Multi-agent medical care teaching virtual tutor system, teaching method and medium

Abstract

The present application belongs to the technical field of medical care teaching system, and particularly relates to a multi-agent medical care teaching virtual tutor system, a teaching method and a medium. The system comprises task arranger, explanation tutor agent, patient simulation agent, observation and evaluation agent, error correction coach agent, business process module, mixed score calculator, knowledge base, evidence log module and other component units. The system of the present application can realize the unity of teaching safety, real-time intervention accuracy, quality traceability, large-scale personalization and long-term dialogue stability through the closed-loop architecture driven by data objects, real-time trigger type tutor intervention, key node forced verification, double-track fusion scoring, role consistency guarantee and personalized error correction path, and significantly improves the reliability, efficiency and teaching effect of the intelligent training system, and has good application prospect.

Description

Technical Field

[0001] This invention belongs to the technical field of medical and nursing teaching systems, specifically relating to a multi-agent medical and nursing teaching virtual tutor system, teaching method, and medium. Background Technology

[0002] Scenario simulation is a core method in medical and nursing education. Teachers typically design patient scenarios around teaching objectives, enabling learners to train their situational awareness, crisis recognition, clinical reasoning, and practical skills under highly realistic conditions. Current practices mainly include: in-person teacher instruction, standardized patient (SP) role-playing, scripted virtual patient question-and-answer sessions, and VR / AR virtual simulation systems. With the development of generative AI, products that use a single large language model as a question-and-answer teaching aid have emerged.

[0003] However, these existing technologies still cannot meet the needs of practical applications, and their shortcomings are as follows: (1) Insufficient multi-role collaboration: Single agent question answering or scripted virtual patients cannot replicate the multi-role relay and division of labor in real clinical practice, and lack a mechanism to support the timing arrangement of "who speaks / when to interrupt / how long to speak".

[0004] (2) Lack of mandatory verification of key processes: Key nodes such as identity verification, allergy history, three checks and seven verifications, and aseptic operation cannot be forcibly intercepted and rolled back, which allows trainees to "pass the test with illness".

[0005] (3) Subjectivity and lack of traceability in scoring: relying on teacher experience or single model scoring, lacking a chain of evidence, scoring consistency (κ) is difficult to guarantee, and it is also difficult to drive subsequent personalized branches.

[0006] (4) Lack of immediate error correction and adaptive explanation: Students’ mistakes are often reviewed in a unified manner afterward, which makes it impossible to correct the mistakes immediately when they occur, which can easily lead to the solidification of errors and the loss of knowledge.

[0007] (5) Long dialogues and context drift: In multi-turn interactions, problems such as role overstepping, context loss, and information duplication frequently occur, interfering with the continuity of learning.

[0008] (6) Lack of standardized orchestration protocols: Existing systems mostly focus on scene rendering and interaction, lacking verifiable and scalable system mechanisms such as "event bus + role consistency protocol + interrupt trigger + hybrid scoring arbitration + evidence log".

[0009] Although the introduction of new teaching systems provides a potential path to solve the above problems, the optimized design of their system architecture and the standardized construction of their operation process remain key technical challenges that urgently need to be overcome. Summary of the Invention

[0010] To address the problems of existing technologies, this invention provides a multi-agent medical and nursing teaching virtual tutor system, teaching method, and medium.

[0011] A multi-agent virtual tutor system for medical and nursing education includes: The task orchestrator is configured to enable control over dialogue and tasks. The instructor's intelligent agent is configured to intervene in the dialogue in real time based on preset trigger conditions, and output layered explanation content, step-by-step instructions, demonstration scripts and comparison tables. The patient-simulated intelligent agent is configured to generate dynamic dialogue content and physiological feedback data based on the case script and scenario parameters. The observation and evaluation agent is configured to extract structured elements from the trainees' statement input and operation logs, map the extraction results to the mandatory slots and guard conditions configured in the current node of the business process module, and output the sub-score results and evidence annotation information. The error correction coach agent is configured to generate itemized error correction prompts, explanations of error reasons, and re-practice paths when it detects missing items in slots or scores below a threshold. The business process module is configured to define the set of nodes in the teaching process, the transition conditions between nodes, the parallel subtasks, the required slots for each node, the guard conditions, and the rollback strategy. The hybrid scorer is configured to combine rule-based scoring with large language model scoring, and output the confidence and difference values ​​for each scoring dimension. The knowledge base is configured to store nursing guidelines, pathways, and explanatory materials. The evidence log module is configured to record events, scoring results, arbitration decisions, and evidence fragments during system operation in an append-only structure. Each record includes a timestamp and an anti-tampering audit log mechanism. The anti-tampering audit log mechanism is implemented as a combination of hash chain evidence logs or read-only log partitions and signature snapshots, and supports full-process playback based on tracking identifiers.

[0012] Preferably, the application scenarios for medical and nursing education include, but are not limited to: pre-hospital emergency care teaching scenarios, perioperative safety check teaching scenarios, intravenous infusion operation teaching scenarios, urinary catheterization operation teaching scenarios, pressure ulcer assessment teaching scenarios, and pharmaceutical prescription review teaching scenarios.

[0013] Preferably, the task orchestrator maintains the business process module, thereby scheduling multi-agent turn-taking, executing priority strategy π and interruption / recovery; The task orchestrator has a built-in event-driven communication mechanism and a global context cache. The event-driven communication mechanism is implemented using a unified event bus or message middleware. The event-driven communication mechanism provides a publish-subscribe mechanism, a priority queue, and a time-series consistency guarantee. Each agent transmits messages and synchronizes its state through the event bus. The global context cache maintains session summaries, key facts, unmet slot queues, and student profile information, and adopts an update strategy that combines sliding windows and hierarchical summaries.

[0014] Preferably, the instructor AI agent has retrieval enhancement generation capabilities and rule gating capabilities, and supports outputting explanation results in a structured tool call manner.

[0015] Preferably, the implementation method of the business process module is selected from: process state machine, process model described by business process modeling markup language, Petri net model or partially observable Markov decision process model.

