Nursing education traceable, interpretable, and verifiable large model intelligent teaching assistant system and method
By employing technologies such as source fragment fingerprinting and causal evidence graph structured interpretation, the problem of tracing and interpretation in high-risk scenarios of the intelligent teaching assistant system for nursing education has been solved. This enables precise tracing and structured interpretation of answer fragments, ensuring the credibility and repeatability of the results and improving the usability and security of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANFU JIANGXI LAB
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-05
AI Technical Summary
Existing intelligent teaching assistant systems for nursing education cannot achieve precise source tracing at the fragment level, structured machine-verifiable interpretation, and deterministic result replay in high-risk scenarios. They suffer from problems such as crude source tracing capabilities, lack of operability in interpretation mechanisms, and insufficient feedback and verification mechanisms.
It employs a combination of technologies including source fragment fingerprinting, executable reference fragment binding, causal evidence graph structured interpretation, and joint verification of evidence consistency and counterfactual stability. Through input capture and semantic parsing, multi-source knowledge retrieval, executable reference fragment binding, causal evidence graph generation, and evidence consistency and counterfactual stability verification modules, it achieves the tracing, interpretation, and verification of answer text.
It achieves precise source tracing at the answer fragment level, structured explanation, and replay of deterministic results, improving the system's credibility and usability in high-risk educational scenarios, reducing the error output rate, and meeting the stringent auditing requirements of medical education.
Smart Images
Figure CN122152991A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent education and medical information technology, specifically to a traceable, explainable, and verifiable large-scale intelligent teaching assistant system and method for nursing education. Background Technology
[0002] Current intelligent teaching assistant systems for nursing education based on Large Language Models (LLM) primarily employ a "Retrieval Augmentation" (RAG) or "Knowledge Base + Dialogue Model" architecture. These systems typically retrieve relevant knowledge from training documents such as textbooks and clinical guidelines using semantic retrieval technology, and then use the LLM to generate natural language responses. In practice, common approaches include: data retrieval and paragraph-level citation suggestions based on vector databases; generating responses using prompting engineering or mind chain suggestions; simply recording dialogues and retrieval logs for basic teaching playback; and providing a visual interface to display "citation sources."
[0003] However, while such systems provide support for teaching, their reliability assurance mechanisms still have significant limitations in high-risk nursing education environments, mainly reflected in the following key deficiencies:
[0004] The system's tracing capabilities are rather crude and difficult to verify. Existing solutions typically only achieve document or paragraph-level citations, failing to reach the fine-grained tracing of sentences or answer fragments. Even under a fixed knowledge snapshot, multiple queries for the same question may not reproduce completely consistent answers and citation relationships, which fundamentally contradicts the stringent requirements of medical education that operational processes must be traceable and verifiable.
[0005] The system's explanation mechanism lacks operability in practical applications. Existing explanations are mostly natural language descriptions generated by large models, which are unstructured free texts. Such explanations are difficult to automatically verify by machines and cannot be effectively compared with contradictory evidence. When answer deviations occur, teachers find it difficult to quickly pinpoint whether the problem stems from retrieval errors, evidence quality, reasoning flaws, or inappropriate expression, which poses a substantial obstacle to error correction and guidance in the teaching process.
[0006] The system's feedback and verification mechanisms are significantly inadequate. A key issue is the lack of a "mandatory evidence binding" mechanism for important conclusions, potentially leading to system outputs lacking sufficient supporting evidence. Furthermore, the system lacks an independent evidence consistency verification function and has failed to establish a stability testing framework for counterfactual scenarios, such as verifying whether minor input adjustments lead to significant changes in conclusions. These deficiencies make it difficult for the system to effectively identify and control potential output risks, failing to meet the high standards of safety and accuracy required in nursing education. Summary of the Invention
[0007] This invention provides a large-scale intelligent teaching assistant system for nursing education that is traceable, interpretable, and verifiable. It solves the problem that existing technologies cannot achieve accurate traceability of answer fragments, structured machine-verifiable interpretation, and replayability of deterministic results in high-risk nursing education scenarios. This system combines source fragment fingerprint encoding, executable reference fragment-by-fragment binding, causal evidence graph structured interpretation, joint verification of evidence consistency and counterfactual stability, and deterministic replayability of results.
