Per reporter: Chen Xing, Wang Jiafei
Before the advent of generative AI technology, AI medical treatment was more of a "fantasy". Often after a series of multiple-choice questions, patients can get a plausible diagnosis, even less than 5 minutes after going to the hospital to see a doctor in line.
But generative AI offers a glimmer of possibility for the world in this blueprint. In May this year, the Internet medical platform "Medical Union" developed the first medical big language model in China - MedGPT (Med refers to medicine, GPT is a pre-trained language model based on the Transformer network structure). Wang Shirui, founder and CEO of the Medical Federation, said in an exclusive interview with the "Daily Economic News" reporter that this model may bring new answers to solve the "impossible triangle" of medical care (the medical system is difficult to improve the quality of medical services, increase the accessibility of medical services and reduce the price of medical services at the same time).
On June 30 this year, the Medical Federation set up an offline free clinic in the hospital, conveying the patient's complaints to real doctors and AI doctors respectively, and completing the whole process of billing diagnosis and issuing treatment plans. Finally, the expert scoring results show that the agreement between AI doctors and top three attending doctors in the score results reaches 96%.
Generative AI understands the patient
Before the advent of generative AI, AI medical treatment was a wish for "beautiful imagination and skinny reality". In 2018, the Internet medical platform Medical Alliance applied AI technologies such as natural language processing (NLP) and computer vision (CV) to land a series of medical application scenarios. These include smart health terminals, smart triage, etc. - this is the "prototype" of intelligent medical assistance.
But at that time, the attempt to intelligent medical assistance "failed". Since the collection of diagnostic decision-making information at that time was mainly completed through multiple-choice questions, the lengthy content of forty or fifty items did not allow users and doctors to pay. The medical association team realized that AI may be the door to open the door to the time, space and manpower constraints of medical services, but they couldn't find the key to open this door.
When the team was frustrated, ChatGPT appeared. The big language model is like the "last piece of the puzzle", and the medical association team developed MedGPT, which was officially released in May this year.
In an interview on July 20, Wang Shirui described MedGPT as follows: "In all the reports and academic journals we see, MedGPT (the value of appearance) should be: for the first time, an AI doctor can conduct multiple rounds of interrogation and differential diagnosis of patients like human doctors, issue test opinions and read reports, and finally give accurate treatment plans." This process mimics real scenes throughout the process, which is not only the first in China, but also the first in the world. ”
"I'm immersed in this every day now because that's the future." Wang Shirui couldn't hide his excitement.
But before the advent of generative AI, the natural and smooth AI disease diagnosis and treatment process was once regarded as a bottleneck that was difficult to break through.
"One of the most difficult points should be the understanding of the real natural semantics of human beings by AI technology." Wang Shirui said. If the natural speech processing is not good, the machine cannot understand the patient's words, and can only rule out possible causes through lengthy multiple-choice questions, and finally make a suspected diagnosis, and the user experience will be greatly reduced. In addition, even if a machine can read human language, it may not be able to directly do elimination like a real doctor, using intuition and experience to make judgments and inferences, a process we call the chain of thought. How to find a most likely diagnosis from several possibilities is a very big problem.
The innovation of the underlying technology first broke through the difficulty of "reading", and after the large language model based on the Transformer architecture came out, the problems related to the communication ability and recognition ability of natural language were naturally solved. But in healthcare, accuracy and consistency in the application of AI technology is critical. To improve this, more algorithms and instructions are needed to require machines to converge judgment and avoid misdiagnosis or overdiagnosis. "For example, medical test reports may have images, texts, including different doses, units, symbols, involving pathology, biochemical indicators, how to accurately read the test report involves a large number of unified standard work. And the medical standard guidelines are constantly updated, so this puts forward high requirements for timely updating of the database. Wang Shirui said.
It is understood that MedGPT is mainly composed of two systems: a large model system and an expert system. Among them, the large model completes 60% of the work, and the expert system completes the final 40%, so as to continuously optimize the accuracy and effectiveness of medicine. So far, MedGPT still tests more than 3,000 cases every month, and then 100 human doctors score the diagnosis of each case. "I receive more than 2,000 pieces of feedback every month."
