How Large Models Are Redefining Health Management: Scene-Driven Innovation, Multimodal Interaction, and Validated Business Models

Entering 2025, a consensus has emerged within the industry: the key determinant of large language models’ competitiveness lies in their ability to truly integrate into daily clinical workflows and undergo continuous iteration. As one of the scenarios that has long been explored for AI applications, health management is also experiencing rapid development in the era of large language models.

So, how can large model technology inject new vitality into health management? And what new trends can it set off? VCBeat andXu Hongxia, Chief Scientist of WeDoctor AI Research Institute; Liu Jian, Senior Vice President of DeepWise; Zhang Zhiyun, Deputy General Manager of Guangzhou Nandafei Medical and Health Technology Co., Ltd.Three medical industry experts with deep expertise in large model applications for health management engaged in a dialogue for industry reference.

The main points of this article are as follows:

Against the backdrop of practical implementation and commercial value conversion becoming the focal point of industrial competition, accurately identifying and penetrating highly compatible scenarios has become a key proposition for the sustainable development of enterprises.

From the sharing by Liu Jian, Senior Vice President of DeepWise,DeepWiseDeepWise’s decision to enter the health management sector through the physical examination scenario is the result of synergies among market demand, technological accumulation, and policy guidance.。

On one hand, rising public health awareness and the national “early screening and early treatment” strategy have jointly driven sustained growth in demand for health checkups. On the other hand, as a leading enterprise in AI-based medical imaging, DeepWise has established substantial barriers to entry through its algorithm development and clinical validation in CT, ultrasound, and other medical imaging modalities. Health checkup scenarios inherently rely on imaging technologies across various examinations—including CT, MR, and DR—making them naturally aligned with DeepWise’s core technological strengths.

Scenarios are the litmus test for technology. Liu Jian shared that DeepWise’s large-model products have already demonstrated value validation in projects targeting lung cancer, colorectal cancer, and other conditions, successfully identifying over one million high-risk individuals and achieving structural optimization of medical insurance expenditures.

Guangzhou Nandafei Medical and Health Technology Co., Ltd. leverages weight management as a breakthrough point, utilizing an AI large model for "three-specialist co-management" to support the prevention and control of chronic diseases.

Zhang Zhiyun, Deputy General Manager of Nandafei, introduced that Nandafei has been deeply engaged in the fields of weight management and chronic disease management for over a decade, initially adopting a multidisciplinary team model to deliver weight management services. It was through this process that the company continuously recognized the limitations of traditional manual models—such as non-standardized services, quality fluctuations, and delayed responses—which constrain the scalable development of medical and health services.

The application of AI technology has readily resolved these challenges. Leveraging a combination of self-developed and open-source models, along with over a decade of data accumulation and technological advantages, Nandafei has launched the “Three-Physician Co-Management” AI large language model and the first AI agent for weight loss, metabolism, and chronic disease management, thereby enhancing the quality, efficiency, and accessibility of chronic disease care. Currently, Nandafei’s “Three-Physician Co-Management” AI large language model is widely deployed both within and outside hospitals as well as in health examination institutions, serving over 100,000 users.

At the VBEF Medical AI Large Model Innovative Application Forum held on May 9–10, multiple companies engaged in in-depth discussions regarding the logic behind scenario selection.

One panelist stated that, leveraging their team’s background and technical expertise, they chose to focus on research scenarios outside the standard clinical workflow. Another panelist adhered to the principles of “addressing pain points, meeting rigid demand, and ensuring high frequency,” prioritizing improvements in physician efficiency. A third panelist focused on consumer-facing health management, enhancing the value of medical services through 24/7 companion-style support. An investor pointed out that large language models should penetrate areas with scarce supply, such as primary care general practitioners, rather than merely optimizing resource allocation.

Despite their divergent paths, all companies converge on a core logic:The deployment of large models in the healthcare sector must be premised on a deep understanding of clinical scenarios, leveraging enterprises’ technological advantages and commercialization capabilities to build differentiated value barriers.This scenario-adaptability-based strategic choice is becoming the key leap for large model technology to transition from “usable” to “highly effective.”

Numerous enterprises have entered the fray, seeking to leverage this technology to deliver more precise and efficient health management services. However, significant challenges lie ahead on this path.

From a technical perspective, the issue of hallucinations in large language models remains a significant challenge. VCBeat has observed that some companies are addressing this problem through methods such as Retrieval-Augmented Generation (RAG), collaboration between generative and discriminative AI, and model improvements.

Xu Hongxia, Chief Scientist at the WeDoctor AI Research Institute, pointed out that,Hallucination is a practical challenge facing large model research at the current stage, but Micro Medical Group (Zhejiang) Co., Ltd. has adopted multiple strategies to address this. On one hand, during the input phase, it adjusts the hyperparameters of large language models and employs technical approaches such as Retrieval-Augmented Generation (RAG), Chain-of-Thought prompting, and fine-tuning with specialized medical knowledge to guide the model toward outputs that better meet healthcare industry requirements. On the other hand, during the output phase, it implements an audit model to rigorously review the large language model’s responses. Meanwhile, Micro Medical Group conducts retrospective analyses and organizes insights on responses that fail the audit, collaborating with experts to explore solutions, thereby continuously enhancing the reliability of the model.