[0016] Preferably, when the difference value output by the hybrid scorer exceeds a preset threshold, the arbitration engine is invoked to make a decision. The arbitration engine is configured to: when the discrepancy between the rule score and the large language model score exceeds a threshold τ, make a decision in the order of "evidence priority > rule priority > LLM priority" and record the path.

[0017] Preferred options also include: The security gating module is configured to perform process compliance review and sensitive language detection on the output content of the instructor intelligent agent, patient simulation intelligent agent, observation and evaluation intelligent agent, and error correction coach intelligent agent, and to perform rejection or require rewriting operations on the illegal content.

[0018] Preferably, the task orchestrator creates and maintains a session state vector, which is used to record the current node identifier, the bitmap representation of the slots that have been satisfied, the status of each guard condition, the cumulative score vector, and the context hash value. The session state vector is updated in the observation and evaluation agent and the hybrid scorer, and is read and used in the explanation tutor agent, the error correction coach agent, and the business process module. The observation and evaluation agent generates an evaluation evidence package, which includes the original input fragment, extracted slot values, the mapping rule identifier used, the confidence level of each extracted item, slot satisfaction status, guard condition pass status, missing slot list, failed guard list, and evidence hash value. After the evaluation evidence package is published via an event-driven communication mechanism, it is used by the hybrid scorer to perform score calculation, by the evidence log module for persistent storage, and by the explanation tutor agent to determine the explanation content. The hybrid scorer generates a trigger decision token based on the score results and preset trigger conditions. The trigger decision token includes trigger type, priority value, target agent identifier, interruption policy and context snapshot reference. The trigger decision token is used to notify the task orchestrator to start the intervention process of the explanation tutor agent or the error correction coach agent. The instructor agent generates an explanation output package, which includes the intervention type, explanation granularity, explanation content, tool call result list, knowledge item reference list, and gate pass flag. After being reviewed by the security gate module, the explanation output package is forwarded to the front end for presentation by the task orchestrator and recorded in the evidence log module.

[0019] This invention also provides a method for medical teaching using the aforementioned multi-agent medical and nursing teaching virtual tutor system, comprising the following steps: Scene initialization: The task orchestrator loads the case script and business process module configuration, creates and initializes the session state vector, obtains student profile information, writes it to the global context cache, and the instructor's intelligent agent selectively executes the opening explanation and publishes the explanation output package; Dialogue relay: The front end pushes the student's input to the event-driven communication mechanism. The task orchestrator encapsulates the assessment request package and distributes it to the observation assessment agent and the patient simulation agent. Each agent performs role consistency checks to detect context conflicts. Real-time evaluation: Observe and evaluate the agent's response to student input, perform named entity recognition and slot mapping, perform slot verification and guard condition verification based on the current node configuration, generate and publish an evaluation evidence package containing extraction results, slot satisfaction status, guard pass status, missing slot list and failed guard list. Tutor Intervention: Upon receiving a trigger decision token containing missing slots, failure guards, or comprehensive scores via a subscription mechanism, the tutor agent first integrates contextual information, including student profiles and historical error patterns. Then, the tutor agent executes a dual-gating process to generate explanation content. The first layer is a retrieval-enhanced generation gating system, which retrieves and filters relevant knowledge from the knowledge base. The second layer is a rule-based gating system, which uses a large language model to generate a content draft and verifies and rewrites it according to key rules (e.g., three checks and seven verifications) to ensure compliance. Next, the tutor agent dynamically selects the explanation granularity based on the student's professional level and the specific trigger type. This granularity includes step-by-step breakdown, key point hints, or quick hints. Finally, the generated information is packaged into an explanation output package including explanation content, knowledge references, and tool call results. After final review and purification by an independent security gating module, it is published to the unified event bus. Hybrid scoring: The hybrid scorer receives the evaluation evidence package, performs weighted fusion of rule-based scoring and model scoring to generate a comprehensive score, and the scoring arbitration module performs consistency verification and manual review triggering judgment on disputed scores; the hybrid scorer generates a trigger decision token based on the scoring results and preset triggering conditions; Status Update: The task orchestrator receives the explanation output packet and updates the session summary in the global context cache, and determines the subsequent flow control strategy of suspension or parallel injection based on the interruption policy; Evidence Recording: The evidence log module receives the evaluation evidence package, arbitration results and explanation output package, and generates log entries containing hash chain structure to support full process playback and integrity verification; Branch migration: The task orchestrator performs node migration or locking based on whether the guard conditions are met and the slots are satisfied. For guard conditions that are not met, a rollback strategy is executed and a re-practice plan is generated, realizing the linkage control between scoring and branching. Report generation: The task orchestrator generates a structured teaching report containing personalized improvement suggestions based on the evidence log statistics node pass rate and score distribution, and provides a full-process replay interface to support post-event auditing.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon for implementing the above-described method.

[0021] This invention provides a system and method for medical and nursing education. The technical solution of this invention has the following beneficial technical effects: 1. This invention establishes clear data dependencies and data flow paths between modules by defining four core data objects: session state vector, evaluation evidence package, trigger decision token, and explanation output package. Each data object performs creation, transformation, aggregation, and consumption operations in a predetermined order in the designated module, forming a complete data processing closed loop, which enables the system to have good scalability and maintainability.

[0022] 2. This invention achieves an organic combination of automated closed-loop teaching and controllable tutor intervention. Through a trigger decision token mechanism, the system can immediately trigger the intervention of the instructor's intelligent agent when a student makes a mistake, reducing the persistence of errors and improving learning efficiency.

[0023] 3. This invention implements mandatory verification of key nodes by using a slot satisfaction bitmap and a guard condition state array. When a required slot is not satisfied or the guard condition is not passed, node locking and rollback operations are performed, which significantly reduces the omission rate of key operations such as identity verification, allergy history inquiry, and sterile field maintenance.

[0024] 4. This invention integrates rule-based scoring and large language model scoring through the dual-track scoring of the hybrid scorer and the priority chain adjudication mechanism of the arbitration engine, and supports the traceability of scoring results through the evidence hash chain, thereby improving scoring consistency and interpretability.

