The three-wheel drive of anchoring technology, data and ethics, the medical model that has emerged like mushrooms after a rain in recent years shows that the medical field is becoming one of the commercial blue oceans with the most application potential of artificial intelligence. Industry insiders interviewed said that since the accuracy of the model generation content is directly related to the patient's life safety, and promoting the implementation of medical big models, it is still necessary to enhance decision transparency and interpretability, break the bottlenecks of data integration and quality, and ensure the reliability of diagnostic suggestions. At the same time, more work is needed in terms of privacy protection and ethical norms.

  From "available" to "reliable" gradually solve the "illusion" problem

  "The medical big model has a wide range of applications, and its implementation directions include: being a growth tool for doctors, reducing the training costs of doctors through the use of medical big model; being a patient management tool, and using medical big model to create AI healthy agents directly for consumers; releasing doctors' time and energy, allowing doctors to participate more in the review process rather than full-process operations." said Wang Guoxin, chief scientist of JD Health Exploration Research Institute.

  The “availability” of medical mockups is closely related to the gradual resolution of the “illusion” problem. At the technical level, traditional AI models are prone to "illusions" due to data noise and knowledge blind spots. Once the model deviates when generating content, incorrect inference in medical scenarios will lead to serious consequences.

  In this regard, technicians are eliminating the "illusion" of big models in various ways: First, set up a "safety fence" to limit the model to deal with problems beyond its capabilities; Second, widely use external tools, with real-time content as the core, to supplement artificial intelligence's understanding of the current situation; Third, start from the underlying capabilities of the model, such as constantly verifying itself in the inference process, and cross-verify their own conclusions from different angles.

  "The first two methods are to suppress illusions, and the latter is to achieve the white-boxing of the reasoning process, which can be identified even if there are errors." Wang Guoxin told reporters that at present, these three methods can cooperate with each other to suppress the "illusions" problem to a certain extent.

  Taking the new upgraded version of the DS-Xiaobu Doctor 2.0 system launched by the National Children's Medical Center and the Affiliated Pediatric Hospital of Fudan University as an example, Zhang Xiaobo, deputy director of the Affiliated Pediatric Hospital of Fudan University, introduced that in terms of suppressing the "illusion" of the big model, relying on the self-built "pediatric enhancement search knowledge base", it can accurately match authoritative medical knowledge and improve diagnosis and treatment reasoning capabilities.

  How to further improve the safety and credibility of medical big models in complex clinical environments? Yang Tong, a researcher at the School of Information Science and Technology of Peking University, suggested that defense can be combined from inside and outside the model.

  "The security of the model itself is improved through adversarial training, using data preprocessing to weaken the effect of adversarial attacks, introducing fairness constraints to reduce algorithm deviations, etc., improve the security of the model itself. At the same time, an independent secure interconnection system is deployed between the model and the user, detect and intercept abnormal input data in real time, filter sensitive or malicious requests, and review and correct model outputs." Yang Tong said.

  DeepSeek helps high-quality data become the key support

  Reduce technical thresholds and optimize model deployment efficiency... As a domestic open source big model, DeepSeek provides an important breakthrough for the implementation of medical big models: hospitals can use the method of "large model base + small sample fine-tuning + professional knowledge integration" to directly fine-tune based on DeepSeek.

  "This is not a traditional sense of training from scratch, but a quick development of AI applications suitable for its own scenarios by requiring only small data and computing resources." Zhang Xiaobo introduced. But she also realized that if you want to train a high-quality medical model, you still need to comprehensively use large-scale high-quality data resources.

  In order to improve the ability to accurately identify and reason complex cases, the DS-Xiaobu Doctor 2.0 system relies on the hospital's big data management and control platform to integrate multi-modal data such as electronic medical records, laboratory examinations, medical imaging, genomic data, wearable device monitoring, etc. to achieve standardized storage, unified management and efficient call. Through data governance technologies, such as data cleaning, semantic analysis and intelligent annotation, we ensure the accuracy, timeliness and consistency of data, and combine expert review to accurately extract key clinical information.

