At Nanfang Hospital, Clinicians Begin Building AI Tools Without Code

Leifeng.com reported that the change addresses a longstanding problem in medical AI: clinicians who understand clinical workflows cannot easily build tools, while engineers who can write code often need long periods to understand nuanced diagnostic decisions. Traditional outsourcing costs money and scheduling time, and many clinical ideas were abandoned before they were proposed, doctors told Leifeng.com.

In the nephrology department, Dr. Xie Di described a case of a 24-year-old woman with systemic lupus erythematosus who was hospitalized for breathing difficulty. Her blood tests pointed to atypical hemolytic uremic syndrome, or aHUS, a rare disease in which microthrombi block small blood vessels; acute-phase mortality can reach 25 percent. Effective intervention is limited to 24 to 48 hours after symptoms appear, and rapid identification and treatment within seven days produces better outcomes than treatment after seven days, according to the report.

aHUS diagnosis is an exclusionary process. Doctors must first identify thrombotic microangiopathy, then rule out thrombotic thrombocytopenic purpura, malignant hypertension, drug-induced causes and other possibilities, involving cross-judgments from eight or nine specialties including hematology, nephrology, infectious diseases and rheumatology. In the past, Xie said, the process relied on personal experience and manual coordination of multidisciplinary consultations. She said a traditional IT project for a high-risk screening clinical decision support system would require substantial outsourcing fees, and communication barriers and scheduling delays often meant the final product differed from what doctors wanted.

Using HAIP, Xie's team converted that diagnostic path into an agent architecture. A data perception agent retrieves key indicators such as blood routine results and schistocytes from the HIS; a layered reasoning agent follows a path of identifying TMA, confirming TMA and excluding other diseases; and an MDT collaboration agent simulates cross-examination by eight specialty doctors. The team added an arbitration agent to handle disagreements, with four levels: tracing evidence and checking for missing or inconsistent data; evaluating different opinions by evidence level, timeliness and the patient's situation; presenting specialty opinions, evidence and major disagreements transparently to a human doctor for a final decision; and recording disputed points, decision bases and conclusions as a traceable case library.

Xie also highlighted two designs she sees as defensible advantages. One is human-machine dialogue: unlike general large models, the system allows doctors to click and correct an erroneous AI judgment, after which the correction enters a knowledge iteration process involving expert review, version management and validation before updating the knowledge base. The other is a memory flywheel: because AI can err when parsing the knowledge base, offline MDT summaries continually feed back into the system; reviewed real-world cases accumulate and strengthen similar-case retrieval and reference, showing doctors historical cases, treatment processes and outcomes. Xie said she started from zero programming knowledge after attending lectures and training. With Nexent's zero-code, visual capabilities, she uploaded knowledge bases and the platform automatically parsed them, then she revised the framework step by step. From zero base to completing a preliminary competition prototype took her team ten days. Next, the team hopes to embed the system in the hospital's medical record system, using southbound interfaces to capture HIS data and northbound interfaces to return opinions, so that information scattered across departments can be integrated in real time for screening and warning within the golden window.

In obstetrics, Dr. Huo Zhifeng faced a different problem. Low-frequency but high-risk emergencies such as postpartum hemorrhage require rapid team coordination and repeated simulation training, but traditional case design takes a teacher 40 to 50 minutes, and debriefing relies on memory. Using HAIP, Huo turned clinical experience into an AI-assisted case generation and intelligent debriefing system. After a teacher uploads de-identified case materials and sets teaching goals and difficulty, Nexent generates a structured medical record draft in the background; the teacher reviews and publishes it, and the drill begins.

During a demonstration observed by Leifeng.com, students saw only an initial scenario and a countdown. The system did not volunteer extra information, and students had to assess and request data themselves, such as monitoring vital signs to receive blood pressure and heart rate feedback. The agent let the condition evolve under time compression: correct handling led to improvement; failure to intervene in time led to deterioration, with blood loss rising from hundreds of milliliters to more than a thousand, accompanied by pallor, agitation and other signs. A deterministic state machine engine controlled progression, but Huo said the rules were not generated by AI; they were set by doctors based on more than a decade of clinical and teaching experience and repeatedly validated. In debriefing, a process that previously took 15 to 20 minutes now produced a draft in about four seconds, recording what students did and offering targeted improvement suggestions. Huo later added voice recognition so the agent could distinguish participants' voices and strengthen team collaboration training.

Huo, who has no programming experience, described the experience as technology equality. He said he can turn teaching ideas directly into a runnable system by stating requirements and asking the AI to revise until satisfied. He argued the agents may be more valuable in primary hospitals than in large ones, because primary doctors may encounter severe postpartum hemorrhage only once every several years; simulation can maintain familiarity. He has received inquiries from ICU and emergency colleagues asking whether the agents can be opened to them. The system is expected to expand to eclampsia, emergency and ICU scenarios, with exploration of image recognition and even VR.

In hepatology, Liu Hongyan's team has worked on liver cancer multidisciplinary treatment for 15 years. They believe the gap in five-year survival between China and Japan lies less in surgical technique than in whole-process management: Japan performs well in early detection, early diagnosis and early treatment, while many Chinese patients are diagnosed at an advanced stage and lack continuous follow-up. Each consultation resembled organizing a meeting, with lead doctors preparing large amounts of material and slides, experts gathering at fixed times and places, and limited numbers of patients served. Follow-up management often broke off after the consultation. Liu said the team wanted online consultations ten years ago and tried an EMDT product in 2016, but it failed because of high technical barriers and operating costs. After Huawei and Nanfang Hospital began their cooperation in April, the team built a demo within more than a month based on HAIP. They converted early warning, quality control, specialty avatars and follow-up monitoring into a closed loop of 11 sub-agents.

Editor's Summary

Nanfang Hospital clinicians have used Huawei's HAIP and Nexent zero-code agent platform to build clinical AI tools for aHUS diagnosis, postpartum hemorrhage training and liver cancer MDT management, according to Leifeng.com. The projects shift some medical AI development from engineers to clinicians and are expected to extend into hospital systems and more emergency scenarios.