[0025] 5. This invention effectively reduces context drift and role overreach in multi-turn dialogues by using a sliding window and hierarchical summary strategy based on role consistency verification and global context caching, thus maintaining the continuity of the dialogue and the fidelity of the roles.

[0026] 6. This invention improves the efficiency of improving weak points by explaining the granularity selection mechanism and the re-practice path generation of the error correction coach agent, thereby realizing personalized teaching feedback based on student profiles and sub-item scores.

[0027] 7. This invention provides complete data support for teaching quality control and teaching research review through an append-only evidence log and hash chain replay mechanism, which facilitates post-audit and course optimization.

[0028] In summary, this invention achieves a balance between teaching security, real-time intervention accuracy, quality traceability, scalable personalization, and long-term dialogue stability through a data object-driven closed-loop architecture, real-time triggered tutor intervention, mandatory verification of key nodes, dual-track integrated scoring, role consistency assurance, and personalized error correction paths. This significantly improves the reliability, efficiency, and teaching effectiveness of intelligent training systems.

[0029] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0030] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0031] Figure 1 This is a diagram showing the overall system structure and module connections. Figure 2 A diagram showing the flow and transformation of core data objects between modules; Figure 3 A flowchart illustrating the multi-agent message timing and interruptible interruption process; Figure 4 A schematic diagram showing the configuration of the nursing process state machine, node slots, and guard conditions; Figure 5 This is a flowchart for branch migration and rollback control.

[0032] The components are as follows: 101-Task Orchestrator; 102-Lecturer / Tutor Agent; 103-Patient Simulation Agent; 104-Observation and Evaluation Agent; 105-Error Correction Coach Agent; 106-Process State Machine; 107-Hybrid Scorer; 108-Knowledge Base; 109-Evidence Log Module; 110-Security Gating Module; 111-Global Context Cache; 112-Unified Event Bus; 113-Arbitration Engine. 201 - Session State Vector; 202 - Evaluation Evidence Packet; 203 - Trigger Decision Token; 204 - Explanation Output Packet; 301 - Student Input; 302 - Evaluation Request Package; 303 - Arbitration Request; 304 - Arbitration Result; 305 - Log Entries. Detailed Implementation

[0033] It should be noted that the algorithms for data acquisition, transmission, storage and processing steps not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described, can all be implemented using content already disclosed in the prior art.

[0034] Example 1: A Virtual Tutor System for Medical and Nursing Education Based on Multi-Agent Orchestration This embodiment provides a system for medical and nursing education, which is applicable to various medical and nursing operation scenarios, including but not limited to: pre-hospital emergency care teaching scenarios, perioperative safety check teaching scenarios, intravenous infusion operation teaching scenarios, urinary catheterization operation teaching scenarios, pressure ulcer assessment teaching scenarios, and pharmaceutical prescription review teaching scenarios.

[0035] Specifically, the system composition of this embodiment is as follows: Figure 1 As shown, it specifically includes: Task orchestrator 101: The core control unit for dialogue and tasks, maintains the process state machine 106, schedules multi-agent turn-based dialogue, executes priority policy π and interrupt / recovery of interruption; built-in unified event bus 112 and global context cache 111.

[0036] The instructor AI agent 102 has the ability to generate enhanced retrieval data and rule gating. It can intervene in the dialogue in real time based on preset trigger conditions, output layered explanation content, step-by-step instructions, demonstration scripts and comparison tables, and supports outputting explanation results in a structured tool call method.

[0037] Patient simulation agent 103: Generates dynamic dialogue content and physiological feedback data based on case scripts and scenario parameters, including vital sign values, pain scores, etc. The patient simulation agent can respond to the trainee's operation behavior and provide corresponding feedback.

[0038] Observation and evaluation agent 104: Extracts structured elements from trainees' statement inputs and operation logs, maps the extraction results to the mandatory slots and guard conditions configured in the current node of the process state machine, and outputs sub-item scoring results and evidence annotation information.

[0039] Error Correction Coach Agent 105: When a missing item is detected in a slot or the score is below the threshold, it generates itemized error correction prompts, explanations of the error reasons, and re-practice paths.

[0040] Process state machine 106: Defines the set of nodes in the teaching process, the transition conditions between nodes, parallel subtasks, the required slots for each node, the guard conditions, and the rollback strategy.

[0041] Hybrid scorer 107: Integrates rule-based scoring and large language model scoring, outputs the confidence and difference values ​​of each scoring dimension, and calls the arbitration engine 113 to make a decision when the difference exceeds a preset threshold.

[0042] Knowledge Base 108: Stores nursing standards, pathways, and explanatory materials. The sources include nursing standards, pathways, and explanatory materials, presented in both free and structured formats, and supports RAG retrieval and rule validation.

[0043] Evidence Log Module 109: An append-only structure records events, scoring results, arbitration decisions, and evidence fragments during system operation. Each record includes a timestamp and hash digest, and supports full-process playback based on tracking identifiers.

[0044] Security Gating Module 110: Performs process compliance review and sensitive word detection on the output content of each intelligent agent, and performs rejection or requires rewriting operations on non-compliant content.

[0045] Global Context Cache 111: Maintains session summaries, key facts, unmet slot queues, and student profile information, employing an update strategy that combines sliding windows and hierarchical summaries.

[0046] Unified Event Bus 112: Provides publish-subscribe mechanism, priority queue and timing consistency guarantee. Each agent passes messages and synchronizes its state through the event bus.

[0047] Arbitration Engine 113: When the discrepancy between the rule score and the large language model score exceeds the threshold τ, the arbitration is carried out in the order of "evidence priority > rule priority > LLM priority" and the path is recorded.

[0048] To enable data association and flow between the aforementioned modules, this embodiment defines four core data objects, such as... Figure 2 , Figure 3 As shown, the four core data objects include: The first type of core data object is the session state vector, which is created and maintained by the task orchestrator. It records the current node identifier, the bitmap representation of the slots that have been satisfied, the status of each guard condition, the cumulative score vector, and the context hash value. The session state vector is updated in the observation and evaluation agent and the hybrid scorer, and is read and used in the explanation tutor agent, the error correction coach agent, and the process state machine.

