Author | Zhou Lei Editor | Cen Feng
Dr. Xie Di from the Nephrology Department of Nanfang Hospital still remembers a 24-year-old female patient admitted a few years ago for systemic lupus erythematosus and respiratory distress. Her blood test results, however, pointed to a rapidly progressive rare disease—aHUS (atypical hemolytic uremic syndrome). During onset, microthrombi block small blood vessels throughout the body, with the kidneys bearing the brunt; the mortality rate during the acute phase reaches 25%. Dr. Xie has participated in too many such consultations. What makes her feel most powerless is that while there are effective treatments, integrating the information is extremely time-consuming. The effective intervention window for this disease is only 24 to 48 hours after symptoms appear. If the disease is quickly identified and treatment is initiated within 7 days of onset, the prognosis is far better than if treated after 7 days. Administering medication within the golden time window can effectively prevent irreversible damage to organ function.
However, the reality is that patients often have to shuttle between multiple departments, with the diagnosis cycle measured in days. By the time doctors from nephrology, hematology, infectious diseases, and other departments are gathered for a consultation and the data is integrated, the optimal intervention window may have already passed. The clinicians who understand the disease logic best have a complete set of screening pathways in their minds, yet they cannot build a tool to help them quickly integrate multidisciplinary information. They can only rely on manual coordination, communicating with each department one by one, and making decisions based on experience under immense time pressure. This dilemma is not an isolated case. At Nanfang Hospital, doctors encounter similar situations in their consulting rooms every day: those who know exactly where the problem lies cannot build the tools to solve it. Creating tools usually requires programming, scheduling, development costs, and getting engineers to understand complex clinical workflows.
Many doctors told Leifeng.com that many of their ideas are always "abandoned before they are even proposed due to this or that reason." It wasn't until April this year that Nanfang Hospital signed an agreement with Huawei, and the HAIP platform, built on Huawei's DCS AI solution, was officially launched. Its core component, the Nexent Agent Platform, allows doctors to generate agent prototypes in a zero-code, visualized manner without programming, simply by describing clinical logic in natural language—many clinical innovations that were completely impossible to put on the agenda in the past have suddenly become ideas that doctors can verify on the same day. When the threshold for creating tools drops from "knowing how to code" to "being able to clearly articulate requirements," those who understand clinical problems best are gaining an unprecedented ability: directly defining the logic of AI tools.
When "Requirements That Would Never Be Raised" Become Reality
As mentioned at the beginning of the article regarding the aHUS diagnosis process, doctors actually already have a complete set of aHUS screening pathways in their minds, but they also clearly know that turning this pathway into a runnable auxiliary tool was almost impossible in the past. Dr. Xie told us that she tried to understand the market rates for traditional IT development, but just to build a clinical decision support system for high-risk screening, the first step would require substantial outsourcing costs. The more fundamental obstacle lies in communication: those doing data collection don't understand AI, and those doing AI don't understand the business workflow. Doctors have complete clinical logic in their minds, yet engineers need a long time to understand those subtle diagnostic and treatment decision nodes. By the time both sides finally align on the requirements, the scheduling takes months, and the final product often differs greatly from what the doctors wanted. In obstetrics, Dr. Huo Zhifeng faces another form of silence. When some low-frequency but high-risk acute emergencies occur clinically, the processing time window is very short, which severely tests team collaboration capabilities and therefore requires a lot of advance simulation training.
However, in traditional simulation teaching, case design is time-consuming, and debriefing relies on memory, making it difficult to increase the training density for resident doctors. Doctors have thought about optimizing this, but he has no programming experience and doesn't know how much it would cost to go through the traditional IT development process. Before large models were widely used, this kind of simulation teaching relied almost entirely on "manual crafting." This situation can be described as an "old ailment" in the field of medical AI. Doctors possess clinical intuition based on experience, while engineers possess system thinking based on code. There is a lack of an effective translation layer between the two languages, and the communication gap has never been bridged.
The industry has a massive amount of "dead data" and "living ideas," but medical AI innovation has long been led by the technical side: engineers define problems and design products, while doctors, whose daily work is already extremely busy, can only passively adapt. From raising requirements to scheduling development, and then repeatedly communicating business logic, there is an insurmountable wall in between. Those needs that grow out of real consulting rooms, because they cannot be accurately expressed or quickly verified, are mostly abandoned before they are even voiced. Liu Hongyan's team at the Liver Disease Center has even experienced a specific frustration in realizing technical requirements. They tried an EMDT system, investing a lot of effort, but it ultimately failed to survive due to high technical barriers and heavy continuous operational costs. After that failure, this concept was shelved for a full ten years. Different departments and different clinical scenarios share the same locked-in state behind them: technology is developing rapidly, but it hasn't stopped to wait for the people who truly need it.