"These feedbacks include, but are not limited to, the system asks too much, the questions are not targeted, the drugs that are not available in China are mentioned, the patient's drug preference is not noticed, whether the test must be done all at once, and so on." Now we think MedGPT still has a lot to optimize. Wang Shirui said frankly.
On June 30 this year, the Medical Federation did such a thing - an offline free clinic was set up in the hospital, where the doctor's assistant communicated with the patient face-to-face, and the patient's main complaint was conveyed to the real doctor and the AI doctor respectively, and after multiple rounds of communication, the "doctors" issued a checklist or diagnosis for the patient, and the patient returned to the clinic after completing the examination on the spot, and then the AI doctor and the real doctor provided clinical diagnosis and treatment plan.
Finally, 7 experts and professors from Peking University People's Hospital, China-Japan Friendship Hospital and other hospitals scored these valid cases from multiple evaluation dimensions. The results showed that the comprehensive score of real doctors was 7.5 points, and the comprehensive score of AI doctors was 7.2 points - the agreement between AI doctors and top three attending doctors in the score results reached 96%.
"If measured from the four dimensions of disease coverage, intelligence, accuracy and sales, the scores of MedGPT should be 9, 6, 8.5 and 9 respectively." Wang Shirui introduced.
Machines will never care more about humans than they do
William Kissick, a professor at Yale University in the United States, has proposed the "impossible triangle" theory in the field of health care, which states that under established constraints, it is difficult for a country's medical system to simultaneously improve the quality of medical services, increase access to medical services and reduce the price of medical services.
But in Wang Shirui's view, generative AI may provide a solution to the "impossible triangle".
He believes that generative medical AI can receive tens of millions of patients, anytime, anywhere, and unlimited supply. With continuous training, the level of medical AI will increase on a monthly basis, "Now we think that MedGPT has basically reached the level of doctors with 10 to 15 years of clinical experience, and in the future, it may improve clinical experience every month for one to two years." "As a result, access to and quality of health services can be balanced with the help of AI as a tool.
In terms of cost, Wang said that the cost of completing a complete consultation process with MedGPT now does not cost more than $1, and the cost will be halved every 18 months thereafter.
For internet healthcare, generative AI brings new possibilities. Wang Shirui said that Internet medical care is about to enter the era of digital medicine. In the initial state, Internet medical treatment used the platform as a link to achieve the aggregation and distribution of information, but the help to the "impossible triangle" was limited. The outbreak of AI has changed Internet medical care from a linker to a creator of productivity, an original productivity that can cover more patients, which can truly solve the problem.
But one problem that AI has always been unable to avoid is the relationship between AI and people. In the medical field, what is the relationship between AI technology and human doctors?
Wang Shirui believes that there are two levels of human doctors that can never be replaced.
The first level is that only human doctors can do really in-depth research. To make medical progress, it depends on solving incurable diseases, rare diseases, and emerging diseases, and with the current level of AI technology, real human medical experts are also needed to set the algorithm rules behind it. Clinical guidelines, case studies, this must be done by human medical practitioners. But in turn, AI can help human medical experts quickly collect disease cases and data, and bring them together for expert reference and overcome.
The second level is for a large number of grassroots and young medical workers, and medical AI can play the role of database and knowledge base.
Wang Shirui introduced: "It takes many years to train a qualified doctor, and it takes decades from undergraduate to doctoral and standardized training, and it takes thousands of hard work to reach a qualified level. With the help of medical AI, young doctors can grow up quickly and improve the rate of diagnosis and treatment. ”
At the end of the interview, Wang Shirui said: "Machines will never care more about humans than humans. The machine can play the role of a powerful assistant, but the one who finally confirms the diagnosis and treatment plan and signs the plan must be our qualified professional doctors, who have to cover the patient."
Reporter|Chen Xing, Wang Jiafei
Editor|Wendo
Video Editor|Han Yang
Visual Design|Shuai Lingxi
Coordinating Editor|Yi Qijiang
Daily economic news