Data, the “fuel” for large language models, is likewise an unavoidable challenge.Zhang Zhiyun of Nandafei candidly stated that chronic disease management initiated through weight control does not fall strictly within the realm of serious medical care, and faces challenges in lifestyle-related data collection due to scarce volume and suboptimal quality. Consequently, Nandafei has had to rely on its own painstaking accumulation, investing significant effort and cost into data collection and organization. After more than a decade of accumulation, Nandafei’s “Three-Professional Co-Management” AI large model has deeply integrated millions of records from obesity and chronic disease patient management, authoritative expert consensus, cutting-edge research findings, and self-constructed datasets.

Business collaboration scenarios are even more complex. On one hand, there is a supply-demand mismatch between application vendors and foundation models. For enterprises such as Micro Medical Group (Zhejiang) Co., Ltd. and Guangzhou Nandafei Medical and Health Technology Co., Ltd., which focus on application development, collaboration with foundational large models is essential to deliver healthcare services. During this process, there may be a mismatch between the iteration cycles of applications and the upgrade cycles of the foundational models. Therefore, companies will give greater consideration to their selection of foundational model partners.

As the technology provider, DeepWise is currently focusing on the seamless integration of medical applications with large language model (LLM) technologies. Citing the example of whole-course management of lung cancer, Liu Jian shared that the incorporation of LLMs has enabled a leap from addressing individual steps to providing comprehensive, end-to-end coverage. “There are differences in AI application scenarios between health management and in-hospital disease diagnosis, so solutions cannot be directly transposed. We need to gain a deep understanding of scenario adaptability in health management to deliver smarter health management services to the public.”

Additionally, there is the collaboration between application vendors and hospitals.Xu Hongxia from WeDoctor noted that, due to limitations in computing power and deployment costs, hospitals are more inclined to adopt lightweight and intelligent models. To address this, WeDoctor and Zhejiang University have been exploring model distillation techniques, achieving notable progress, publishing multiple academic papers, and accumulating substantial technical expertise. Currently, WeDoctor has also introduced technologies similar to Mixture of Experts (MoE), effectively controlling costs by segmenting medical scenarios, thereby facilitating smoother collaboration with hospitals.

Finally, there is the gap between the corporate mission of empowering patients and its practical application.Current practices of large medical models in terms of user experience indicate that the original aspiration to “empower patients” still faces multiple challenges, including technological implementation, service design, and user trust. Enterprises need to identify more concrete pathways to break through these barriers by striking a balance between technical precision and humanistic care, between service boundaries and reverence for medical practice, and between scenario adaptation and user needs.

Micro Medical Group adopts the “AI health management tools + human health managers” model. The AI system collects data from wearable health devices, continuously monitors patients’ various indicators, and instantly issues health alerts along with reminders for exercise and dietary plans, thereby assisting health managers in providing medication guidance or hospitalization prompts to patients. This model of Micro Medical Group has already achieved significant results in the Tianjin Digital Health Community.

Guangzhou Nandafei places great emphasis on humanistic care for its users. For instance, in the Guangdong-Hong Kong-Macao Greater Bay Area, its products are specially equipped with a Cantonese-speaking AI health management assistant that provides proactive services with voice-responsive interactions, thereby better serving the elderly population.

DeepWise has iterated its technical service philosophy to address the differences between in-hospital medical diagnostic services and out-of-hospital personal health checkup scenarios. For instance, in health checkup services, users—unlike hospital patients who receive single diagnostic results focused solely on disease detection—place greater emphasis on comprehensive, multi-dimensional services based on health reports for checkup clients. In response, DeepWise has developed a technical solution tailored to health examinations, enhancing both service professionalism and efficiency. This also illustrates that as medical AI expands from serious clinical care into consumer health, it requires an adjustment in service philosophy and technological boundaries, rather than a simple transfer of existing technical capabilities.

In health management scenarios, intelligent wearable devices enable enterprises to monitor users’ conditions outside of hospital settings. Interviews reveal that, compared with traditional AI-powered wearables, large language model (LLM)-based wearable devices deliver significantly better performance in data processing dimensions, depth of personalized services, breadth of scenario coverage, and technological integration capabilities.

Zhang Zhiyun from Nandafei shared that traditional wearable devices are constrained by computing power and algorithms, typically capable of collecting only a single physiological metric, with data exhibiting an "isolated island" characteristic.Wearable devices powered by large language models can, on the one hand, integrate with a broader range of devices to enable more diverse data processing dimensions; on the other hand, they offer a higher degree of personalization by establishing user memory, thereby providing accurate data support for subsequent health management.