[0049] The second type of core data object is the evaluation evidence package, which is generated by the observation and evaluation agent. It includes the original input fragment, the extracted slot values, the mapping rule identifier used, the confidence level of each extracted item, the slot satisfaction status, the guard condition pass status, the missing slot list, the failed guard list, and the evidence hash value. After the evaluation evidence package is published via the unified event bus, it is used by the hybrid scorer to perform score calculation, by the evidence log module for persistent storage, and by the explanation tutor agent to determine the explanation content.

[0050] The third type of core data object is the trigger decision token, which is generated by the hybrid scorer based on the scoring results and preset trigger conditions. It includes the trigger type, priority value, target agent identifier, interruption policy, and context snapshot reference. The trigger decision token is used to notify the task orchestrator to start the intervention process of the explanation tutor agent or the error correction coach agent.

[0051] The fourth type of core data object is the explanation output package, which is generated by the explanation tutor intelligent agent. It includes intervention type, explanation granularity, explanation content, tool call result list, knowledge item reference list, and gate pass mark. After the explanation output package is reviewed by the security gate module, it is forwarded to the front end for presentation by the task orchestrator and recorded in the evidence log module.

[0052] The modules mentioned above establish data associations through the four core data objects to form a complete data processing closed loop. Each data object performs creation, transformation, aggregation, and consumption operations in a predetermined order within the designated module.

[0053] In a preferred embodiment, the parameters are preferably set as follows: The parameters for the priority scheduling strategy are configured as follows: the weight of the task orchestrator's overall guard condition failure, the number of missing slots, the response delay duration, the conflict level, and the score confidence are set to dynamically prioritize, and the recommended value range for the response delay threshold is six to twelve seconds.

[0054] The arbitration threshold τ is configured as follows: The arbitration threshold is set adaptively according to the importance of the scoring dimension. The recommended value for the arbitration threshold of critical security items is 0.05, and the recommended value for the arbitration threshold of ordinary communication items is 0.1 to 0.15.

[0055] The global context cache parameters are configured as follows: the recommended window size for the sliding window is between 20 and 40 rounds of dialogue, and the update frequency of the hierarchical summary is adaptively adjusted according to the node complexity.

[0056] The trigger condition parameters are configured as follows: It is recommended that the threshold for the number of consecutive rollbacks be set to two. Once this threshold is reached, the system will automatically enter the reinforcement and re-practice mode.

[0057] As an equivalent technical solution, some technical means in this embodiment can be replaced as follows: The process state machine can be replaced by a process model described by a business process modeling markup language, a Petri net model, or a partially observable Markov decision process model. After the replacement, the configuration and verification functions of nodes, slots, and guard conditions can still be realized.

[0058] The knowledge base can be replaced with an ontology knowledge base or a knowledge graph. After the replacement, it can still support enhanced retrieval generation and rule-validation queries.

[0059] The unified event bus can be replaced by a message middleware implementation, and the publish-subscribe mechanism and timing consistency guarantee can still be provided after the replacement.

[0060] The hash chain evidence log can be replaced with a combination of read-only log partitions and signature snapshots, while still maintaining immutability and traceability.

[0061] The enhanced retrieval generation gate can be replaced by a combination of offline rule templates and lightweight retrieval, which is suitable for scenarios where a complete knowledge base cannot be deployed.

[0062] The scoring fusion can be replaced with a hierarchical Bayesian model or an end-to-end learning fusion model. Even after the replacement, the function of fusing rule-based scoring and model scoring can still be achieved.

[0063] Example 2: Medical and nursing teaching method using a virtual tutor system for medical and nursing education This embodiment uses the virtual tutor system for medical and nursing education provided in Embodiment 1 for medical teaching. The specific process is as follows: S1, Scene initialization and data object creation: The task orchestrator loads the case script and process state machine configuration, and reads the node set, the required slot definitions for each node, the guard condition definitions, and the transfer rules from the configuration. The task orchestrator registers the subscription relationships of each agent on the unified event bus and sets priority scheduling parameters for each agent.

[0064] The task orchestrator creates a session state vector and performs initialization, sets the current node identifier as the starting node, initializes all bits of the slot satisfaction bitmap to the unsatisfied state, initializes the guard condition state array to the state to be evaluated, initializes the score vector to null, and calculates the initial context hash value.

[0065] The task orchestrator obtains student profile information from the student management system, writes the student profile information into the global context cache, and writes the target node identifier and the initial queue of unsatisfied slots into the global context cache.

[0066] The instructor agent decides whether to perform the opening lecture based on the configuration. If the opening lecture is performed, an explanation output package containing teaching objectives and key points of the standard is generated and published to the unified event bus.

[0067] S2, dialogue relay and role consistency: When the front end receives student input, it pushes the student input to the unified event bus, injecting role identifier, tracking identifier, timestamp and sequence number into the message header.

[0068] The task orchestrator receives the student input via a subscription mechanism and reads the current session state vector from the global context cache. The task orchestrator encapsulates the student input and the session state vector into an evaluation request packet.

[0069] The task orchestrator distributes the assessment request packets to the observation assessment agent and the patient simulation agent via a unified event bus. The observation assessment agent receives the assessment request packets to perform feature extraction and verification operations, while the patient simulation agent receives the assessment request packets to generate response dialogue content.

[0070] At the message receiving end, each agent performs a role consistency check operation. The role consistency check operation includes: the receiving agent verifies whether the state hash value in the message header is consistent with the state hash value cached locally. If they are inconsistent, it is determined to be a context conflict. The receiving agent refuses to process the message and returns conflict exception information to the task orchestrator. The conflict exception information includes the expected hash value and the received hash value.

[0071] S3, Real-time Assessment and Interruption Trigger: The observation and evaluation agent performs named entity recognition and slot mapping operations on the student input and outputs an extraction result set. Each element in the extraction result set includes the slot key name, extraction value, confidence score, and source text fragment position.

[0072] The observation and evaluation agent loads the configuration information of the current node from the process state machine. The configuration information includes a list of required slots, a list of guard conditions, and a list of scoring rules.