[0008] This invention is achieved through the following technical solution:
[0009] Firstly, this application provides a traceable, explainable, and verifiable large-scale intelligent teaching assistant system for nursing education, characterized by comprising:
[0010] The input capture and semantic parsing module is used to receive and parse the user's multimodal input, generate standardized question templates, and obtain the risk level information of the questions;
[0011] The multi-source knowledge retrieval and version tracing module is connected to the input capture and semantic parsing module. It is used to perform retrieval based on the standardized question template and output a set of candidate evidence with source fragment fingerprints for the retrieved knowledge fragments.
[0012] An executable reference and segment binding module is connected to the input capture and semantic parsing module and the multi-source knowledge retrieval and versioning tracing module. It is used to divide the answer into the smallest semantic unit according to the risk level information, and bind each smallest semantic unit to the source segment fingerprint in the candidate evidence set to generate the answer text and reference mapping relationship.
[0013] The causal evidence graph and structured interpretation generation module is connected to the executable reference and segment binding module, and is used to construct a causal evidence graph and generate a corresponding causal evidence graph identifier based on the answer text and reference mapping relationship;
[0014] The evidence consistency and counterfactual stability verification module is connected to the executable reference and segment binding module and the causal evidence graph and structured interpretation generation module. It is used to perform logical consistency verification and counterfactual stability testing on the answer text and the causal evidence graph, and output the verification results.
[0015] The replayable decoding and evidence storage module is connected to the executable reference and segment-by-segment binding module, the causal evidence graph and structured interpretation generation module, and the evidence consistency and counterfactual stability verification module. It is used to store the replayable decoding information, the answer text, the causal evidence graph, and the verification result. The replayable decoding information includes at least the reference mapping relationship and the causal evidence graph identifier.
[0016] A further optimization is to implement forced binding in the executable reference and fragment binding module as follows:
[0017] By using pointer network technology, the smallest semantic unit in the answer text is associated with one or more source fragment fingerprints, and the support strength score of each association is recorded in the reference map;
[0018] Gating processing is configured differently based on the risk level information received from the input capture and semantic parsing module: for high-risk scenarios, outputting answer fragments without binding evidence is prohibited; for important scenarios, answer fragments without binding evidence are displayed with warning labels; for ordinary scenarios, answer fragments without binding evidence are displayed with visual degradation.
[0019] A further optimization scheme is that, in the multi-source knowledge retrieval and version tracing module, the data structure of the source fragment fingerprint includes document identifier, version number, content hash value, and character offset.
[0020] A further optimization scheme is as follows: in the causal evidence graph and structured interpretation generation module, the node types of the causal evidence graph include at least claim nodes, evidence nodes, rule nodes, and constraint nodes, and the edge types include at least support edges, rebuttal edges, causal edges, subordinate edges, and time-limited edges; the structured description of the causal evidence graph is generated through controlled syntax based on the nursing domain ontology, and the corresponding graph structure hash value is generated.
[0021] The further optimized solution is as follows: the counterfactual stability test in the evidence consistency and counterfactual stability verification module is specifically as follows:
[0022] Construct counterfactual perturbation inputs covering at least one of the key nursing dimensions: gender, pregnancy status, age, liver and kidney function, and allergy history;
[0023] Resubmit the counterfactual perturbation input to the input capture and semantic parsing module, and trigger chained processing in subsequent modules;
[0024] Calculate the logical consistency score between the original output and the perturbed output. If the score is lower than the preset threshold, automatically trigger the reduction of the result confidence level or start the manual review process.
[0025] A further optimization scheme is that the replayable decoding information recorded in the replayable decoding and evidence storage module includes at least the large model version identifier used, the Merklegen hash value of the knowledge base snapshot, the random seed controlling the randomness of generation, the complete reference mapping data, and the graph structure hash value of the causal evidence graph.
