Author Zhang Ling Source Economic Observer
At 3:00 a.m., Beijing Tsinghua Changgeng Hospital (hereinafter referred to as Changgeng Hospital), the lights in an operating room have just gone out, a surgeon did not rest, but took out his cell phone and used DeepSeek, but also forwarded the news of a hospital access to DeepSeek to his colleague Li Dong.
Li Dong, also awake, is the director of the Medical Data Science Center at Chang Gung Hospital and is busy writing a DeepSeek training tutorial, hoping to get doctors up to speed and ensure error-free use.
This scene epitomizes the transformation DeepSeek is triggering in China’s medical community.
On February 23rd, Chang Gung Hospital completed the localized deployment of DeepSeek. Before and after Chang Gung Hospital, major hospitals across the country also joined the race. According to incomplete statistics, as of February 27, nearly 100 hospitals in more than 20 provinces have announced the completion of the localized deployment of DeepSeek, exploring its application in clinical diagnosis and treatment, medical records quality control, image analysis, scientific research, hospital management and other scenarios. This list is still getting longer fast.
China’s healthcare sector is undergoing a reshuffle. According to Li Dong, hospitals have access to DeepSeek, which will force hospitals to improve efficiency and data governance capabilities, and hospitals with low efficiency and procrastination and lack of forward-looking layout will be gradually out of the game.
Doctors are both excited and apprehensive. In the outpatient clinic operating room laboratory, they embrace DeepSeek, it is really a good assistant In social media In the exchange with peers, there are also people who are afraid of DeepSeek Doctors will really be replaced?
China’s top neurosurgeon once described the state of mind when exploring the human brain with a scalpel as walking on thin ice, as if facing an abyss, and going all out to do his best. Li Dong said that when doctors embrace AI, the feeling of walking on thin ice even exceeds that of doing brain surgery.
The doctor’s fear is very simple Surgery is only facing a specific lesion, the opposite of AI, is hundreds of millions of living patients.
And DeepSeek narrowly.
Not long ago, Zhang Hongwei, director of neurosurgery at Sanbo Brain Hospital of Capital Medical University, encountered a patient with an ultra-complex brain tumor in the outpatient clinic. The patient had already undergone two craniotomies, one transnasal surgery, one radiotherapy, and one Gamma Knife treatment, and the tumor was still uncontrolled.
Zhang Hongwei sent the patient’s situation to DeepSeek and asked what it should do next. Just 18 seconds later, it gave a clear plan, with everything Zhang Hongwei could think of in the literature, including the latest clinical trials.
Zhang Hongwei was shocked, and on February 22, he posted the experience online, which sparked a lot of empathy. He believes that for more complex problems such as brain tumors, DeepSeek is at least equivalent to provincial tertiary hospital experts. Since then, he has often used DeepSeek to access clinical problems.
On the same day, a Guangdong medical blogger, Lonely Fame, also hit the hot search with a tweet The sky is falling! The patient questioned my treatment plan after DeepSeek, and I was so angry that I checked the guidelines again, only to find that they had been updated.
Several interviewed doctors told the Economic Observer that in the past, peers met mostly to discuss cases, but now, more and more people are discussing whether DeepSeek is good or not How to use it. In peer groups, discussions about DeepSeek always make doctors hit the forward button.
At the cadre meeting on the next day after the completion of DeepSeek’s localization and deployment, hospital leaders set aside 40 minutes for Li Dong to explain DeepSeek’s application scenarios. As soon as the meeting was over, seven or eight people asked him for the full slides, and four other departments asked him to go back to the department to explain in detail. The hospital office simply sent the slides to the whole hospital, and the demand was great.
Early the next morning, Li Dong was surrounded by several doctors in the cafeteria and asked How to write prompts to let AI optimize research bids? How to access the latest international treatment programs? Breakfast took more than 20 minutes, and Li Dong was answering questions the whole time.
Over the past few weeks, the technical leader has been accustomed to late-night thinking Access to DeepSeek and secondary development Nanny tutorials for doctors At least two times a week, doctor training Li Dong has not been able to rest before 1:00 a.m. for a long time.
In-hospital DeepSeek cost account
During the Spring Festival, Li Dong submitted a preliminary access feasibility report to the hospital within a week of the DeepSeek public beta release. After resuming work, Chang Gung Hospital immediately met to discuss how to access it.
Accessing DeepSeek is a systematic project that involves a lot of resource deployment. The arithmetic is very energy-consuming and cannot affect the medical power supply. In order to ensure the safety of the power supply, the hospital has specially remodeled the floor circuits. The server room should be located in a higher position, taking into account the building load-bearing data security, ventilation and heat dissipation, which requires coordination with the logistics and finance departments.
There are three main ways to use DeepSeek Direct access to applications Calling cloud service interfaces Localized deployment Customized development. Hospitals using DeepSeek need to localize the deployment of custom development. Li Dong introduced, on the one hand, the hospital data can not leave the hospital On the other hand, to ensure that patients get more professional advice, you need to use the hospital professional real medical data to train, while ensuring that the data is not leaked.