Doctors Begin to "Build It Themselves and Reap the Rewards"
As those who understand clinical workflows best, how do doctors transform the recurring dilemmas in their consulting rooms into runnable agent logic? To find out, we spoke with three clinicians involved in agent development.
aHUS Rare Disease Diagnosis: Connecting Clues Scattered Across Eight Departments
The aHUS mentioned at the beginning of the article actually has a relatively clear etiology: the patient's complement system, originally the immune guards used to fend off bacterial invasion, is overactivated when it shouldn't be, attacking vascular endothelial cells inwardly, leading to systemic microthrombosis.
As long as this rare disease can be diagnosed, targeted treatment can be administered. However, the early symptoms of this disease lack specificity. Patients might present with anemia and thrombocytopenia in the hematology department, or with elevated creatinine in the nephrology department. Each department only sees fragments and can only associate them with diseases related to their own specialty; the diagnostic and treatment records of each specialty are also scattered in different systems, making it difficult to aggregate them in time to catch the golden window for treatment. Xie Di also described a case that left a deep impression on us: a pregnant woman was admitted with aHUS. Although her life was saved after terminating the pregnancy, she was left with irreversible renal insufficiency—if the disease could have been quickly diagnosed and targeted drugs initiated in the early stages, it would not only avoid the high mortality rate in the acute phase but also protect the patient's long-term renal function, preventing subsequent progression to uremia or even the risk of losing the transplanted kidney after a kidney transplant.
The problem lies not in the medical judgment itself, but in the speed of information integration. However, the diagnosis of aHUS is a highly complex diagnosis of exclusion. Doctors first need to determine that the patient has thrombotic microangiopathy (TMA), and then rule out thrombotic thrombocytopenic purpura (TTP), malignant hypertension, drug-induced causes, and other possibilities one by one. This involves cross-judgments across eight or nine specialties, including hematology, nephrology, infectious diseases, and rheumatology/immunology.
In the past, this screening process relied entirely on the doctor's personal experience and manual coordination of multidisciplinary consultations. Now, Xie Di's team, based on the HAIP platform, is attempting to transform this complex clinical thinking into a runnable agent architecture: the data perception agent is responsible for extracting key indicators such as complete blood count and schistocytes from the HIS system; the stratified reasoning agent advances according to the pathway of "identifying TMA → confirming TMA → ruling out other diseases one by one"; and the MDT collaborative agent simulates the cross-examination of doctors from eight specialties. More crucial is the divergence handling mechanism. Xie Di told Leifeng.com that they specifically added an arbitration agent for divergence handling—rather than eliminating differences, precisely defining ambiguous problems is the focus.
When agents from different specialties provide suggestions in different directions, the system designs a four-level handling logic: Level 1 traces back evidence and checks for omissions to determine if the divergence stems from missing key data or inconsistent information; Level 2 structurally evaluates the basis of different opinions by combining evidence levels, timeliness, and the patient's actual condition; Level 3 presents the opinions of each specialty, the evidence basis, and the main divergences to the real doctor in a transparent manner for the doctor to make the final decision; Level 4 completely records the points of contention, decision basis, and final conclusions, precipitating them into a traceable case library to provide auxiliary reference for future similar cases.
Xie Di specifically mentioned two designs she considers to have moat value: one is human-machine dialogue. Unlike general large models, this system allows doctors to directly click to correct when they find the AI's judgment is wrong. After the doctor's correction, the content enters the knowledge iteration process, which is updated to the knowledge base after expert review, version management, and verification. The second is the memory flywheel. Having knowledge base constraints is not enough; AI can also make mistakes when decomposing the knowledge base. The memory flywheel allows the summary results of every offline MDT to continuously feed back into the system. As reviewed real-world cases continue to accumulate, the agent can enhance its ability to retrieve similar cases and provide auxiliary references, presenting historical similar cases, handling processes, and outcomes to doctors in new cases to better assist clinical judgment.
The above agent design seems complex, but Xie Di told us that she started from scratch, began using the HAIP platform to build after attending lectures and training, and relied on the zero-code visual capabilities of the Nexent Agent Platform: "You just upload the knowledge base directly, no programming is needed, it automatically helps you decompose and generate the knowledge base." During debugging, she first roughly stated what she wanted, the platform generated a rough framework, and then she gradually corrected it according to clinical needs—which indicators to focus on, which logic chain to change, correcting it step by step into the desired form.