[0073] The observation and evaluation agent performs slot verification. For each slot key name in the current node's required slot list, the observation and evaluation agent determines whether the slot key name exists in the extraction result set. If it exists and the corresponding confidence level is greater than or equal to the slot position confidence threshold, the bitmap position corresponding to the slot key name in the session state vector is set to the satisfied state; otherwise, the slot key name is added to the missing slot list.

[0074] The observation and evaluation agent performs guard condition verification. For each guard condition in the current node's guard condition list, the observation and evaluation agent evaluates whether the guard condition passes based on the current slot's satisfaction bitmap and the extraction result set, and writes the evaluation result to the corresponding position in the guard condition state array of the session state vector.

[0075] The observation and evaluation agent generates an evaluation evidence package. The evaluation evidence package contains the following fields: a tracking identifier, whose value is inherited from the evaluation request package; a current state identifier, whose value is the identifier of the current node; a raw input hash value, whose value is obtained by performing a hash operation on the student input; an extraction result, whose value is the set of extraction results; a slot satisfaction status, whose value is the updated slot satisfaction bitmap; a guard pass status, whose value is the updated guard condition status array; a missing slot list; a failed guard list, whose elements are the failed guard conditions in the guard condition status array; a timestamp; and an evidence hash value, whose value is obtained by performing a hash operation on the above fields.

[0076] The observation and evaluation agent publishes the evaluation evidence package to a unified event bus, with subscribers including a hybrid scorer, a task orchestrator, and an evidence log module.

[0077] S4, Mentor Intervention and Rule Gating: The instructor agent receives trigger decision tokens through a subscription mechanism and performs context extraction operations. It reads student profile information and historical error pattern information from the global context cache, and extracts the missing slot list, failed guard list, and scoring vector from the trigger decision tokens.

[0078] The instructor agent performs a retrieval enhancement generation gating operation, which is the first gating operation. Based on the missing slot list, the failed guard list, and the current node topic, the instructor agent constructs a retrieval query and sends it to the knowledge base for retrieval. The knowledge base returns the top few knowledge entries ranked by relevance. The instructor agent calculates the relevance score of each returned knowledge entry to the retrieval query; if the relevance score is lower than the relevance threshold, the entry is filtered. If the set of filtered knowledge entries is empty, the instructor agent reverts to using a local rule template.

[0079] The instructor's agent performs a rule-based gating operation, which is the second gating. Based on the filtered set of knowledge items and the session state vector, the instructor's agent calls a large language model to generate a draft of the explanation content. The instructor's agent reads the set of key rules corresponding to the current node type. This set of key rules includes rules such as the three-check-seven-verification rule and aseptic operation rules. For each rule in the key rule set, the instructor's agent checks whether the draft explanation content conforms to that rule. If a rule does not conform, the instructor's agent calls the large language model to rewrite the draft explanation content based on the violation feedback information of that rule. The rewriting operation is performed a maximum of a preset number of times. If the rule still does not conform after the preset number of rewrites, a rule template is used to replace the draft explanation content.

[0080] The instructor agent performs the granularity selection operation. The instructor agent determines the granularity based on the professional level field in the student's profile and the trigger type in the trigger decision token. If the professional level is "beginner" or the trigger type is "key deficiency," the granularity is set to step-by-step decomposition mode, which includes step-by-step explanations and sample scripts. If the trigger type is "low score trigger," the granularity is set to key point prompt mode. Otherwise, the granularity is set to quick prompt mode.

[0081] The instructor's intelligent agent generates an explanation output package. This output package contains the following fields: tracking identifier; intervention type (its value is the trigger type); explanation granularity; explanation content (its value is the gating-processed explanation content); a tool call result list, where each element contains the tool function name, call parameters, and return result. The tool functions include a checklist verification tool, a guard condition check tool, a scoring rule tool, and a demonstration generation tool; a knowledge item reference list, containing the identifiers of filtered knowledge items; a gating pass flag, which records the pass status of the retrieval enhancement generation gating and rule gating; and an output hash value, obtained by performing a hash operation on the above fields.

[0082] The security gating module reviews the content of the presentation package. It incorporates a sensitive word list, process compliance rules, and a risk classifier to perform sensitive term detection and process compliance checks on the presentation content. If any violations are detected, the security gating module cleans up or rewrites the presentation content and updates the security gating field in the gating pass flag to a rewritten state; if no violations are detected, the security gating field is updated to a pass state.

[0083] The instructor agent publishes the explanation output package, which has been processed by security gating, to the unified event bus.

[0084] S5, Hybrid Scoring Fusion and Arbitration: In the i-th discrimination, the hybrid scorer combines the rule score R and the large model score L to give the rule scoring results. Scoring results with large models The overall score is a weighted sum: , in, , As weight, This is a comprehensive score.

[0085] The confidence level of the rating is determined by whether the difference between the two scores exceeds a threshold. .

[0086] Rating confidence for: when When the score is 1, the arbitration engine initiates the arbitration decision, determines the final score according to "evidence priority > rule priority > LLM priority", and records the arbitration reasons and path.

[0087] The specific implementation involves the following three steps: First priority: the principle of evidence first.

[0088] This is the highest priority judgment in arbitration. The arbitration engine first retrieves and analyzes the assessment evidence package relevant to the current judgment scenario. The system will determine whether there is directly verifiable objective evidence that can clearly support a certain score result. The objective evidence includes, but is not limited to: accurate system error logs, specific status codes returned from authoritative third-party interfaces, and tamper-proof user operation records.

[0089] Ruling Action: If such objective evidence exists, and the logical direction of the evidence perfectly matches the judgment basis of a certain scoring rule, then the arbitration engine will unconditionally adopt the rule score generated by that rule. This will serve as the final score for this determination, terminating subsequent arbitration procedures. The arbitration reasoning will be recorded as "Score Based on the Adoption Rule of Verifiable Evidence".

[0090] Second priority: Rule priority principle.

