results.
Zhang Shaodian introduced two cases. The first case is the cooperation between Senyi and Shanghai Children's Medical Center to establish the best diagnosis and treatment plan for children with congenital heart disease before surgery.
"Our system is capable of establishing an optimal treatment plan including surgery, anesthesia, cardiopulmonary bypass, etc. It also predicts the risk of bleeding after surgery, the amount of bleeding, the duration of stay in the ICU, and the postoperative syndrome. Risks, etc." Zhang Shaodian said, "When doctors need to change the parameters of the surgical plan, the system can automatically calculate the changes in the risk factors after the parameters are modified."
"In fact, our system functions similarly to IBM Watson. But Watson is an imported product, using foreign people's data sets. We use Chinese local data to better match the physical characteristics of Chinese patients." The current model of this system has been trained. At the end, Senyi is looking for a cooperative hospital to try to land.
In addition to the diagnosis and treatment system for pediatric congenital heart disease, Senyi also trained a risk prediction model for anticoagulant therapy using data from 37 municipal hospitals in Fuzhou.
"After the anticoagulant therapy is completed, some patients will be embolized again, and some patients will bleed. For different patients, the possible situation after surgery is completely different." Zhang Shaodian said. What Senmu’s system does is to predict the risk of different reactions in patients after anticoagulant therapy.
Huang Hong believes that while using artificial intelligence to develop treatment plans, it is necessary to define what is a "good" diagnosis and treatment plan. "The best clinical solution does not mean that this patient is the best," Huang said. "There is a saying that you choose to die with dignity or choose to live without dignity."
Huang Hong believes that because each patient's family situation is different, the ability to pay, religious beliefs, etc., treatment options may also be different. Therefore, today's artificial intelligence technology should not only stay in imaging, histology, and patient history itself, but also integrate social data to make the final solution more realistic.
Ten thousand obstacles to data for medical AI
The primary problem facing medical artificial intelligence is still at the data level.
"No matter what the terminal application is, data is the foundation." Zhang Shaodian said. "The data problem is not a technical issue, but a system issue."
Zhang Shaodian mentioned that there are some successful cases in the medical artificial intelligence industry in the United States, but China does not currently. In retrospect, there is a great relationship with the data. "The domestic medical institutions are still in a relatively dispersed state. The data standardization and structure are very low, and they are relatively incomplete. The interconnection between hospitals is not good." He said, "You have no way to get one." Comprehensive historical data of patients."
Mayo Hospital of the United States with 2,800 IT staff has achieved some results in the field of medical artificial intelligence. "When I was in Mayo's exchange last week, I found that the entire Mayo system had only 1200 beds combined. This size is not big in China.
As far as I know, Shanghai Ruijin Hospital has about 1,600 beds. Zhang Shaodian said, "But when comparing the effects of Mayo's 1200 beds and the data collected by our 1600 beds, the artificial intelligence system that was trained last, you will find that the two are not at an order of magnitude."
This means that more data does not necessarily lead to good artificial intelligence, high-quality and high-value data can train good artificial intelligence.
"Now many operations are performed under endoscopes, such as cholecystitis, gallstones, etc. In fact, the mirror data collected by doctors during surgery is carried by a doctor using a hard disk. In fact, the current medical system. Far from realizing dynamic data sharing," Huang Hong said.
Zhao Sunguang believes that the current demand for data standards in the industry is greater than the demand for data.
The artificial intelligence model is based on hardware acquisition data. Take CT equipment as an example. There are 7-8 mainstream CT equipment manufacturers on the market. However, all CT-related models are now built for a set of equipment. If you want to transplant the model to other hospitals, you need to retrain the model, which becomes a bottleneck that hinders the wide application of artificial intelligence technology in the industry.
In addition, Zhao Sunguang also mentioned that in the process of collecting data, the physician's technique will directly affect the effect of the model. Taking an electrocardiogram as an example, sometimes a doctor needs a patient to wear an ECG instrument for 24 hours to monitor the dynamic performance of the patient's heart rate.
At this time, how the patient wears the device, the location of the wire connection, and even the fatness of the patient will affect the final monitoring results. However, after the patient leaves the hospital, the 24-hour wear is invisible to the doctor. In this process, the doctor has no way to control, so the training data is even more useless.
In addition, in the fields of pathology, ECG, etc., each manufacturer basically follows its own proprietary data format. Zhao Sunguang believes that the industry needs to actively convert private formats into public formats in order to accumulate data available to neural networks.
"I believe that the technicians who do artificial intelligence know very well that with good data, the algorithm is really not a particularly complicated matter." Zhang Shaodian said, "Whether you are engaged in artificial intelligence, data analysis, or data mining, at least 80% of the time is spent on data cleaning."
In addition to data problems, there are still problems with the pattern and system of artificial intelligence in the medical industry.
"At present, if artificial intelligence products are sold to medical institutions in a sales manner, it is difficult to achieve both in terms of qualifications and classification of products." Zhao Guangyang said, "For these artificial intelligence products, the future will be researched through hospitals. It is more feasible to land the subject."
In addition, Zhao Sunguang also mentioned legal issues. If the medical artificial intelligence system diagnoses a patient and causes the patient to die, who should bear this responsibility? Just like the Uber car accident that happened in the driverless field not long ago.
"Now medical devices have a classification of Category 2 and Category 3. If artificial intelligence is classified into Category 3, strict clinical verification is required. The state still attaches great importance to this aspect," said Zhao Guangyang.
Starting a journey of a thousand miles
Medical artificial intelligence has just started, facing the future, there are still many problems to be solved.
For example, in medical imaging, the current practice in the industry is to analyze only images, without multimodal fusion. "In the future, we must analyze it in a multi-modal manner," said Zhao Guangyang. "To combine a variety of patient information, such as clinical information, follow-up medical records, etc., to form a comprehensive multimodal system."
In addition, although artificial intelligence imaging technology has been able to achieve 4-6mm microscopic nodule diagnosis, it has already shown better sensitivity to some extent. But in the future, in the diagnosis of nodules, the industry needs to consider more than the size factor, but also need to be able to identify other characteristics including root cause, spur, split, calcification.
Also, current medical artificial intelligence systems lack historical retrospective analysis. That is to say, diagnosis is only performed for a single image, and data of a time dimension is lacking. For diseases such as cerebral infarction, films at different time points are very important for the treatment plan after the date.
As well, there are about 100,000 gaps in pathologists in China, and the period of training a pathologist is very long. This unsolvable problem in a short period of time needs to be alleviated by artificial intelligence technology.
However, the pathology film is much larger than the CT, MRI and other films involved in medical artificial intelligence. Finding tiny lesions in hundreds of millions of pixels is a challenge to algorithms and computational power. In addition, the pathological diagnosis requires not only the observation of local features, but also the joint analysis of the overall features, so the challenge is even greater.
In addition, Huang Hong also mentioned brain-computer interface, targeted therapy, individualized medication, etc., all of which are important positions for the future development of artificial intelligence. As you can see, for medical artificial intelligence, everything is just beginning. The goal is clear, the future is bright, and the road is long.
Our company offers rubber, and specialty seal and o-ring for molding electronic and electrical Connectors, both circular and rectangular types. We have in house capabilities and quick turn around.
Cable Grommets with various Grommet Shapes, Rib Styles, Slot Shapes, Inner Flanges, and Colors - To have access to our 3D Cable Grommet Design Tool
Silicone Rubber Products,Cable Silicone O-Ring,Rubber Seal,Custom Silicone Seal,Waterproofing O-Ring,Tpe Grommet
ETOP WIREHARNESS LIMITED , https://www.oemmoldedcables.com