[0026] A further optimized solution is to implement the evidence consistency verification in the evidence consistency and counterfactual stability verification module using a natural language reasoning model. Logical conflicts between evidence from different sources are detected by calculating the cosine similarity of sentence embedding vectors. When the similarity is lower than a set threshold, it is determined that there is a contradiction.
[0027] A further optimized solution is that the system is applicable to nursing education, skills assessment, patient safety training, pharmaceutical education, and emergency care training scenarios.
[0028] Secondly, this application provides a large-scale intelligent teaching assistant method for traceable, interpretable, and verifiable nursing education, characterized by the application of the large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described above, including the following steps:
[0029] The system parses user multimodal input, identifies entities and intents, determines risk levels, and generates standardized question templates.
[0030] Based on the standardized question template, a set of candidate evidence is retrieved from the version-managed nursing knowledge base using a fusion retrieval strategy;
[0031] By using pointer network technology, each smallest semantic unit is simultaneously bound to the fingerprint of a knowledge fragment in the candidate evidence set during the answer generation process, generating a reference mapping that records the supporting relationship, thereby determining the binding evidence that supports the answer;
[0032] Based on the relationship between the answer and the binding evidence, a structured causal evidence graph is constructed;
[0033] Perform evidence consistency verification and counterfactual stability testing on the answers and causal evidence diagrams, generate verification results including confidence scores and risk indicators, and record deterministic replayable decoding information;
[0034] It outputs the final answer by integrating and binding evidence, structured causal explanations, and multi-dimensional verification results, and simultaneously provides users with fragment-level citations, causal evidence diagram visualizations, and verification result explanations.
[0035] A further optimized solution is to include multimodal input, including text, voice, and images;
[0036] The smallest semantic unit is bound through a pointer network;
[0037] Evidence consistency verification is performed using a pre-trained natural language reasoning model.
[0038] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0039] The system parses multimodal user input, generates standardized question templates and risk level information, and retrieves candidate evidence sets from the knowledge base based on these. During answer generation, it enforces the binding of answer fragments to evidence, generating reference maps and answer text, and constructs a machine-verifiable causal evidence graph to provide structured explanations. The system performs consistency verification and counterfactual stability testing on the answers and the causal evidence graph, outputting verification results including confidence scores and risk indicators. By recording complete generation context information, including model version, knowledge snapshot identifier, random seed, reference map, and causal evidence graph identifier, the system ensures complete reproducibility of output results under identical conditions, forming an auditable trust loop, thereby systematically improving the credibility and usability of intelligent teaching assistants in nursing education. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:
[0041] Figure 1 Functional module block diagram of the large-scale intelligent teaching assistant system for traceable, explainable, and verifiable nursing education provided in the embodiments of this application;
[0042] Figure 2 A flowchart illustrating the traceable, interpretable, and verifiable large-scale intelligent teaching assistant method for nursing education provided in this application embodiment. Detailed Implementation
[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.
[0044] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.
[0045] LLM: Large Language Model;
[0046] RAG: Retrieval-Augmented Generation;
[0047] CEG: Causal Evidence Graph;
[0048] DRR: Deterministic Replayable Decoding;
[0049] JSON: JavaScript Object Notation;
[0050] BM25: Best Match 25;
[0051] NLI: Natural Language Inference;
[0052] RoBERTa: Robustly Optimized BERT Pretraining Approach.
[0053] BERT: Bidirectional Encoder Representations from Transformers;
[0054] OCR: Optical Character Recognition;
[0055] ResNet: Residual Network;
[0056] SHA: Secure Hash Algorithm;
[0057] AES: Advanced Encryption Standard;
[0058] CNTSS: Chinese Nursing Terminology Standard System;
[0059] URI: Uniform Resource Identifier;
[0060] ID: Identification;
[0061] SQL: Structured Query Language.