[0091] If the first priority condition is not met (i.e., there is no directly verifiable strong evidence), the arbitration engine will proceed to this step. The system will analyze the rules that triggered this scoring. The system determines the specific set of rules and whether it contains a predefined "deterministic rule." A deterministic rule refers to a rule with clear business logic, no room for ambiguity, and proven high reliability in historical applications. The decision-making action is as follows: If the current score matches at least one deterministic rule, the system will assume that the deterministic rule has higher judgment power than the reasoning score of the large model and will adopt the rule's score. As the final score, the arbitration reason will be recorded as "Score based on adoption rule based on deterministic rule".

[0092] Third priority: Model-weighted arbitration principle (reflecting the priority of LLM).

[0093] If neither of the aforementioned priority conditions is met, it indicates that the current discrimination scenario falls under a complex or ambiguous situation that the rule system cannot precisely cover. In this step, to reflect the idea of ​​"LLM priority," that is, to acknowledge and utilize the advantages of large language models in handling scenarios with undefined rules and requiring deep contextual understanding.

[0094] Scores and evidence are written to the evidence log, and the global context cache is updated simultaneously.

[0095] S6, Branch Migration and Rollback: The task orchestrator receives the tutorial output package via a subscription mechanism and performs the following operations: The task orchestrator writes a summary of the explanation content from the explanation output package to the global context cache and updates the session summary field in the global context cache. The task orchestrator uses a strategy combining a sliding window and hierarchical summaries to update the global context cache. The sliding window size is the preset number of dialogue turns, and the hierarchical summaries are updated adaptively according to the node complexity.

[0096] The task orchestrator forwards the explanation output package to the front end for presentation.

[0097] The task orchestrator determines the subsequent process based on the interruption policy in the trigger decision token. If the interruption policy is to suspend the current turn, the task orchestrator waits for the student's response to the explanation before resuming the original turn; if the interruption policy is parallel injection, the task orchestrator allows the student to continue the current turn while sending the explanation.

[0098] S7, Reporting and Retrospection: The evidence log module receives evaluation evidence packages, arbitration results, and explanation output packages through a subscription mechanism and performs append-only recording operations.

[0099] The evidence log module generates a log entry for each record. The log entry contains the following fields: entry identifier; tracking identifier, used to associate all records in the same session; timestamp; sequence number; executor identifier, whose value is the identifier of the agent that generated the record; status identifier, whose value is the node identifier when the record was generated; action description; scoring result; confidence level; evidence hash value; and preceding hash value, whose value is the evidence hash value of the previous log entry.

[0100] The evidence log module establishes a hash chain relationship between log entries through the preceding hash value field, supporting full-process playback and integrity verification based on the tracking identifier.

[0101] Step S8, Branch Migration and Rollback Control: like Figure 5 As shown, the task orchestrator performs branch migration decision operations based on the updated session state vector.

[0102] If all guard conditions for the current node are met and all required slots are satisfied, the scoring and branching linkage control is executed. If the scores for all scoring dimensions are higher than the passing threshold for that dimension, the task orchestrator reads the transition rules for the current node from the process state machine, determines the next node based on the transition rules, and updates the current node identifier in the session state vector to the next node identifier.

[0103] The task orchestrator performs linked control of scoring and branching. If the score of a certain scoring dimension is lower than the passing threshold for that dimension, the task orchestrator marks the skill item corresponding to that scoring dimension as a re-practice state. The trainee must complete the re-practice of that skill item and reach the passing threshold before the migration to the subsequent node can be triggered.

[0104] If a failed guard condition exists, the task orchestrator performs a node locking operation, preventing the process from migrating to subsequent nodes. The task orchestrator reads the rollback strategy of the current node from the process state machine, determines the rollback target node based on the rollback strategy, and calls the error correction coach agent to generate a re-practice plan. The re-practice plan includes targeted practice content and practice paths for the failed guard conditions.

[0105] Step S9, Report Generation and Traceability Support: When the teaching process ends or a student requests it, the task orchestrator triggers a report generation operation.

[0106] The task orchestrator reads all log entries corresponding to the current tracking identifier from the evidence log module and calculates the pass rate of each node, the score distribution of each scoring dimension, the number of arbitrations, and the distribution of arbitration results.

[0107] The task orchestrator generates structured teaching reports based on statistical results and student profiles. These reports include node pass rates, sub-item scores, arbitration record summaries, evidence links, and personalized improvement suggestions based on weakness analysis.

[0108] The evidence log module provides a full-process replay interface based on tracking identifiers, supporting teaching administrators in conducting post-event audits and teaching research reviews.

[0109] The teaching method of this embodiment will be illustrated below through three specific application scenarios: Example 1: Objective Structured Clinical Examination Scenario for Acute Chest Pain Patient Reception Scene initialization configuration: like Figure 4 As shown, the process state machine configuration includes five nodes: reception node, initial assessment node, treatment node, education node, and shift handover node. The mandatory slots for the reception node include the identity verification slot, allergy history slot, and chief complaint duration slot. The guard conditions for the reception node require that both the identity verification slot and the allergy history slot be satisfied. The trigger parameters are configured as follows: response delay threshold of eight seconds, arbitration threshold of 0.05 for critical items, and arbitration threshold of 0.15 for general items.

[0110] Example of data processing procedure: The trainee entered dialogue content at the consultation node but omitted asking about allergy history. The observation and evaluation agent performed slot mapping on the trainee's input. The confidence level of the dual identity cores for the slots in the extracted result set was 0.92. The allergy history slot did not appear in the extracted results, and the confidence level of the chief complaint duration slot was 0.85. The observation and evaluation agent added the allergy history slot to the missing slot list, generated an evaluation evidence package, and published it to the unified event bus.

[0111] The hybrid scorer receives the evaluation evidence package. The rule-based scoring results show a perfect score for identity verification, a zero score for allergy history, and the chief complaint assessment dimension is pending evaluation. The large language model scoring results show a communication clarity score of 0.7. The hybrid scorer detects that the missing slot list is not empty, generates a trigger decision token, with the trigger type being critical missing, a priority of 0.95, the target agent being the instructor agent, and the interruption strategy being to suspend the current turn.

