2025.02.20
本文字数：2274，阅读时长大约4分钟

Introduction: More than a few major healthcare companies have announced integrations, yet enterprises still need to proactively consider various challenges including safety, ethics, registration filings, and health insurance access.

Following the official announcements by the first batch of companies, including Yidu Tech, Eagle Eye Technology, Wanda Information, and Zhiyun Health, the latest data shows that over a hundred major health companies have announced integrations, covering areas such as drug development, image analysis, diagnostic screening, pathological testing, and chronic disease management. Due to the characteristics of large models being open-source and cost-effective, the integration of artificial intelligence (AI) with the healthcare and wellness industry will give rise to more scenarios and products.

"But we must also face the market recognition behind the scenes and products, as well as challenges in payment, access, security, ethics, and other aspects," a senior industry insider told First Financial.

Two mature application scenarios
记者了解到，数据的数量规模、质量等级决定了AI背后的模型训练状态，这也是AI大模型率先在医疗、金融等领域爆发的原因。以医疗领域为例，“智慧医院”是一大主要应用场景。比如，深圳市人民医院日前宣布已本地化部署DeepSeek。该院信息技术部主任丁万夫称，现阶段，AI在智慧医疗方面的应用场景主要是辅助诊断，医院与腾讯合作开发的AI预问诊服务已得到应用，患者在挂号缴费后可以收到预问诊推送，医生则结合患者的回复信息自动化一键生成电子病历。

Similarly, the Fourth People's Hospital of Shanghai recently announced that it has completed local deployment. The next step is to rapidly provide doctors with precise decision-making support based on over ten thousand typical cases and years of accumulated treatment plans. Additionally, Ruijin Hospital in Shanghai has partnered with Huawei to launch the "Ruizhi Pathology Model," which empowers clinical diagnosis with millisecond-level, seamless image reading capabilities.

As another major application scenario, "diagnostic equipment" is more mature compared to smart hospitals. For instance, Neusoft Medical's stroke assessment software can complete ischemic penumbra analysis within seconds with an accuracy rate of over %, and it has been included in the Chinese expert consensus. Another example is United Imaging Healthcare, which integrated algorithms into equipment such as CT (Computed Tomography) and PET-CT (Positron Emission Tomography-Computed Tomography) many years ago, enabling imaging diagnostics with smaller doses and lower radiation for patients.

The question is whether the aforementioned application scenarios can ultimately enhance efficiency and reduce costs. In this regard, Zhang Yuming, Director of the Medical Big Data Research Center (East China) at the China Academy of Information and Communications Technology, told reporters that this needs to be viewed categorically. For scenarios such as medical record writing, assisted diagnosis, and remote consultations, which mainly involve natural language processing, multi-modal and multi-dimensional data integration, and normalization processing, large models can make the reasoning process relatively transparent, thereby improving the efficiency of doctors' identification and confirmation. As for scenarios like pathological analysis and image navigation, in addition to improving efficiency, these scenarios can also compensate for issues such as lack of concentration due to changes in human energy and emotions. Compared to the previous type of scenarios, this type of scenario places more emphasis on subtle differences and precision.

Zhou Yuefeng, Vice President of Huawei and President of the Data Storage Product Line, also stated that at this stage, integration with various industries still faces numerous obstacles. These include: First, data engineering is time-consuming, as data needs to go through a series of processes such as collection, cleaning, enhancement, and evaluation, which accounts for a significant percentage of the time spent on model development and training. Second, the difficulty in model training and application implementation is evident in the large number of paradigms for large models, making development and debugging challenging. Third, factors such as waiting for computing power, task tidal waves, and resource fragmentation can all lead to lower cluster availability.

risks and challenges
随着AI在医疗领域加速落地，产业背后正在面临哪些风险与挑战？厦门大学医学人工智能研究院负责人王连生则表示，除了业界一直关注的数据安全和隐私保护等风险，AI大模型也会存在包括推理过程、责任归属、公平性等在内的诸多挑战。推理过程的挑战涉及模型是否可信、数据输入限制；责任归属的挑战涉及相关部门需制定AI大模型监管规则，以及处理数据时需要遵循的法律法规；公平性的挑战则涉及大模型可能出现的“幻觉”“偏见”等问题。

In addition to the risks and challenges inherent in the development of a company's own business, medical devices (including both software and hardware) face numerous difficulties during the pre-market registration and medical insurance access stages. Zhang Yuming stated that before being marketed, medical devices must undergo rigorous safety and efficacy research, testing, and verification. Only after approval can they be used as equipment, software, and system products for medical diagnosis and treatment. Post-market, continuous attention must be paid to the occurrence and analysis of adverse events.

In this regard, Zhang Yuming stated that currently, on one hand, there is an urgent need to advance the formulation of relevant standards. This includes developing industry standards for medical devices, standardizing data interfaces, algorithm evaluations, safety certifications, etc., to promote interoperability and sustainable healthy development. On the other hand, it is necessary to clarify the interpretability and reliability of the algorithms behind medical devices, which can make the decision-making process understandable to both doctors and patients.

Regarding the registration strategy for medical devices, Wang Jing, the founder of Silicon Intelligence, suggests the following: First, companies should clearly define the classification of medical devices. If the related products are used for disease diagnosis, treatment, prevention, or have a direct impact on patient health, they should be classified as medical devices. In this case, it is necessary to submit complete technical documents and clinical data in accordance with relevant regulations. Second, companies should select reasonable training and validation sets. Specifically, the training set should cover diverse clinical cases and scenarios to ensure the representativeness and diversity of the data; the validation set is used to assess the model's performance on unseen data, and should ensure that its sample size and distribution are reasonable. The algorithm behind the training set is particularly important, as the rigor of its logic will directly affect the reliability of the medical device's effectiveness. Third, for the "black box" issue of the model, it is essential to demonstrate its safety and effectiveness through adequate clinical validation and performance testing. Even if the internal workings of the algorithm are not interpretable, its reliability can be proven through long-term use of real data and validation across multiple scenarios.

In terms of medical insurance access, the National Healthcare Security Administration stated in [year and month] that it would establish an "AI-assisted" extension in radiology, ultrasound, and rehabilitation projects; at the same price level, hospitals can choose to train medical staff for diagnosis and treatment, or opt to use artificial intelligence to participate in medical practices, but no additional charges will be applied at this stage.

Currently, hospitals include the costs of image viewing and interpretation in the overall charges, and it is rare to see these costs itemized separately. These have always been practiced as auxiliary functions," a provincial medical insurance department official told reporters. In addition to auxiliary diagnostics, intracranial deep brain stimulation pacemakers, which have been in use for many years, also feature natural language decoding capabilities. "For such medical devices, if they apply for related medical insurance charges, it might be appropriate to consider them," the official added.