Collective下沉 Files IPO Prospectus Highlighting Grassroots Healthcare AI Solutions

Since 2025, DeepSeek has accelerated the deep integration of algorithm development with clinical scenarios through its open ecosystem. Large medical models have moved away from a “technology-first” mindset and gradually entered a phase of pragmatism. In this context, primary healthcare, which has an urgent need to improve diagnostic and treatment efficiency and quality, has become a key service target for large models.

How well do large language models align with primary healthcare? What is the real-world feedback from patients, physicians, and hospital administrators in primary care settings? VCBeat andGairui Technology, Huamei Haolian, Weimei Health, Winning Health(Alphabetical Order) A Discussion with Four Innovative Companies That Have Deployed Large Models in Primary Healthcare, for Industry Reference.

The main points of this article are as follows:

In the face of challenges such as uneven distribution of medical resources, a shortage of primary care physicians, and the severe burden of chronic disease prevention and control, artificial intelligence technologies represented by large language models are leveraging their capabilities to provide innovative solutions for “improving quality and efficiency” in China’s primary healthcare services.

Based on feedback from multiple interviewees, the application scenarios for large language models in primary healthcare include AI-powered intelligent triage and assisted diagnosis, automated medical record generation and quality control, personalized patient management, and public health services. Furthermore, the integration of AI health monitoring with smart wearable devices plays a significant role in township health centers and in the medical and health management of elderly individuals lacking adequate care.

Among them,AI-assisted diagnosis is considered one of the scenarios with the fastest implementation speed.Given the practical challenges faced by primary care institutions, such as weak diagnostic and treatment capabilities and scarce technical resources, large language model (LLM) products can rapidly analyze patients’ chief complaints and generate preliminary diagnostic recommendations, making them particularly suitable for diagnosing common and chronic diseases frequently encountered at the primary care level.

Wang Jun, co-founder of Weimei Health, added,The promotion of automated medical record generation and quality control has also yielded favorable outcomes in grassroots county-level regions., “Primary care physicians are often in a training phase, resulting in suboptimal quality of case generation and documentation. In the context of medical insurance cost containment under DRG (Diagnosis-Related Groups) and DIP (Big Data Diagnosis-Intervention Packet) payment systems, AI-assisted automated medical record generation and quality control can help physicians mitigate the risk of insurance reimbursement deductions.”

“Large Language Model Technologies Have Also Yielded Significant Feedback in the Personalized Management of Patients.” Hao Zhonghua, General Manager of the Regional Health R&D Center at Winning Health, stated that analyzing residents' health status to provide customized chronic disease management plans or health education content, which are then sent to residents after physician confirmation, can effectively reduce physicians' workload and enhance residents' sense of health gain.

Wei Qun, Vice President of Gerui Technology and an expert on the AI large model project for primary care, stated that large models for grassroots healthcare can enhance quality and efficiency from three perspectives: physicians, patients, and hospital administration. They enable stakeholders to perform tasks that should be done but are met with low willingness (enhancing efficiency), as well as tasks that were previously unfeasible but are now achievable through technology (improving quality).Significant progress has been made in the implementation of functions such as patient pre-consultation, automated medical record documentation and quality control, rational drug use, post-diagnosis follow-up management, and chronic disease management protocols.However, “if large models can provide more comprehensive and personalized recommendations and plans in treatment protocols and chronic disease management, users will favor them more.”

Wu Lei, AI R&D Director at Huamei Haolian, sharedThe Value of Large Language Models in Cost Control for Primary Healthcare. The in-depth application of large language models in primary healthcare not only reconstructs diagnostic and treatment workflows but also reduces medical costs across multiple dimensions. For instance, improved precision in primary care diminishes the need for referrals and triage, directly alleviating patients’ financial burden. On the provider side, AI assistance helps general practitioners bridge knowledge gaps between general and specialized medicine, lowering misdiagnosis rates and thereby driving a sharp decline in ineffective medical expenditures, such as redundant tests and duplicate prescriptions.

Interviewees also added that public health services constitute a major component of primary care physicians’ responsibilities. AI technologies, through automated follow-ups and medical record quality control, liberate primary care physicians from repetitive public health tasks, and can alsoSignificantly Enhance the Coverage Quality and Implementation Efficiency of Primary Public Health Services, the feedback on improving quality and efficiency for large models will be more pronounced.