From zero foundation to completing the preliminary prototype, Xie Di and her team took only ten days. In the past, outsourcing similar systems would be expensive, time-consuming, and no one would truly understand the clinical workflow; now, simply by describing the screening pathway in natural language, the "rule-out one by one" differential diagnosis thinking in her mind can be implemented as a runnable system. She revealed that the team's next step is to embed this system into the hospital's medical record system, capturing HIS data through southbound interfaces and feeding back opinions through northbound interfaces, so that information scattered in various departments can be integrated in real time, completing screening and early warning within the golden window.
A Doctor Who Doesn't Code Built an Obstetric Acute and Critical Care Training System
The rescue of aHUS races against the time window, and the same is true for obstetric critical care training. For example, postpartum hemorrhage progresses rapidly and tests team collaboration, but it is not commonly encountered in clinical practice, so resident doctors see it less often. Dr. Huo Zhifeng told us that in traditional teaching formats, a decent case requires the teacher to manually organize it for forty to fifty minutes. Once the training starts, the teacher has to observe, record, and judge simultaneously, spreading themselves too thin. Debriefing can only rely on after-class memory, making it difficult to give precise feedback on the spot about what the students did well and what they missed. More importantly, traditional training struggles to replicate the time pressure of real rescues.
Students might perform well in a relaxed environment, but in a real environment, pressure surges, and human judgment significantly declines. That sense of urgency cannot be replicated by conventional training. This need to simulate "real reactions under high pressure" can only be truly realized by those who have long stood on the clinical front line and personally taught. Based on the HAIP platform, Huo Zhifeng transformed his clinical experience in postpartum hemorrhage rescue into an AI-assisted case generation and intelligent debriefing system. The teacher only needs to upload desensitized case data and tell the system the teaching objectives and difficulty level. The Nexent Agent Platform will generate a structured medical record draft in the background, which is published after the teacher's review and confirmation, and the drill officially begins.
Leifeng.com noticed at the agent demonstration site that the entire usage process is designed to be as close to a real rescue scenario as possible. The student side only sees an initial scenario and a countdown; the system will not proactively provide additional information—there is no teacher prompting them at all times in a real rescue either. Everything relies on the students to evaluate and request on their own. For example, the system only feeds back blood pressure and heart rate when the student requests vital sign monitoring. The agent will cause the patient's condition to continuously evolve under time-compressed pressure: if the student handles it correctly, the system feedback shows the condition improving; if the student fails to intervene in time, the condition automatically deteriorates, with blood loss rapidly rising from hundreds of milliliters to over a thousand milliliters, accompanied by physical sign changes such as pale complexion and restlessness—this design forces students to maintain active thinking and decision-making under high pressure, rather than passively waiting for instructions.
The deterministic state machine engine in the system is responsible for controlling the evolution of the condition, but Huo Zhifeng explicitly stated that the rules are not generated by AI, but formulated by doctors based on over a decade of clinical and teaching experience—under what circumstances the condition should progress and to what extent has been repeatedly verified. The changes in the debriefing session are equally obvious: previously, the teacher would slowly compare and organize with a checklist, taking about 15 to 20 minutes; now the system can generate a debriefing draft in about 4 seconds, recording what the student did at what time and providing targeted improvement suggestions. Huo Zhifeng also made an upgraded version. With the addition of speech recognition, the agent can distinguish the voices of different participants, strengthening the training on team collaboration.
Huo Zhifeng told Leifeng.com that this platform-based development experience made him realize what "technological equity" truly means. He has no programming experience, yet he can directly turn teaching ideas into a runnable system. "I don't know how to code, but I don't need to now; as long as the AI knows how, I just need to raise requirements, and if I'm not satisfied, I'll ask it to change." Frontline clinical educators are experiencing the ability to directly create tools for the first time. More importantly, this system is generating outward technological flow. Huo Zhifeng believes that the value of such agents in primary hospitals might be much higher than in large hospitals.
Students in large hospitals generally have strong capabilities, encountering more patients and more complex situations; but for primary hospitals, they might not encounter a single case of severe postpartum hemorrhage for years, and once they do, it is doubly dangerous—training with agents can at least maintain a certain level of familiarity. He even received messages from ICU and emergency colleagues asking if the agent could be opened to them—from Huawei's technical foundation, to the clinical doctors' teaching innovation, to the capability extension of primary and brother departments, technological equity has formed a clear flow chain here. Next, this system will migrate to scenarios such as eclampsia, emergency, and ICU, and explore adding image recognition or even VR, extending the boundaries of simulation training from obstetrics to more high-pressure acute scenarios.
