The medical large model Med-Go under "Shuo Le Information" is integrated into DeepSeek, and the accuracy rate of medical record diagnosis is increased by 10%. | Early-stage project

After the explosion of DeepSeek in popularity, it seems that competing medical enterprises' rush to deploy it has become a top priority since the resumption of work after the Spring Festival. According to incomplete statistics from 36Kr, since the resumption of work after the festival, at least more than 20 enterprises in the medical field have publicly announced that they are introducing DeepSeek. Among them, there are traditional pharmaceutical companies such as Hengrui Medicine and Yunnan Baiyao; established IVD companies such as KingMed Diagnostics and Sansure Biotech, but more active are various AI medical concept companies, such as Zhiyun Health, Airdoc Technology, Yidu Tech, and Shukun Technology, etc.

In the communication with many enterprises that have already connected to DeepSeek, we learned that, on the whole, DeepSeek may not be called a "disruptive technological innovation", but after fine-tuning, its excellent performance in reasoning and decision-making scenarios can already provide stronger support for each company's own AI products in processing complex medical data, or supporting precise decision-making and other aspects.

Shuole Information is one of the medical enterprises that connected to DeepSeek earlier. CEO Zhang Hanxiang introduced that in November last year, Shuole Information just jointly launched the AI medical large model Med-Go with Shanghai East Hospital.

Before the model was released, the Shuole Information team selected a total of more than 6,000 domestic and foreign medical textbooks to "feed" Med-Go, "and now it has reached a basis supported by 20 billion high-quality medical data". Currently, the medical model Med-Go after connecting to DeepSeek-R1 (671B) has also been applied in clinical institutions such as Shanghai East Hospital, and its ability has been verified in the ICU environment.

The following is the more detailed content (edited) that Zhang Hanxiang shared with 36Kr:

Q: What are the considerations for Med-Go to connect to DeepSeek, and what effects are expected to be achieved?

Our main consideration is its excellent reasoning ability and mathematical ability. Combined with the Med-Go medical knowledge base, it can better exert the value of medical data. For the DeepSeek of the Med-Go version, we have conducted a large number of case tests with the medical team, The accuracy rate of medical record diagnosis has increased by more than 10%, especially in the diagnosis accuracy of complex cases, which is particularly obvious.

Q: Please combine specific scenarios to talk about your experience in using the DeepSeek of the Med-Go version?

Currently, Med-Go has been used in hospital clinical practice. From the usage of ICU doctors, the DeepSeek of the Med-Go version in the "deep thinking mode" has greatly enhanced the doctors' ability to comprehensively analyze the patients' conditions, especially playing an important role in the differential diagnosis of complex or multiple diseases.

Because in the ICU, when doctors treat critically ill patients, they often have to make key decisions in a short period of time. We once encountered a patient with multiple organ failure, whose condition was complex and changed rapidly. Doctors faced 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 a comprehensive decision support framework for doctors.

In the diagnosis process of some critically ill patients, it can quickly analyze the patient's various data, including vital signs, laboratory test results, imaging data, etc., and provide multiple possible diagnosis plans, which helps doctors shorten the decision-making time and improve the accuracy of diagnosis.

In addition, DeepSeek can predict the impact of different treatment plans on the prognosis of patients, helping doctors choose the most suitable treatment plan for the patients. It can also dynamically adjust the treatment recommendations according to the real-time data of the patients to ensure the timeliness and effectiveness of the treatment plan. In this way, DeepSeek not only improves the decision-making efficiency of doctors, but also reduces medical risks and improves the treatment effect of patients.

Q: For medical enterprises, what might be the differences between using DeepSeek and previously deploying other generative AI large models?

The greatest value of DeepSeek lies in providing a "open source + localized" strong reasoning model solution for the medical industry. This enables enterprises to more independently control technology and data, and at the same time be able to conduct in-depth customization according to specific business needs. This not only solves the concerns about data security, but also provides possibilities for infinitely diverse AI application scenarios in the future.

Q: One of the core contradictions in the application of large models in the medical field is data and privacy. In this regard, how do you view the performance of DeepSeek? What promoting effects might it have on promoting the progress of AI medicine?

DeepSeek provides a practical and feasible approach for the implementation of large models in the medical field. Through the local deployment and open source transparent model, the open source of DeepSeek effectively alleviates the concerns of medical institutions about privacy and compliance, while enabling more application scenarios and innovation opportunities, thereby comprehensively promoting the progress and development of the AI medical industry.

The Med-Go implemented in the hospital is privately deployed and does not connect to the external network, so the problem of privacy leakage is solved from the source by not allowing the data to leave the hospital.

Q: DeepSeek may be a more excellent and cost-effective tool. From a commercial perspective, do you think DeepSeek will perform better?

From a technical perspective, the open source and local deployment characteristics of DeepSeek mean that medical enterprises can significantly reduce their reliance on third-party cloud platforms. On the one hand, it saves the cost of cloud services, and on the other hand, it also better controls data compliance and security.

For many medical scenarios that highly value user privacy and compliance, this model has obvious cost-effectiveness advantages. More importantly, the open source strategy of DeepSeek allows enterprises to deeply customize the model, and derive more specialized AI applications in multiple business lines such as clinical diagnosis, drug development, and patient management, thereby increasing the market competitiveness of products.

Take Med-Go as an example. We are strongly coupled with doctors. From the selection of corpora to multiple rounds of training and deep thinking, it is distilled based on doctors. Doctors are the product managers of our Agent, while the corpora of the base large model in the medical field is not professional, and the ETL will not be very careful, so the accuracy rate in serious medical treatment will be low. This is the difference in our deep customization. This time, DeepSeek mainly improves our ability in mathematical calculation problems.

Q: Previously, when large models were applied in the medical field, they were often questioned for not being able to find a reasonable application scenario. On this issue, will DeepSeek bring some differences?

The core reason why large models cannot find application scenarios in the medical field is that the medical capabilities of the native large models are not up to standard, and their capabilities have not been recognized by professional groups such as doctors. Now the emergence of DeepSeek has greatly improved the capabilities of the basic model. The DeepSeek model that is deeply optimized by the medical team will definitely be able to effectively solve the pain points in the medical field.