Localized deployment requires DeepSeek installed to the local server, and then disconnected from the network to use. Li Dong said that if open-source, exogenous data will be mixed with the hospital’s data, and the reliability of the consultation cannot be guaranteed.
After completing the localized deployment, the answer to using DeepSeek in the hospital must be more professional. Li Dong said that in addition to more professional, but also customized, such as the development of some features tailored to the key departments. The secondary development can make DeepSeek more powerful and flexible, and combine it with the medical scene more deeply.
In the process of large model deployment, training and secondary development, the importance of doctors is emphasized. Software engineers are responsible for fine-tuning the direction of the parameters that match the medical scenarios, the hospital information system provides high-quality data resources for training, and most critically, the doctor has to teach the machine to optimize itself by explaining the specific diagnosis and treatment logic and decision-making process; otherwise, the secondary development will be difficult to succeed. This process requires time, experience accumulation and repeated professional-level polishing of large models, so much so that some people jokingly refer to this process as How much artificiality is how much intelligence .
Access to DeepSeek is not difficult, if the data engineers are skilled enough, the fastest deployment can be completed in an afternoon. But after accessing DeepSeek, there is still a long way to go to make it reach the standard of medical grade application What part of data should be mobilized for training? How to do a good job of data governance? How much arithmetic power is needed? How much funding and manpower is needed? How to make the model can really understand the diagnosis and treatment logic and decision-making process?
Li Dong said that the most complex part of the two points First, to debug the hospital customized version of the generic version, although easy to access, but because of the inability to adapt to the complexity of the medical scenarios, it is difficult to give highly professional and accurate answers Second, after the deployment of the big model to the pre-hospital prevention, in-hospital diagnosis and treatment of the hospital and the post-hospital management, as well as community rehabilitation and other full-cycle scenarios is not an easy task.
Embracing DeepSeek, hospitals have to calculate the cost account.
The full-blooded version of DeepSeek R1 model parameters amount to 671B, only the model file requires 404GB of storage space, and 1300GB of video memory is required for operation. Peking University First Hospital Hereinafter referred to as Peking University Hospital Information Center Director Zhou Guopeng revealed that the deployment of full-blooded version of the hardware equipment alone will cost 1.5 million 3 million yuan, the hospital is generally completed at the end of the year to declare the budget for the following year, and re-reporting the budget is not easy.
Some hospitals with sufficient arithmetic reserves like Chang Gung Hospital deployed the full-blooded version, and many hospitals deployed smaller versions. The 32B version deployed by Peking University Hospital’s hardware investment can be controlled within 500,000 yuan.
DeepSeek is a versatile tool for physicians
Even if the hospital does not deploy DeepSeek, doctors will still use it to write bids. Li Dong told the Economic Observer that the recent end of the National Natural Science Foundation of China declaration, doctors have an urgent need for DeepSeek.
Tertiary hospitals are chronically overcrowded, with doctors’ outpatient visits reaching manpower limits while patients continue to increase. Doctors must transition from an experience-based model to one that combines experience and technology, and DeepSeek can just help.
Zhang Hongwei told the Economic Observer that when checking the room, he often encountered such clinical diagnosis and treatment problems as rare types of pituitary tumor treatment Adenoid cystic carcinoma radiotherapy chemotherapy latest progress. In the past, he would usually let graduate students to search the literature, read it and then summarize it to form a review, and the students could only tell him the relevant results after a few days. Now, he gives this job to DeepSeek, which can complete it in a few tens of seconds with high quality.
In addition to being a doctor outside of God, Zhang Hongwei is also a dean, and he will now use DeepSeek to assist him in managing the hospital. He remembers the shock DeepSeek brought him when it first appeared, he asked I am the president of Kunming Sanbo Brain Hospital, the hospital is opening a new hospital, can you provide a framework plan for the construction of a stroke center in a brain hospital? After 27 seconds, DeepSeek clearly set out the process plan, including quality control indicators, framework, construction timeline, etc., and the overlap of the plan after consultation and communication with experts could reach 80%.
For most doctors, writing medical records can save a lot of work when the hospital is connected to DeepSeek.
In the past, doctors are often in the medical record sampling due to irregular writing sentenced to failure, now DeepSeek can be automatically prompted before the submission of the problem, not up to standard can not be submitted. What is more surprising to doctors is that the consultation information can be intelligently filled into the standard medical record template, and even compressed the workload of one hour to half a minute.
There are also hospitals that are cautious about letting DeepSeek help write medical records. Zhou Guopeng said that medical records are mainly written by residents in training, and many teaching experts are worried about AI interfering with training, and believe that the process of writing medical records is conducive to deepening the memory of the development of clinical thinking The importance of this thinking to diagnosis, as muscle memory is to play the piano.
In Zhou Guopeng’s eyes, it is a waste to use DeepSeek only as a question and answer system, and it is necessary to explore practical application scenarios through secondary development.
In the past week, an intelligent decision-making system for clinical application of innovative drugs developed by Beida Hospital based on DeepSeek is being applied on a pilot basis. This system integrates clinical data, such as electronic medical records, with a knowledge base of medication, and through dialogue and interaction, it can provide doctors with scientific and accurate medication recommendations, especially in the treatment of complex diseases, such as tumors and rare diseases.
