Recently, Sun Yat-sen University Cancer Center (SYSUCC) and YIDU TECH jointly organized a research event for investment institutions. The event featured no flashy PowerPoint presentations or exaggerated demonstrations claiming to "outperform human doctors." Instead, the information technology team directly accessed real outpatient workstations within the hospital to showcase core AI assistant functions, including automated medical record generation, intelligent tumor staging assessment, and standardized treatment plan recommendations. This candid, "bare-faced" approach is relatively uncommon in China's medical AI sector, especially during investor research sessions.
Coinciding with this institutional research window, YIDU TECH (02158) officially announced its successful bid for a major SYSUCC project, with the smart clinical and settlement service upgrade construction project valued at 9.08 million yuan. The underlying collaboration is highly indicative of industry trends: YIDU TECH and SYSUCC have been engaged in deep, collaborative development for a full decade since 2015.
This decade of dedicated implementation has brought a fundamental industry logic to the forefront: top-tier hospitals do not need AI healthcare solutions with the most powerful parameters or the most sensational hype. What they truly require is practical AI that can integrate into clinical workflows, meet essential research needs, ensure medical safety, and be genuinely usable, reliable, and capable of long-term iteration.
**Three Major Pitfalls for Large Models "Entering the Hospital": Why Do Many AI Products Fail at the Department Door?**
Since 2025, with the rapid advancement of large model capabilities from players like DeepSeek and Tongyi Qianwen, hundreds of top-tier tertiary hospitals across the country have announced "integration of large models." However, behind the buzz of announcements, few products have been genuinely implemented at the clinical frontline for routine use by doctors, with actual penetration rates far below industry claims.
The SYSUCC research event reflected a common reality: even with well-developed AI tools, significant barriers remain to integrating them into doctors' daily work. Clinical doctors possess high professional barriers and entrenched diagnostic habits, often holding a cautious or even conservative attitude towards new AI tools, with a naturally weak willingness to adopt and learn them.
More challenging than user acceptance, the entire medical AI industry is currently grappling with three major structural pitfalls during hospital implementation:
**Pitfall 1: Emphasizing Models Over Data, Leaving AI as "Water Without a Source"**
Data from specialized cancer hospitals is inherently complex: multimodal, massive in volume, with long dynamic follow-up cycles, fragmented heterogeneous systems, a high proportion of unstructured text, and highly inconsistent descriptions for the same condition among doctors (e.g., over a thousand expressions for lung cancer diagnosis alone).
SYSUCC's solution was to invest several years in data governance first, establishing single-layer data standards, specialized cancer datasets, and disease-specific field sets. They structured and standardized the full-course data of over 2 million patients, ultimately forming more than 8,000 structured fields.
YIDU TECH's role in this process was not merely that of a "model provider" but a deep collaborative partner from data governance to AI middleware platform construction. This is precisely the capability most general large model companies lack—they do not understand the real, "semi-structured, unstructured, and variably described by doctors" data within hospitals, nor do they possess the ability to perform specialized data organization and value accumulation.
**Pitfall 2: One-Size-Fits-All Approach, Ignoring Individualized Doctor Needs**
SYSUCC's IT department revealed that when the hospital's self-built AI diagnostic assistant was first launched, it sparked an AI craze across the hospital, with four to five hundred doctors signing up for training. However, the IT department soon received 50 different application requests—each specialty and each doctor had different requirements for medical record generation formats, detail levels, and focus areas. A one-size-fits-all universal model simply could not be effectively implemented.
Consequently, SYSUCC and YIDU TECH jointly launched a "personalized" intelligent agent platform: doctors need only three steps—naming, writing prompts, and selecting data scopes—to quickly create their own dedicated AI assistants. To date, approximately 170 doctor-built intelligent agents have been deployed hospital-wide, covering scenarios such as medical record writing, medical order management, and patient education. Some agents are even named after the doctors themselves, truly achieving AI adaptation to individuals and integration into daily clinical routines.
**Pitfall 3: Lack of Traceability, Making Doctors Hesitant to Use AI**
"How can we trust AI when it has hallucinations?" This is the ultimate challenge facing all medical AI. YIDU TECH's solution centers on full-chain traceability. In the live demonstration at SYSUCC, all AI-generated medical record content could be traced with one click, directly linking to corresponding original medical records such as discharge summaries and test reports, ensuring clear sources and verifiable evidence.
In a practical VTE thrombosis risk assessment test, among 59 samples where AI and manual scoring differed, cross-review by senior experts determined that AI assessment was more accurate in 50 cases, achieving a comprehensive accuracy rate of 90%.
