Using DeepSeek for medical treatment has saved the "AI healthcare".
Interview | Hu Xiangyun
Article | Hai Ruojing Hu Xiangyun
The "big model heatwave" triggered by DeepSeek is still ongoing.
Some people use it for programming, some for fortune-telling, and some for medical treatment.
On February 18, the "AI + Medical" sector in the Hong Kong stock market and the "AI + Pharmaceutical" sector in the A-share market continued to rise. Since February, the stock price of Yidu Technology has nearly doubled, and the stock prices of JD Health and Ali Health have been climbing one after another. The long-silent AI medical track has once again shown sparks.
In addition to serving as a "health consultant" for ordinary people outside the hospital and providing medical advice for common diseases such as influenza and allergies, big models such as DeepSeek have entered hospitals, especially intensive care units (ICUs), emergency departments, and pediatrics.
In the past, the treatment of critically ill and emergency patients depended entirely on doctors; now, the fine-tuned DeepSeek can quickly analyze various types of data, help doctors sort out complex cases, and provide treatment ideas and suggestions.
It is equivalent to doctors having a smart and on-call AI assistant, and this assistant is also "evolving" day and night.
Recently, South China Hospital Affiliated to Shenzhen University, Kunshan First People's Hospital, and the Second Affiliated Hospital of Army Medical University have announced that they have deployed the DeepSeek model; the medical model Med-Go after being connected to DeepSeek-R1 (671B) has also been applied in clinical institutions such as Shanghai East Hospital, and its capabilities have been verified in the ICU environment.
In fact, AI in the medical field has attracted many rounds of public attention and investment booms. From 2016 to 2023, many enterprises have tried to use AI, to solve the problems of insufficient supply of high-quality doctors and uneven regional diagnosis and treatment levels. As a result, many unicorns with a valuation of tens of billions have been spawned.
Many AI products for clinical diagnosis and treatment have obtained medical device registration certificates, but they have repeatedly encountered difficulties in the commercialization links of entering hospitals, patient/medical insurance payment. "The technical barrier is not high enough (intense industry competition), and the clinical demand is not strong enough (lack of payers)", a partner of a medical investment institution summarizes the reasons for the difficulties of previous AI medical treatments.
Indeed, to achieve an excellent AI medical application, all aspects such as technology, products, doctor-patient education, business, regulation, and ethics are essential. Up to now, the understanding and reasoning capabilities demonstrated by big models such as DeepSeek in the medical field have excited the industry.
So, what might AI medical companies that integrate big models like DeepSeek change?
Can the breakthrough extension of the long board of AI technology drive other links to quickly follow up and use the power of silicon-based life to solve the dilemma of "difficult and expensive medical treatment" for carbon-based life?
What has changed after integrating DeepSeek?
In February, medical and pharmaceutical companies such as Yidu Technology, Airdoc Technology, and Zhiyun Health have all integrated DeepSeek to enhance their old business capabilities of medical data insight, AI image diagnosis, and chronic disease management with this "national-level" big model.
One of the Hong Kong-listed medical listed companies that integrated DeepSeek revealed: After integrating DeepSeek, it is carrying out business empowerment and project expansion, and the technical team is very busy, "There are many hospital customers asking us, and the top priority is to implement it."
Another Tencent-affiliated Internet medical company also reported that the technical team is studying DeepSeek and is expected to come up with a detailed plan in two weeks.
Behind the enthusiasm, most technical leaders actually know that DeepSeek may not be called a "disruptive technological innovation", but after "fine-tuning", its excellent performance in reasoning and decision-making scenarios can indeed provide stronger support for their own products in processing complex medical data or supporting precise decision-making.
And this is precisely the urgent need of AI medical companies at the moment.
"We mainly consider its reasoning ability and mathematical ability when integrating DeepSeek-R1 (671B), which can increase the accuracy of medical record diagnosis by more than 10%, especially for the diagnosis of complex cases." Zhang Hanxiang, CEO of Shuole Information, told 36Kr. In November last year, they just jointly launched the AI medical big model Med-Go with Shanghai East Hospital.
Zhang Hanxiang introduced that in the intensive care unit with huge diagnosis and treatment pressure, the medical big model integrated with DeepSeek can quickly analyze various data of patients, including vital signs, laboratory test results, image data, etc., and provide multiple possible diagnosis plans.
"For example, when dealing with a case of a patient with multiple organ failure, doctors face multiple treatment options, including mechanical ventilation, blood purification, and drug treatment. In this case, the fine-tuned DeepSeek can combine the patient's medical history, examination results, and the latest medical research to provide doctors with a comprehensive decision-making support framework."