Full-Cycle Liver Cancer MDT: Freeing Experts from the Conference Room
Obstetrics emphasizes team collaboration training for young doctors, while for senior doctors with high seniority and extremely busy schedules, collaboration efficiency is a pain point on another level. There is a significant gap in the five-year survival rate of liver cancer between China and Japan. Liu Hongyan's team has been focusing on this issue for a long time. They believe the gap lies not in surgical techniques, but in whole-process management—Japan does well in the "three early" (early detection, early diagnosis, early treatment), while many patients in China are diagnosed at a late stage and lack continuous follow-up management. But after 15 years of liver cancer MDT, Liu Hongyan's team increasingly feels a kind of organizational fatigue: every consultation is like organizing a meeting. The attending doctor has to organize massive amounts of data and make PPTs, and experts need to be present at a fixed time and place, serving a very limited number of patients. What's more troublesome is that the end of the consultation is often the endpoint. Whether the patient followed the treatment plan, what the follow-up results were, and subsequent follow-up management are completely disconnected.
Dr. Liu Hongyan told Leifeng.com that ten years ago they wanted to do online consultations, and in 2016 they even tried an EMDT product, which ultimately failed to survive due to high technical barriers and operational costs. The cooperation established between Huawei and Nanfang Hospital in April this year made Liu Hongyan's team feel the "east wind" of technology. Based on the HAIP platform, the team built a demo in just over a month, realizing a concept that had been unfulfilled for 15 years. They transformed the full-cycle links of early warning, quality control, specialty avatars, and follow-up monitoring into a collaborative closed loop of 11 sub-agents.
Before the consultation, the system automatically extracts the patient's medical history and completes quality control; during the consultation, AI assistants from various specialties output guideline-based diagnosis and treatment suggestions in parallel, and the chief AI summarizes them to form the consultation conclusion; after the consultation, the system automatically formulates a personalized follow-up plan and triggers graded early warnings when patients fail to return for follow-up on time, reminding patients, family members, attending doctors, and nurses to form a layered response. For doctors, this means every expert has a 24/7 AI assistant. The agent is like a secretary: "Equipping each expert with such an AI assistant can reduce their burden"—it organizes the guideline-recommended opinions, and doctors only need to review, supplement, or delete, without needing to think about everything from scratch.
Originally, initiating a consultation took hours to organize data and make PPTs; now, it can be automatically generated in ten minutes after data upload. The team can even implement it in real scenarios at any time: with two consultations a week, the agent and the real consultation can run synchronously for continuous training. The team plans to let the agent continuously accumulate real cases and gradually understand the tacit experience not written in the guidelines. This improvement in collaborative efficiency not only means a reduction in the organizational cost of a single consultation but also means that the full-cycle management from "screening → diagnosis → treatment → management → application" has a closed loop that can operate automatically for the first time.
Redefining the Relations of Production
Looking back, these three cases actually point to the same change: the relations of production in medical AI are being redesigned. In the past, those who defined the problems and those who solved them were disconnected. Engineers led the tool design, and if doctors had requirements, they could only write requirement documents, queue up for scheduling, and wait for replies from the IT department. The final delivered product might be miles away from their expectations, and time constraints made rework and iteration impossible.
Now, doctors directly participate in the design of the application layer, transforming clinical experience into system rules; technology vendors retreat to the foundation layer, doing a good job in the infrastructure of computing power, models, security, and governance. This division of labor abandons the simple allocation of Party A and Party B roles, instead empowering those who understand clinical problems best with the ability to create tools, allowing technology to truly serve frontline needs.
Healthcare is an industry with an extremely low tolerance for errors. During the development process, the doctors at Nanfang Hospital have always maintained their insistence on rigor: the aHUS agent strictly limits the knowledge base to retrieve only authoritative literature, setting up multi-agent cross-validation and final review by doctors; the obstetric training system explicitly has rules formulated by teaching doctors, with all condition evolutions clinically verified; and every sub-agent in the liver cancer MDT is based on the latest guidelines, with conclusions finally reviewed and confirmed by real experts. AI can assist in organizing and generating drafts, but the bottom line of medical logic must be built on guidelines, expert consensus, and continuously updated real cases, with the final decision-making power always remaining in the hands of doctors.
When doctors gain the ability to create tools, the innovation logic of medical AI is shifting from "what technology can do" to "what the clinic needs." These attempts at Nanfang Hospital are just early explorations. We have reason to believe that innovation will grow naturally from consulting rooms and teaching sites, and an excellent computing power foundation will allow this demand-driven innovation to continue happening, turning frontline voices directly into runnable tools, thereby achieving true technological equity.