A Knowledge Graph-Based Intelligent Guidance System for Competency-Oriented Anesthesiology Residency Training: Design and Evaluation

In the high-stakes environment of an operating room, anesthesiologists manage a complex web of information that spans from the moment a patient is assessed before surgery to their recovery afterward. This field demands a mastery of hundreds of drugs, intricate procedures, and emergency protocols, all while maintaining a zero-tolerance policy for error. For doctors in training, known as residents, learning this vast system is often a fragmented experience. They might piece together knowledge from scattered textbooks or rely on sporadic, one-on-one teaching from busy senior physicians, leaving gaps in their understanding of how to handle critical emergencies. While artificial intelligence has begun to offer new ways to teach, many existing tools rely on cloud-based systems that send sensitive patient data over the internet, raising serious privacy concerns that prevent their use in secure hospital networks.

A team of researchers at Ningxia Medical University and its affiliated hospital in China has developed a new kind of digital teaching assistant designed specifically to solve these problems. They created a system that lives entirely within a hospital's own computers, keeping all data private and secure. Instead of simply answering questions, this tool guides residents through a step-by-step learning process that mimics the way a senior doctor would think through a complex problem. By combining a structured map of medical facts with advanced language technology, the system offers a way to practice high-risk scenarios without ever putting a real patient at risk.

The core of this new system is a specialized knowledge graph, which the researchers built by organizing thousands of facts about anesthesia into a logical structure. Imagine a vast library where every book, drug, and procedure is connected by invisible threads that show exactly how they relate to one another. The researchers mapped out the entire journey of a surgical patient, from preoperative assessment to postoperative pain management, and filled this map with 805 specific medical entities and over 6,500 relationships between them. This structure allows the computer to understand not just isolated facts, but the logical flow of clinical decision-making. Unlike general medical tools that focus on diseases, this map is built around the workflow of anesthesiology, reflecting how doctors actually work in the operating room.

To make this knowledge useful for learning, the system employs a method called progressive dialogue. When a resident asks how to handle a specific emergency, such as toxicity from a local anesthetic, the system does not simply dump a long list of instructions. Instead, it breaks the solution down into five sequential steps. It presents the first action, asks the resident to confirm they understand, and then moves to the next step only after receiving that confirmation. This approach forces the learner to actively engage with the material, building a mental habit of checking and verifying each stage of a procedure. The system also tracks the user's questions and history, adapting its guidance to fit the learner's pace and ensuring that the feedback is always tied to authoritative medical guidelines.

A major innovation in this work is how the system finds the right information to answer a question. The researchers built a dual-engine retrieval system that acts like a two-pronged search. One part looks through the structured knowledge graph to find precise, guideline-based facts, such as exact drug dosages or contraindications. The other part searches through unstructured text documents to find detailed explanations and context. The system then combines these two sources, giving more weight to the structured facts to ensure accuracy. In testing this approach against 300 questions created and graded by senior anesthesiologists, the system achieved a guideline compliance rate of 98.5 percent. This means that nearly every answer it gave was fully consistent with official medical standards, a significant improvement over standard language models that often hallucinate or invent facts.

The researchers also addressed the practical challenge of running this technology in a hospital setting. Many powerful artificial intelligence tools require massive, expensive servers or a constant connection to the internet, which is often impossible in secure medical environments. This new system was designed to be lightweight and fully local. By compressing the software, the team made it possible to run the entire application on a standard hospital workstation with a graphics card containing just 6 gigabytes of memory. The system operates completely offline, meaning no data ever leaves the building. In tests, it generated a complete, guided response in less than 15 seconds, a speed fast enough for a resident to use during a study session or while preparing for a shift.

The evaluation of the system revealed that while a standard language model without these specific guides could generate longer, more complete-sounding answers, it frequently included incorrect or unverified information. The specialized system, by contrast, prioritized safety and accuracy, ensuring that every piece of advice was backed by a cited source. The researchers noted that the system was particularly effective at teaching the management of critical events, a skill that is difficult to practice in real life because such emergencies are rare. By simulating these scenarios in a safe, interactive environment, the tool helps residents build the procedural memory and clinical reasoning habits they need before they ever face a real crisis.

While the results are promising, the authors are careful to note that this is a technical demonstration of a new teaching tool, not a final proof that it improves long-term medical outcomes. The current knowledge base relies on public textbooks and guidelines, and the team plans to expand it to include specific hospital cases and subspecialties like pediatric or cardiac anesthesia in the future. They are also preparing for a controlled educational study to see if using this system actually helps residents become better doctors over time. For now, the work stands as a demonstration that it is possible to build a sophisticated, secure, and highly accurate learning companion for medical training that respects the strict privacy rules of the healthcare world.