On social media such as Xiaohongshu, during the Chinese New Year, some parents have used DeepSeek to interpret their children's blood routine test reports and seek medication advice, and finally "DeepSeek gave the same diagnosis as the chief physician of the People's Hospital". When consulting offline, doctors will give a diagnosis result, but the explanation is usually relatively simple. Faced with many professional terms, parents who are concerned about their children often have a series of questions, and the AI medical model can continue to respond.
Of course, the "hallucination" problem of AI also exists in DeepSeek. Although the base big model has a strong reasoning ability, its medical language materials are not professional, and there is a lot of room for improvement in the meticulousness of data processing. When directly used in serious diagnosis and treatment scenarios, the accuracy rate is relatively low, and the medication and treatment suggestions given may mislead patients.
This gives enterprises the space for "deep customization", on the basis of open source, such as applying expert-annotated accurate data sets, the diagnostic thinking chain of the doctor team, etc., to improve the accuracy rate. This is exactly where the accumulation of previous AI medical companies lies.
In addition, the "data privacy and security" problem that has previously plagued AI medical companies and hospitals has also been effectively avoided under the open source + local deployment model of DeepSeek. When the relevant AI agent is implemented in the hospital, it can be privately deployed, which not only solves the problem of privacy leakage from the source, but also saves the cost of cloud services and reduces the enterprise's dependence on third-party cloud platforms.
In addition to diagnosis and treatment, DeepSeek also shows potential in the 2B medical circulation scenario. Shanghai Pharma Yunjiankang, which focuses on the pharmacy retail business, told 36Kr that it has applied DeepSeek to business links such as intelligent Q&A, pharmacist training, and patient personalized operations. For example, in the issue of prescription review, DeepSeek is expected to help pharmacists more quickly identify potential problems such as drug interactions and dose errors in prescriptions, thereby reducing the risk of medication.
Searching for the Medical Version of DeepSeek
AI + Medical has gone through ten years of ups and downs, and many "AI vs. Doctor Competitions" like AlphaGo vs. Ke Jie have been held.
During these ten years, AI products have gradually changed from "toys" to doctors' tools, and then from tools to doctor assistants. With the advent of the technological singularity, smarter "AI intelligent doctors" are getting closer and closer to us.
After entering 2024, the market's value judgment of medical big models has gradually shifted from emphasizing model capabilities to business orientation, that is, finding "killer applications" in the appropriate scenarios. To achieve a medical killer application, it requires the combination of technology, products, doctor-patient education, business, regulation, ethics, and other sectors to form a "deep-water wooden barrel".
The value of DeepSeek lies not only in the significant extension of the long board of AI technology, but also in its strong influence in making a big impact (This is a tentative translation for "出圈" as its exact meaning is not clear), and the "AI usage education" it has conducted for doctors and patients.
Among the doctor group, there are avant-garde doctors who embrace new technologies, and there are also a considerable number of doctors who are more conservative. The capabilities demonstrated by big models such as DeepSeek have made more doctors embrace AI technology. For patients, when AI tools have integrated into daily life and work, it is natural to use AI for health consultation, image reading, and consultation.
Currently, the common conversation bar of Doubao has also added the "AI Health Consultation Assistant" launched by Xiaohe Health, guiding users to have health-related conversations and consultations with it.
However, different from consultation conversations in other life scenarios, medical consultation and cancer screening have high requirements for accuracy, and have low tolerance for "AI hallucinations" and "fabrications". Imagine that if the AI health assistant makes a misdiagnosis, and the user follows its advice to take medicine and experiences certain physical discomfort, medical disputes and liability issues may arise. This not only puts specific requirements on the capabilities of medical products, but also tests the capabilities in "access and regulation".
In addition, enterprises that develop AI + medical applications will eventually face commercialization, and the old problem of "who will pay" is difficult to avoid.
After ten years of exploration, AI products for diagnosis and treatment scenarios have been commercialized, mainly achieving "pay-per-use", so as to bring continuous income to enterprises while meeting the clinical needs of patients. In the past, many AI medical enterprises could provide similar services, and the business model of "one-time software sales fee" in the intense competition has made it difficult for the entire industry to make money.
Of course, the past is just a prologue. With the rapid evolution of technology, what potential will the evolved AI medical products show in hospitals, pharmaceutical chains, and beyond? 36Kr will continue to follow up.