[0062] The purpose of this invention is to provide a traceable, explainable, and verifiable large-scale intelligent teaching assistant system and method for nursing education. This system aims to address the three core problems commonly found in existing technologies: "coarse traceability granularity and lack of verifiability, lack of structured and operable explanation, and inability to replay results".
[0063] Firstly, such as Figure 1 As shown, this application provides a large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education, including an input capture and semantic parsing module 100, a multi-source knowledge retrieval and versioned tracing module 200, an executable citation and segment-by-segment binding module 300, a causal evidence graph and structured interpretation generation module 400, an evidence consistency and counterfactual stability verification module 500, and a replayable decoding and evidence storage module 600. Specifically, each module works together to achieve its function as follows:
[0064] The input capture and semantic parsing module 100 is used to receive and parse the user's multimodal input, generate a standardized question template containing intent and key entities, and classify the clinical or teaching risk level information involved in the question.
[0065] The multi-source knowledge retrieval and version tracing module 200 is communicatively connected to the input capture and semantic parsing module 100. It is used to retrieve knowledge from the version-managed knowledge base according to the standardized question template, generate a source fragment fingerprint containing content hash and version identifier for the retrieved knowledge fragment, and output a candidate evidence set with the source fragment fingerprint attached.
[0066] The executable reference and segment binding module 300 is communicatively connected to the input capture and semantic parsing module 100 and the multi-source knowledge retrieval and versioning tracing module 200. It is used to divide the answer into the smallest semantic unit according to the risk level information, bind each smallest semantic unit to the source segment fingerprint in the candidate evidence set, generate the answer text and reference mapping relationship, and perform risk level-based gating processing on the answer segments without bound evidence.
[0067] The causal evidence graph and structured interpretation generation module 400 is communicatively connected to the executable reference and segment binding module 300. It is used to automatically construct a structured causal evidence graph that represents the causal logic of nursing operations with nodes and edges based on the answer text and reference mapping relationship, and generate the corresponding causal evidence graph identifier.
[0068] The evidence consistency and counterfactual stability verification module 500 is communicatively connected to the executable reference and segment binding module 300 and the causal evidence graph and structured interpretation generation module 400. It is used to perform logical consistency verification and counterfactual stability testing on the answer text and the causal evidence graph, and output verification results including confidence scores and risk indicators.
[0069] The replayable decoding and evidence storage module 600 is communicatively connected to the executable reference and segment-by-segment binding module 300, the causal evidence graph and structured interpretation generation module 400, and the evidence consistency and counterfactual stability verification module 500. It is used to persistently store the replayable decoding information, the answer text, the causal evidence graph, and the verification results, which are composed of model version identifier, knowledge snapshot identifier, random seed, reference mapping relationship and causal evidence graph identifier.
[0070] This application constructs a collaborative technical framework integrating multiple modules to ultimately achieve precise source tracing and machine verifiability at the fragment level of answers, ensuring that each assertion is supported by reliable evidence; it provides structured explanations and automatic verification of the reasoning process, presenting the reasoning logic in the form of clear and reviewable causal evidence diagrams, and supports automatic machine verification of logical consistency; it ensures the replayability and auditability of results under the same conditions, ensuring the deterministic reproducibility of output results by generating key contextual parameters through complete recording. Thus, it systematically improves the usability and credibility of intelligent teaching assistants in high-risk educational scenarios such as nursing.
[0071] The core algorithm of the traceable, explainable, and verifiable large-scale intelligent teaching assistant system for nursing education provided in this application establishes answer fragment identifiers ( ) and knowledge fragment citation ( The system accurately correlates the evidence with the answer fragments, quantifies the support level of the evidence using a support strength score, and employs the following confidence calculation formula optimized based on a large number of nursing scenario samples. This aims to ensure that the confidence assessment of the output results aligns with the safety requirements of high-risk nursing practices:
[0072] ;
[0073] In the formula, This represents the final confidence level. For retrieval rating; For cross encoder consistency; For consistency of natural language reasoning; To ensure counterfactual stability, the weights assigned in this formula are optimized based on training with over 5,000 nursing scenario samples, aiming to ensure that the calculation results are highly consistent with the risk management needs in nursing practice.