Fundamentally, “enhancing quality and efficiency” represents the core value proposition of large medical models. Therefore, evaluating the merit of a large model in specific scenarios ultimately hinges on its practical application capabilities. Wu Lei shared the following dimensions to assess the empowering impact of large models on primary healthcare.

In the short term, medical care efficiency and quality can be observed.Evaluate the efficiency of large language model (LLM) enablement through metrics such as referral rates, reduced misdiagnosis rates, and triage accuracy; monitor changes in coverage data for the management of common and chronic diseases in primary care, along with parameters such as public health follow-up completion rates and health record completion rates, to assess the value of LLMs in enhancing regional healthcare quality.

By extending the observation period, regional population health outcomes can be assessed.such as the percentage reduction in regional diagnosis and treatment costs and healthcare system operational and maintenance costs, changes in medical insurance reimbursement rates and out-of-pocket payment ratios, whether chronic disease control rates have improved, and the decline in the incidence of regional endemic diseases.

Since the beginning of 2025, DeepSeek has surged in popularity and broken into the mainstream, with enterprises and hospitals alike announcing its deployment in anticipation of accelerating their intelligent transformation through this technology. However, reality is far more complex than imagined, and the actual implementation process inevitably faces the dilemma of “high ideals but harsh realities.” In primary healthcare settings, the core contradictions are concentrated in three key dimensions: technical adaptability, business integration, and resource support.

First, the infrastructure development of large language models is constrained by multiple factors, including computing power, algorithms, and data.On the computing power front, primary healthcare institutions face limited resources and funding, generally lacking the hardware infrastructure to build their own supercomputing centers or high-performance local server clusters. Their heavy reliance on cloud-based deployment leads to issues such as inadequate real-time responsiveness, latency, and system stability, which directly impair the continuity of the diagnostic and treatment experience.

From the perspectives of data and algorithms, primary healthcare data is characterized by uneven quality, fragmented distribution, and low standardization, resulting in high data cleaning costs. Consequently, some large language models (LLMs) are trained using data from tertiary hospitals; however, this approach leads to data mismatch with the epidemiological profile of primary care, which is dominated by high-prevalence chronic and common diseases. To some extent, this causes LLMs to perform poorly in primary care settings, exhibiting limited generalization capabilities and imprecise clinical decision support.

Secondly, issues such as hallucinations and poor interpretability in large models impose higher requirements on foundational healthcare LLMs operating within the serious medical domain.The Transformer architecture, a large-scale neural network framework widely adopted by large language models, can automatically capture and learn various features from data. It possesses powerful learning capabilities and can identify complex feature patterns. Compared with large hospitals, primary care institutions have relatively weaker diagnostic and treatment capabilities. For primary care physicians, AI serves as an enabling and supplementary tool that requires higher accuracy and reliability. This necessitates that large medical models deployed in primary care settings minimize the impact of hallucinations.

Next is the mismatch between the supply of technology and business demand.The current industry exhibits, to some extent, a misalignment characterized by “technology leading, demand lagging”—practitioners tend to focus more on “what large language models can do” rather than “what is actually needed at the grassroots level.” In contrast to large hospitals in first- and second-tier cities, clinical needs in primary healthcare primarily center on common and frequently occurring diseases as well as resident health, with a commitment to addressing inclusive healthcare challenges. In primary care settings, large language models require deep integration with clinical workflows and health information systems; however, exploration in this area remains insufficient.

Finally, there is the issue of resource support.The high technical threshold for large language model (LLM) application technologies makes implementation extremely challenging at the primary care level if AI platform capabilities, as well as model development and fine-tuning tools, are lacking. Furthermore, the scarcity of professionals with expertise in data and algorithms, coupled with insufficient engineering experience at the grassroots level, further exacerbates the difficulties in deployment. In addition, current AI technologies face numerous deficiencies in terms of laws and regulations; issues such as standards for LLM applications, data management, and the avoidance and allocation of medical liability require further improvement and refinement of relevant legal and regulatory frameworks.

Meanwhile, physicians should regard large language models as equal partners rather than blindly relying on their outputs. Therefore, promoting the healthy development of AI in primary care requires not only continuous technological advancements but also a process of market education to alleviate public concerns and enhance understanding.