“Micro Medical Group also aims to integrate professional medical knowledge with large models’ data perception of patients, transforming AI into ‘embodied cognitive intelligence,’ akin to embodied intelligence.”.” Xu Hongxia of WeDoctor believes that embodied intelligence is an interdisciplinary field combining artificial intelligence and robotics, while “cognitive intelligence” is the product of integrating artificial intelligence with healthcare. The “cognitive” component consists of two parts: professional medical knowledge on one hand, and large models’ data perception of patients on the other. She introduced that through in-depth cooperation with multiple well-known domestic smart wearable hardware manufacturers, WeDoctor Holdings has successfully broken down data barriers between hospitals and external settings, achieving integrated scenario-based data and closing the “last mile” in the full lifecycle of prevention, screening, diagnosis, treatment, management, and rehabilitation.

The core highlight of this innovative model lies in achieving seamless integration of data between in-hospital and out-of-hospital settings.. Leveraging health data collected in real time by wearable devices, the AI system of Micro Medical Group (Zhejiang) Co., Ltd. can generate personalized health management plans and continuously optimize and adjust them based on dynamic monitoring data, thereby achieving precise health interventions for patients.

In the wave of commercialization of large medical models, Micro Medical Group (Zhejiang) Co., Ltd., DeepWise, and Guangzhou Nandafei Medical and Health Technology Co., Ltd. have each leveraged their respective strengths, relying on their technological advantages and strategic positioning to explore distinct development paths.

WeDoctor Holdings Has Made Early Inroads in the AI Healthcare Sector with Impressive Achievements, Making Its AI Health Consortium Model a Market FocusAccording to Frost & Sullivan, WeDoctor Holdings has become China’s largest provider of AI-driven healthcare solutions by revenue in 2023. Furthermore, within the AI-powered Tianjin Health Community, digital enablement has delivered improved health outcomes while keeping medical insurance funds under control. The surplus rate of per-capita medical insurance quotas for diabetic patients exceeded 25%, and the growth rate of medical insurance expenditures declined significantly, achieving a “four-way win” for primary healthcare institutions, physicians, patients, and health and medical insurance authorities.

DeepWise, building on its B2B foundation, continuously expands service scenarios and advances into the B2C model, leveraging its mature AI technology expertise to construct a multi-dimensional protection system.: First, enhance efficiency and reduce patient waiting times through an intelligent auxiliary diagnostic system. Second, leverage AI quantitative analysis technology to generate high-precision diagnostic indicators in areas such as coronary heart disease risk prediction and bone mineral density testing, enabling physicians to focus on critical aspects of care. Finally, utilize large language models to translate complex imaging reports into layman-friendly interpretations, addressing patients’ comprehension challenges and establishing a comprehensive, multi-dimensional disease prevention and control system that spans from institutional services to patient engagement. In this process, AI shifts the diagnostic threshold earlier, providing all-around technical support for disease prevention, diagnosis, treatment, education, and research.

Guangzhou Nandafei Medical and Health Technology Co., Ltd. pointed out that, against the backdrop of a mature foundational layer for large models, the application end is poised for an explosive growth period. To this end,Nandafei will leverage resources from Grade 3A hospitals to deliver standardized services through its AI agent business middle platform, gradually extending its scope from health management to comprehensive patient journey services. Meanwhile, Nandafei is exploring diverse pricing models to cater to the varying needs of different clients.。

At the aforementioned VBEF Medical AI Large Model Innovative Application Forum, panelists also discussed the commercialization of medical large models. Their viewpoints can be summarized into four categories:

First, it advocates building a business model centered on replacing physicians, arguing that the physician-assistance model struggles to generate profits from health insurance reimbursements or physician income. Second, it emphasizes profitability through direct-to-consumer (C-end) patient payments, leveraging large language model (LLM)-based products to enhance patient experience and encourage payment based on perceived management value. Third, it adopts a business-to-business (B2B) model through private deployment, Software-as-a-Service (SaaS) subscriptions, and embedding models into enterprise workflows. This model has already been validated as successful in the healthcare sector.

Fourth, some argue that the business models of large language models can be divided into two categories:The first is the basic platform model, which charges based on technical API call volume; the second is similar to AI+CRO, providing services based on large language model technology, with a focus on service quality and delivery efficiency.

Future Development Trends of Large Language ModelsMicro Medical Group has explicitly designated “embodied intelligence” and its interactive capabilities as core research directions, striving to endow large language models with multi-dimensional perception and interaction abilities, thereby driving the evolution of AI from single-modality text-based interaction to multi-modal interaction encompassing vision, hearing, and beyond.In response to the rapid development of specialized vertical models such as scGPT and RuiPath, WeDoctor will explore multi-agent systems to accelerate the construction of a multimodal large medical model framework. Meanwhile, it will investigate the synergistic application of large models with diverse tools to further enhance their practicality and efficiency in healthcare scenarios. Additionally, addressing the computational cost challenges of large models, WeDoctor is actively developing lightweight technologies to reduce resource consumption while enhancing model intelligence, thereby promoting the development of large medical models toward greater efficiency and multidimensionality.