[0074] In one embodiment, the multimodal input processed by the input capture and semantic parsing module 100 includes text, speech, and images; the module is able to identify key entities from the input, such as drugs, operations, locations, and times; at the same time, the module also performs intent classification and risk level determination, the defined risk levels include ordinary, important, and high risk, and generates corresponding label vectors accordingly.
[0075] In one embodiment, the exclusive knowledge base constructed by the multi-source knowledge retrieval and versioning traceability module 200 covers nursing guidelines, operating procedures, teaching materials, hospital infection control standards, and pharmacopoeias.
[0076] This module generates a source fragment fingerprint (KnowledgeSpan) for knowledge fragments. Its data structure includes fields such as source identifier (source_id), document path (doc_uri), version number (version_id), hash value (sha256), start and end character positions (start_char, end_char), license tag (license_tag), and update time (updated_at). The sha256 is generated by hashing the knowledge fragment content using the SHA256 algorithm. Combined with the version identifier and character offset, it enables unique identification, version tracing, and tamper detection of knowledge fragments.
[0077] In terms of retrieval strategy, this module adopts a fusion scheme that combines BM25 retrieval, vector retrieval and cross-encoder rearrangement, aiming to output a high-confidence candidate evidence set.
[0078] In one embodiment, the executable reference and segment binding module 300 implements forced binding through pointer network technology, associating the smallest semantic unit in the answer text with one or more source segment fingerprints; the smallest semantic unit includes short sentences or terms.
[0079] The CitationMap generated by this module includes answer segment identifiers, corresponding knowledge segment references, support strength scores, and binding methods. Its structure is defined as: CitationMap={answer_span_id,[KnowledgeSpanRef], support_score, method:"pointer+rerank"}.
[0080] The gating strategy is configured differently based on the risk level: in high-risk scenarios, the output of answer fragments without binding evidence is prohibited; in important scenarios, answer fragments without binding evidence are displayed with warning signs; and in ordinary scenarios, answer fragments without binding evidence are visually downgraded.
[0081] In one embodiment, the causal evidence graph constructed by the structured interpretation generation module 400 is a directed multi-type graph. The node types of this graph include claim nodes, evidence nodes, rule nodes, and constraint nodes; the edge types include supporting edges, rebuttal edges, causal edges, subordinate edges, and time-sensitive edges. Subordinate edges represent hierarchical or inclusion relationships between nodes, and time-sensitive edges represent constraints related to time validity. The causal evidence graph generates its structured description using a controlled syntax based on a nursing domain ontology, and is decoded under the constraints of this ontology, ultimately generating a corresponding graph structure hash value to ensure uniqueness.
[0082] In one embodiment, the evidence consistency and counterfactual stability verification module 500 performs evidence consistency verification and counterfactual stability testing in step S5 in the following ways: The module uses a pre-trained Natural Language Inference (NLI) model (e.g., RoBERTa-base-snli-mean-tokens) for evidence consistency verification. It detects logical conflicts between evidence from different sources by calculating the cosine similarity of sentence embedding vectors. When the similarity is lower than a set threshold (≤0.3), it is determined that there is a contradiction. Simultaneously, the module performs a counterfactual stability test by constructing counterfactual perturbation inputs covering key nursing dimensions such as gender, pregnancy status, age, liver and kidney function, and allergy history. The perturbation inputs are resubmitted to the system to trigger a complete processing flow from input parsing to answer generation, and a logical consistency score between the newly generated result and the original output is calculated. If the score is lower than a preset threshold (<0.7), a mechanism to automatically reduce the confidence level of the result or initiate a manual review process is triggered.