[0112] The instructor's intelligent agent receives a trigger decision token and reads the student's profile from the global context cache, displaying a professional level of "beginner." The instructor's intelligent agent constructs a retrieval query, and the knowledge base returns an entry for the acute coronary syndrome treatment guidelines, identified as knowledge entry "Acute Coronary Syndrome 001." The instructor's intelligent agent determines the explanation granularity to be a step-by-step decomposition mode, generates an explanation output package, and the explanation content includes step-by-step explanations and demonstration questions. The tool call result list includes the call results of the demonstration generation tool, and the knowledge entry reference list includes the entry's identifier.

[0113] The task orchestrator forwards the explanation output package to the front end and simultaneously performs a node locking operation. The error correction coach agent generates a re-practice path, including asking about allergy history, system review, and migrating to the initial review node after the guard conditions are met.

[0114] Report generation results: The teaching report shows that the pass rate for the patient reception point was 90%, and the pass rate was 90% after the first failure. The sub-scores show that the allergy history dimension achieved a perfect score after further practice. The personalized suggestions indicate that the focus of further practice should be on the order of asking key questions.

[0115] Example 2: Standard dressing change procedure scenario Scene initialization configuration: The process state machine configuration includes four nodes: preparation node, sterile draping node, wound cleaning node, and covering node. The guard conditions for the sterile draping node include hand hygiene completion, sterile field maintenance, and no cross-contamination between the instrument area and the contaminated area.

[0116] The data processing procedure is illustrated below: During the aseptic draping procedure, the trainee enters the contaminated area. The observation and evaluation agent performs guard condition checks on the trainee's operation log. The condition that the instrument area and contaminated area do not overlap is evaluated as a failure, and an evaluation evidence package is generated. The failed guard list includes the stated condition.

[0117] The hybrid scorer receives the evaluation evidence package, detects that the list of failed guards is not empty, generates a trigger decision token, the trigger type is critical missing, and the target agent is the lecturer agent.

[0118] The instructor's intelligent agent performs dual-gating processing, retrieving and enhancing the gating return of aseptic operation specification entries. The rule-based gating verifies that the draft explanation content conforms to aseptic operation rules. The instructor's intelligent agent generates an explanation output package, which includes explanations of the principles and risks of cross-zone contamination, and step-by-step instructions for proper towel laying and item retrieval.

[0119] The task orchestrator performs node locking and rollback operations, with the rollback target node being the preparation node. The error correction coach agent generates a retraining path, including repaving the sterile field, practicing the object retrieval action, and migrating to the debridement node after the guard conditions are met.

[0120] Example 3: Three checks and seven verifications for oral medication scenario Scene initialization configuration: The process state machine configuration includes four nodes: verifying medical orders and medications, verifying patients, administering medication and providing education, and observing adverse reactions. The mandatory slots for verifying medical orders and medications include medication name, dosage, time, route of administration, and batch number. The mandatory slots for verifying patients include patient name and wristband. The medication and education node requires all three checks and seven verifications to be met before medication administration can proceed.

[0121] The data processing procedure is illustrated below: Trainees were checking for missing batch number records in the medical orders and medication records. The evaluation agent performed slot verification; the batch number slot was not found in the extraction results. An evaluation evidence package was generated, and the missing slot list included the batch number slot.

[0122] The hybrid scorer performs dual-track scoring. The rule-based scoring result shows a score of zero for the batch number dimension, while the large language model scoring result, based on semantic reasonableness, gives a score of 0.6 for the batch number dimension. The hybrid scorer calculates the absolute value of the score difference to be 0.6, which is greater than the key item arbitration threshold of 0.05, and generates an arbitration request to be sent to the arbitration engine.

[0123] The arbitration engine determines that there is clear evidence in the evaluated evidence package that the batch number slot is missing. Following the principle of evidence priority, the engine uses the rule-based scoring result as the final score, and the arbitration reason is recorded as evidence priority. The arbitration engine outputs the arbitration result, and the evidence chain references the evidence hash value containing the evaluated evidence package.

[0124] The hybrid scorer generates a trigger decision token, and the instructor agent generates a presentation output package containing batch number record examples and electronic form operation instructions.

[0125] The evidence log module generates log entries with the action description of missing batch number slot, a score of zero, a confidence level of 0.99, and establishes a hash chain relationship between the evidence hash value and the preceding hash value.

[0126] As can be seen from the above embodiments, the system of the present invention realizes medical and nursing teaching with multi-agent arrangement, and has the advantages of high reliability, high efficiency and good teaching effect. It has high application value in medical and nursing teaching in various scenarios.

Claims

1. A multi-agent virtual tutor system for medical and nursing education, characterized in that, include: The task orchestrator is configured to enable control over dialogue and tasks. The instructor's intelligent agent is configured to intervene in the dialogue in real time based on preset trigger conditions, and output layered explanation content, step-by-step instructions, demonstration scripts and comparison tables. The patient-simulated intelligent agent is configured to generate dynamic dialogue content and physiological feedback data based on the case script and scenario parameters. The observation and evaluation agent is configured to extract structured elements from the trainees' statement input and operation logs, map the extraction results to the mandatory slots and guard conditions configured in the current node of the business process module, and output the sub-score results and evidence annotation information. The error correction coach agent is configured to generate itemized error correction prompts, explanations of error reasons, and re-practice paths when it detects missing items in slots or scores below a threshold. The business process module is configured to define the set of nodes in the teaching process, the transition conditions between nodes, the parallel subtasks, the required slots for each node, the guard conditions, and the rollback strategy. The hybrid scorer is configured to combine rule-based scoring with large language model scoring, and output the confidence and difference values ​​for each scoring dimension. The knowledge base is configured to store nursing guidelines, pathways, and explanatory materials. The evidence log module is configured to record events, scoring results, arbitration decisions, and evidence fragments during system operation in an append-only structure. Each record includes a timestamp and an anti-tampering audit log mechanism. The anti-tampering audit log mechanism is implemented as a combination of hash chain evidence logs or read-only log partitions and signature snapshots, and supports full-process playback based on tracking identifiers.

2. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, The application scenarios for medical and nursing education include: pre-hospital emergency care teaching scenarios, perioperative safety check teaching scenarios, intravenous infusion operation teaching scenarios, urinary catheterization operation teaching scenarios, pressure ulcer assessment teaching scenarios, and pharmaceutical prescription review teaching scenarios.

3. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, The task orchestrator maintains the business process module, thereby scheduling multi-agent turn-taking, executing priority strategy π and interruption / recovery; The task orchestrator has a built-in event-driven communication mechanism and a global context cache. The event-driven communication mechanism is implemented using a unified event bus or message middleware. The event-driven communication mechanism provides a publish-subscribe mechanism, a priority queue, and a time-series consistency guarantee. Each agent transmits messages and synchronizes its state through the event bus. The global context cache maintains session summaries, key facts, unmet slot queues, and student profile information, and adopts an update strategy that combines sliding windows and hierarchical summaries.

4. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, The described instructor AI agent possesses retrieval and enhancement generation capabilities as well as rule gating capabilities, and supports outputting explanation results in a structured tool invocation manner.

5. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, The implementation methods of the business process module are selected from: process state machine, process model described by business process modeling markup language, Petri net model or partially observable Markov decision process model.

6. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, When the difference value output by the hybrid scorer exceeds a preset threshold, the arbitration engine is invoked to make a decision. The arbitration engine is configured to: when the discrepancy between the rule score and the large language model score exceeds a threshold τ, make a decision in the order of "evidence priority > rule priority > LLM priority" and record the path.

7. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that, Also includes: The security gating module is configured to perform process compliance review and sensitive language detection on the output content of the instructor intelligent agent, patient simulation intelligent agent, observation and evaluation intelligent agent, and error correction coach intelligent agent, and to perform rejection or require rewriting operations on the illegal content.

8. The multi-agent medical and nursing teaching virtual tutor system according to claim 1, characterized in that: The task orchestrator creates and maintains a session state vector, which is used to record the current node identifier, the bitmap representation of the slots that have been satisfied, the status of each guard condition, the cumulative score vector, and the context hash value. The session state vector is updated in the observation and evaluation agent and the hybrid scorer, and is read and used in the explanation tutor agent, the error correction coach agent, and the business process module. The observation and evaluation agent generates an evaluation evidence package, which includes the original input fragment, extracted slot values, the mapping rule identifier used, the confidence level of each extracted item, slot satisfaction status, guard condition pass status, missing slot list, failed guard list, and evidence hash value. After the evaluation evidence package is published via an event-driven communication mechanism, it is used by the hybrid scorer to perform score calculation, by the evidence log module for persistent storage, and by the explanation tutor agent to determine the explanation content. The hybrid scorer generates a trigger decision token based on the score results and preset trigger conditions. The trigger decision token includes trigger type, priority value, target agent identifier, interruption policy and context snapshot reference. The trigger decision token is used to notify the task orchestrator to start the intervention process of the explanation tutor agent or the error correction coach agent. The instructor agent generates an explanation output package, which includes the intervention type, explanation granularity, explanation content, tool call result list, knowledge item reference list, and gate pass flag. After being reviewed by the security gate module, the explanation output package is forwarded to the front end for presentation by the task orchestrator and recorded in the evidence log module.

9. A method for conducting medical and nursing education using the multi-agent medical and nursing teaching virtual tutor system according to any one of claims 1-8, characterized in that, Includes the following steps: Scene initialization: The task orchestrator loads the case script and business process module configuration, creates and initializes the session state vector, obtains student profile information, writes it to the global context cache, and the instructor's intelligent agent selectively executes the opening explanation and publishes the explanation output package; Dialogue relay: The front end pushes the student's input to the event-driven communication mechanism. The task orchestrator encapsulates the assessment request package and distributes it to the observation assessment agent and the patient simulation agent. Each agent performs role consistency checks to detect context conflicts. Real-time evaluation: Observe and evaluate the agent's response to student input, perform named entity recognition and slot mapping, perform slot verification and guard condition verification based on the current node configuration, generate and publish an evaluation evidence package containing extraction results, slot satisfaction status, guard pass status, missing slot list and failed guard list. Tutor Intervention: Upon receiving a trigger decision token containing missing slots, failure guards, or comprehensive scores via a subscription mechanism, the tutor agent first integrates contextual information, including student profiles and historical error patterns. Subsequently, the tutor agent executes a dual-gating process to generate explanation content. The first layer is a retrieval-enhanced generation gating system, which retrieves and filters relevant knowledge from the knowledge base. The second layer is a rule-based gating system, which uses a large language model to generate a content draft and verifies and rewrites it according to key rules to ensure compliance. Next, the tutor agent dynamically selects the explanation granularity based on the student's professional level and the specific trigger type. This granularity includes step-by-step breakdown, key point hints, or quick hints. Finally, the generated information is packaged into an explanation output package including the explanation content, knowledge references, and tool call results. After final review and purification by an independent security gating module, it is published to the unified event bus. Hybrid scoring: The hybrid scorer receives the evaluation evidence package, performs weighted fusion of rule-based scoring and model scoring to generate a comprehensive score, and the scoring arbitration module performs consistency verification and manual review triggering judgment on disputed scores; the hybrid scorer generates a trigger decision token based on the scoring results and preset triggering conditions; Status Update: The task orchestrator receives the explanation output packet and updates the session summary in the global context cache, and determines the subsequent flow control strategy of suspension or parallel injection based on the interruption policy; Evidence Recording: The evidence log module receives the evaluation evidence package, arbitration results and explanation output package, and generates log entries containing hash chain structure to support full process playback and integrity verification; Branch migration: The task orchestrator performs node migration or locking based on whether the guard conditions are met and the slots are satisfied. For guard conditions that are not met, a rollback strategy is executed and a re-practice plan is generated, realizing the linkage control between scoring and branching. Report generation: The task orchestrator generates a structured teaching report containing personalized improvement suggestions based on the evidence log statistics node pass rate and score distribution, and provides a full-process replay interface to support post-event auditing.

10. A computer-readable storage medium, characterized in that, It stores a computer program for implementing the method of claim 9.

Images