As artificial intelligence technology permeates primary healthcare, the deployment strategy of large language models has become a key variable determining the realization of their technical efficacy. In response to the widespread challenges faced by primary healthcare institutions—including limited computational power, fragmented data, and insufficient operational and maintenance capabilities—Weimei Health, Gerui Technology, and Winning Health have each developed scenario-adaptive solutions.It should be noted that the deployment method needs to be combined with actual scenarios and user needs, which does not mean that the enterprise only supports this type of deployment method.

Weimei Health believes thatAll-in-one device deployment is highly compatible with primary healthcare.Ready-to-use and capable of delivering varying levels of computing power tailored to different application scenarios. Currently, Weimei Health provides ready-to-deploy hardware solutions to primary care institutions by integrating domestically produced computing resources with established large language models (such as deploying 70-billion-parameter models). This approach reduces reliance on high-performance servers and primarily supports core primary healthcare needs, including assisted diagnosis and medical record quality control.

The second category isCollaborate with the lead hospital of a medical consortium or the Health Commission, by unifying the local deployment of large language models (LLMs), primary healthcare institutions can adopt a tenant-based model for out-of-the-box usability. Gairui Technology highlighted three major advantages of this approach: First, it provides Retrieval-Augmented Generation (RAG) knowledge bases and data annotation tools to support continuous model evolution. Second, it establishes a dual quality control system that not only audits the quality of medical record data but also verifies the reliability of LLM outputs through comprehensive audit trails, thereby assisting in medication decision-making. Third, the model is optimized for common diseases encountered at the primary care level and can be deeply integrated into Hospital Information Systems (HIS) to enhance diagnostic and treatment efficiency.

The third category isCoordinated and jointly built by the government.Winning Health stated that the technical costs, computing power costs, and post-deployment maintenance costs associated with large language model (LLM) deployment are relatively high. It recommended leveraging unified government resource planning, with government departments taking the lead to enable multiple commissions and bureaus to share AI capabilities under the unified regulation and planning of the Big Data Bureau. For instance, in regions such as Ningxia and Qingdao, Shandong Province, the local Big Data Bureaus have taken the lead in centrally planning AI computing resources, while Health Commissions primarily assume the role of users and consumers, thereby facilitating stronger data security guarantees. Winning Health also noted that localized deployment is not always necessary for certain scenarios; for applications with lower data sensitivity, such as health-related Q&A services, public cloud resources can be directly utilized after proper data de-identification.

Finally, let us discuss the payment models for large language models. Currently, there are three primary payment methods for large model products: government fiscal funding, hospital self-funding, and collaborative ecosystem building by enterprises.

The government is the primary payer for large language model products in primary healthcare.The “Reference Guidelines for AI Application Scenarios in the Health Sector” issued by the National Health Commission, along with the “AI+” Action Plans successively launched in regions such as Beijing and Henan, have provided directional support and policy incentives for the introduction of large language models into primary healthcare. Enterprises also offer diverse payment options—including one-time buyouts, project-based purchases, and annual subscriptions—to meet the needs of different scenarios.

If customers are unwilling to pay, it indicates that the product’s value has not yet been fully demonstrated. In the long run, if large language model (LLM) products can be deeply integrated into medical workflows and address core pain points, their role in improving quality, enhancing efficiency, and reducing costs will become more pronounced, making customers more willing to sustain their payments.In addition to the government, medical consortiums and primary healthcare institutions are also shifting procurement costs, often through mechanisms such as health insurance payments.

Currently, several top-tier tertiary hospitals are proactively procuring large medical AI model products. A search of public records reveals that in March alone, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Changzhou No. 1 People’s Hospital, Baoji Central Hospital, Shaoxing People’s Hospital, and Nanpi County People’s Hospital all published procurement intentions for large medical AI models, with budget amounts reaching up to RMB 4.8 million.

There are also some new innovative payment models emerging. Such asHuamei Haolian explores new pathways through a B2B2C model, reaching primary healthcare institutions via the corporate sector to establish a commercialized channel for health management services.

In fact, AI applications in the healthcare sector primarily serve an auxiliary role rather than completely replacing human physicians. While AI can efficiently process vast amounts of data and provide precise analyses, core clinical judgments in medical decision-making still rely on the experience and wisdom of human doctors. This characteristic mirrors the historical trajectory of numerous medical technological innovations, which were often initially met with misunderstanding and skepticism but ultimately achieved widespread adoption and benefited humanity as the technologies matured and became more prevalent.