YIDU TECH's technical approach does not rely solely on a single large model. Instead, it employs a hybrid intelligent architecture combining large and small models, constrained by medical knowledge graphs, and featuring full-process traceability. It avoids relying on model parameter hype, instead using rigorous technical risk control to strictly manage hallucination risks and anchor medical safety, genuinely enabling doctors to trust and confidently incorporate it into daily diagnostic workflows.
**Industry Transformation: Structural Reshaping of Medical AI**
Currently, China's medical AI sector has formed three main player groups, each establishing clear positioning differences and capability boundaries within top-tier tertiary hospital systems:
* **Traditional Medical Informatization Vendors:** Their advantage lies in mature in-hospital channels and deep understanding of business processes. However, they are relatively weak in large model capabilities, deep data governance, and intelligent agent architecture, serving more for process digitalization upgrades and facing challenges in incremental markets like advanced AI clinical research empowerment.
* **General Large Model Teams:** They possess strong foundational model capabilities and ample computing power. However, the complexity of medical scenarios, lack of specialized data standards, and the multi-year clinical磨合 threshold mean their current implementations are more focused on non-core diagnostic areas like科普, triage, and basic document generation. Cases of真正进入临床决策闭环 remain limited.
* **Vertical Medical AI Vendors (represented by YIDU TECH):** With over a decade of深耕 in healthcare,长期绑定头部医院,沉淀专科数据标准, deep understanding of clinical implicit logic, research demands, compliance, and data localization requirements, they possess the capability for middleware-based continuous iteration. They can manage the complete lifecycle from data infrastructure and AI middleware platform construction to full-scenario intelligent agent deployment and continuous expansion/repurchase, becoming the preferred partner for research-oriented hospitals in advanced AI construction.
These three player groups are not in a "winner-takes-all" competition but rather engage in division of labor and collaboration. However, barriers to high-value core scenarios are rapidly rising.
Simultaneously, hospitals' procurement logic is undergoing a fundamental shift. In the past, hospital AI procurement was primarily project-based and point-specific: one imaging AI, one medical record generation AI, one CDSS, purchased in a fragmented, non-integrated manner. Now, leading tertiary hospitals are shifting towards infrastructure-level investment, treating AI and data middlewares as the new generation of intelligent infrastructure—a shared foundational platform for the entire hospital, empowering multiple scenarios on demand. Concurrently, annual budgets are becoming常态化, computing power investment is increasing year by year, and applications are expanding from points to全院铺开.
SYSUCC mentioned during the research that its proportion of AI-related investment is continuously increasing, which is highly representative. Hospitals are no longer solely pursuing short-term returns but are placing greater emphasis on data asset accumulation, research productivity enhancement, clinical manpower substitution efficiency, and the industrial吸附效应 of clinical trials. In the future, more hospitals will follow the path of "first build the foundation, then deploy intelligent agents, achieve全场景渗透." The industry is entering a new stage of platform-based co-construction and long-term iterative continuous service.
**Capital and Industry Insights: Medical AI Enters a "The Strong Get Stronger" Matthew Effect Era**
**Benchmark Hospital Models Will Be Rapidly Replicated Nationwide**
As a benchmark in oncology, SYSUCC's AI construction model holds strong industry reference value. The SYSUCC data and AI platform has integrated 56 internal business systems, accumulated full-course data for over 2 million patients, supported more than 5,000 clinical studies, facilitated the publication of over 500 high-impact papers, improved单病种智能上报 efficiency by 10 times, and cumulatively saved medical staff 80,000 hours. These quantifiable, implementable, and reviewable practical values will powerfully attract large tertiary hospitals and renowned specialized hospitals nationwide to借鉴与复刻. Companies that first successfully implement落地范式 in leading hospitals will顺势迎来规模化推广的行业红利.
**Institutional Investment Logic Shifts from "Speculating on Models" to "Evaluating Implementation Barriers"**
The core focus of leading research institutions has clearly changed: no longer just listening to模型讲概念, but instead重点考察—是否有顶级医院长期共建,是否有可量化的落地效果,是否有持续复购订单,是否有药企变现的第二曲线,是否形成了专科数据标准的壁垒. Targets based purely on concepts without深耕场景 will gradually be marginalized by capital.
**Industry Thresholds Significantly Raised, Increasing Difficulty for Latecomers to Catch Up**
The real门槛 in medical AI is not the algorithm itself, but rather years of clinical磨合, data standard沉淀, accumulation of doctor diagnostic logic, and experience in compliant implementation. These cannot be compensated for by short-term烧钱 or general large model capabilities. The industry is formally entering the Matthew Effect stage: leading benchmarks锁定先发优势, while latecomers are mostly left to compete in non-core scenarios.
Comments