  As a technology company that is also committed to applying AI technology to Grade 3 hospitals, Li Tao, chairman and CEO of Qilin Hesheng Network Technology Co., Ltd., said that DeepSeek is a general model. To make it have medical capabilities, it is necessary to use a lot of professional knowledge for incremental training, especially combining the hospital's own cases and knowledge base, such as rare disease case data in specific hospitals, special case data in specific regions, etc., to further optimize and fine-tune the model so that it can better adapt to the needs of specific scenarios.

  Industry insiders interviewed said that the most needed real medical data for medical big models are the application data of clinical experts and the training data of clinicians. This part of the data is often stored in different medical institutions in a multimodal form.

  "Advanced data distillation technology can greatly improve model performance." Yang Tong suggested that the data format should be unified, the data interoperability should be improved, and medical experts should be invited to participate in the data distillation process in depth.

  "The expert team must accurately record the patient's symptoms, signs, diagnostic process, treatment plan, and treatment effects, and label and interpret the key information, and optimize the model's diagnostic capabilities through human-machine collaboration." Li Tao said that through medical research projects, experts can be invited to participate and collect more valuable medical data.

  Improve the governance system and build a solid line of security and ethical defense

  In the process of applying large models to medical care, ethical risks have always received key attention from the industry. Dong Jiahong, an academician of the Chinese Academy of Engineering and dean of the School of Clinical Medicine of Tsinghua University (Beijing Tsinghua Chang Gung Hospital), analyzed that the big model is essentially a statistical model, and it is inevitable that factual errors, logical errors, etc. The "black box" characteristics of the model make it difficult to understand the medical decision-making logic and increase the difficulty of ethical review.

  Industry insiders interviewed said that the complexity of the ethical risk governance of medical big model originates from the need to consider both the two dimensions of medical ethics and scientific ethics. They have different perspectives and involve complex and extensive issues, and there are intersections between them, which urgently need "multi-dimensional" norms.

  First, in areas such as medical care that require transparency, big models need to improve interpretability, help users manage the decision-making process and build trust, and ensure the fair performance of the model among different groups of people by using diversified data sets and developing new algorithms, and eliminate potential bias and discrimination.

  Yang Tong said that the latest big model has powerful thinking chain capabilities, which can gradually refine and present the diagnostic reasoning process. With this capability, the model can be required to output detailed and structured inference steps that explain the logical basis behind each diagnosis or suggestion. At the same time, based on the refined thinking chain, the model can conduct multiple rounds of interactive communication with doctors and patients to further explain the diagnostic logic and decision-making reasons.

  Second, a closed loop of ethical governance throughout the entire life cycle of technology should be built, and a new paradigm of smart medical care should be created based on the principle of "technology for good".

  "In building the DS-Xiaobu Doctor 2.0 system, we deeply embedded ethical governance into the practical paradigm of the technological innovation chain, and built a knowledge and trust scale for clinical application of medical AI. Through sampling and survey of 332 medical workers, we systematically sorted out ethical governance issues. We built a two-way feedback channel for acceptance and satisfaction in actual operation, so that a positive cycle is formed between system iteration and patient trust." Zhang Xiaobo said.

  Third, most medical data involves patient privacy, and the future development trend of large models will inevitably strengthen data privacy protection measures.

  "Healthy personnel should be enhanced in their awareness of data compliance and privacy protection to ensure that ethical standards are followed in the AI ​​application process." Dong Jiahong suggested using data encryption, anonymity processing and differential privacy technologies to prevent unauthorized access and data leakage, and adopt emerging technologies such as blockchain and privacy computing to enhance the transparency and traceability of data governance. (Reporter Li Wenzhe and Gong Wen)

[Editor in charge: Ran Xiaoning]