Zhou Guopeng explained that although China encourages the use of innovative drugs, the actual utilization rate is low, firstly, the knowledge rate of new drugs is low, and secondly, there are many indicators to determine whether a patient is suitable for the use of a certain innovative drug, which many doctors do not understand enough, so the intelligent decision-making system for the clinical application of innovative drugs is equivalent to a good consultant. In addition, the system is docked with the clinical trial system of Peking University Hospital, which can accurately match suitable clinical trials based on patients’ diagnostic results and pathology information, and promote the clinical research and translational application of innovative drugs.
DeepSeek is also useful before surgery. Li Dong takes hepatobiliary surgery as an example, when liver cancer is resected, it is necessary to maximize tumor removal while preserving liver function. In the past, how much to cut depends on the doctor’s experience, and every cut is a challenge. Now, with the intelligent diagnostic system, doctors can accurately input all the patient’s data, and DeepSeek can quickly call up all the treatment methods in the world, and select the best among the best to help doctors determine the precise surgery plan based on individualization. DeepSeek is also useful in helping hospitals reduce costs and increase efficiency and helping medical students acquire new knowledge. Li Dong introduced, for example, a certain drug in the past into the 1000, DeepSeek accurate calculation may only need to enter the 750, which reduces the inventory cost Many medical textbooks knowledge is outdated, DeepSeek can instantly push the electronic version of the latest treatment program.
Ordinary people use DeepSeek to get the answer although there will be no significant deviation, but the professionalism may be slightly weaker, after the hospital localized deployment of DeepSeek will be more professional. Li Dong expects that with DeepSeek as the representative of the big model access, the future will be a lot of hospitals in the outpatient hall set up ports for patients to experience the hospital version of DeepSeek patients queuing up before and after the consultation can go to DeepSeek for a while .
Can it replace doctors?
For many doctors, a soul-searching question is Can DeepSeek replace doctors?
According to Li Dong, whether it can replace doctors depends on whether DeepSeek can reach or exceed the level of expert doctors in disease judgment, and whether it has the ability to take responsibility, the latter depending on whether it has a sense of autonomy.
Li Dong judges that doctors who cannot use modern technology to improve their diagnosis and treatment in due course are likely to be eliminated. As for whether DeepSeek can completely replace doctors in the future, it is still a very serious question, just like whether AI can completely replace human beings .
Primary care doctors will definitely have a sense of crisis, and if they don’t keep up with the times, they will be robbed of their jobs by AI. Zhou Guopeng has a similar view.
When outpatient clinic, Zhou Guopeng met patients who came to the clinic with DeepSeek’s suggestions, and he found that these suggestions are often more scientific. He feels that the widespread use of DeepSeek in the medical field can improve the standardization of diagnosis and treatment, and reduce overmedication and even medical corruption. For example, if a doctor uses a treatment plan that doesn’t make enough sense, the patient is likely to discover the problem with the help of DeepSeek.
The Economic Observer has learned that some patients have recently complained about hospitals with imperfect and flawed treatment plans in the past.
In Zhang Hongwei’s view, DeepSeek can not replace the doctor, but can become a good assistant to the doctor, a good bridge between doctors and patients.
Zhang Hongwei patients with DeepSeek to the analysis of the clinic to ask the doctor as a good thing, which will allow patients to better understand their own disease state and treatment, the doctor to save a lot of popularization of the explanation of the work can improve the efficiency of the clinic. Doctors do not have to worry about being replaced, because DeepSeek is often based on international and domestic diagnostic and treatment guidelines, literature, cases to give conclusions, and a good doctor will be guided by the guidelines, combined with many years of experience in comprehensive judgment and analysis, to make a personalized treatment plan for patients. In addition, patients do not have professional knowledge, the way of asking questions may not be accurate, the conclusion may be biased, and can not judge the accuracy of the conclusion by themselves, but also need to go to the doctor for further consultation.
Zhou Guopeng also suggests that patients can usually use DeepSeek more access to disease science, but in an emergency, should go to the emergency room as soon as possible, because once the AI judgment error, it may make the patient’s life is in danger.
Can DeepSeek replace doctors, there is another issue that can not be ignored Policy red line. Zhou Guopeng told the Economic Observer that, at present, AI can not directly carry out diagnosis or issue medical prescriptions, which involves medical ethics. Whether this ethics will be broken in the future is uncertain and depends on the attitude of the legislature.
As early as 2022, the National Health Commission stipulated that artificial intelligence software and other artificial intelligence software shall not be used to replace the physician himself to provide diagnostic and treatment services Strictly prohibit the use of artificial intelligence and other automatic generation of prescriptions .On February 11, at the same time as a number of hospitals competed for the layout of DeepSeek, the Hunan Provincial Bureau of Medical Insurance also issued a document to stipulate that the use of AI to automatically generate prescriptions is strictly prohibited.
END
For more information, please visit Finance World Magazine.
Leave a Reply