[0083] In one embodiment, the replayable decoding and evidence storage module 600 records at least the large model version identifier, the Merkle root hash value of the knowledge base snapshot, a random seed, complete reference mapping data, and the graph structure hash value of the causal evidence graph. This module stores the above information, along with other output results, in an encrypted database that supports access control, with permissions divided into at least three levels: administrator, teacher, and student. Through this mechanism, the system ensures that, under the same input and knowledge snapshot conditions, the answer content and its reference relationships can be completely reproduced to meet auditing and traceability requirements.
[0084] In one embodiment, the technical parameters of each module of the large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education provided in this application are as follows:
[0085] The input capture and semantic parsing module 100 uses the BERT-based model for entity recognition and intent classification, with an F1 score of no less than 0.92; image input uses ResNet50 combined with OCR technology to identify device labels, with a recognition rate of over 0.93; the risk level determination rules are generated based on training from more than 200 nursing expert labeled samples.
[0086] The multi-source knowledge retrieval and version traceability module 200 has built a knowledge base containing more than 20,000 medical and nursing knowledge points compiled by nursing experts, more than 200 industry guidelines, and a catalog of more than 100 textbooks. It uses Elasticsearch to implement BM25 retrieval, with a response time controlled within 100 milliseconds. The cross-encoder adopts the cross-encoder / ms-marco-MiniLM-L-6-v2 model, and the reordering time does not exceed 200 milliseconds.
[0087] The pointer network hidden layer dimension of the executable reference and fragment binding module 300 is set to 512. The training data contains more than 20,000 pairs of annotated nursing answer fragments and evidence. The minimum semantic unit length is configured to be 1 to 5 short sentences. The gating strategy defaults to the rule of prohibiting output of missing evidence in high-risk scenarios and displaying warnings in ordinary scenarios.
[0088] The CEG ontology of the Causal Evidence Graph and Structured Interpretation Generation Module 400 contains 8 types of nodes and 5 types of edges, the definitions of which are based on the Nursing Terminology System (CNTSS). This module uses the NetworkX library to construct the graph structure and generates hash values from the JSON representation of the graph structure using the SHA-256 algorithm to ensure its uniqueness.
[0089] The NLI model inference speed of the evidence consistency and counterfactual stability verification module 500 does not exceed 50 milliseconds per sentence, and the conflict detection threshold is 0.3, which has been verified by more than 1,000 nursing conflict samples. The counterfactual perturbation test involves more than 20 core dimensions, such as changing the pregnancy status from non-pregnant to late pregnancy, and changing liver and kidney function from normal to failure. The logical consistency threshold is set at 0.7 and has been optimized through testing in more than 500 nursing scenarios.
[0090] The replayable decoding and evidence storage module 600 uses a PostgreSQL database to store DRR data, uses the SHA-256 algorithm to build a Merkle tree, and generates a knowledge snapshot root hash every hour to ensure that the knowledge version cannot be tampered with.
[0091] Secondly, such as Figure 2 As shown, this application provides a traceable, interpretable, and verifiable large-scale intelligent teaching assistant method for nursing education, including the following steps:
[0092] Step S1: Parse the user's multimodal input, identify entities and intents, determine the risk level, and generate standardized question templates;
[0093] Step S2: Based on the standardized question template, retrieve the candidate evidence set from the version-managed nursing knowledge base using a fusion retrieval strategy;
[0094] Step S3: Using pointer network technology, each smallest semantic unit is simultaneously bound to the fingerprint of a knowledge fragment in the candidate evidence set during the answer generation process, generating a reference mapping that records the supporting relationship, thereby determining the binding evidence that supports the answer;
[0095] Step S4: Based on the relationship between the answer and the binding evidence, construct a structured causal evidence graph;
[0096] Step S5: Perform evidence consistency verification and counterfactual stability test on the answer and causal evidence graph, generate verification results including confidence score and risk label, and record deterministic replayable decoding information including model version, knowledge snapshot root hash, random seed, reference mapping and causal evidence graph hash and store it in an encrypted database to ensure that the results are replayable;
[0097] Step S6: Output the final answer that integrates and binds evidence, structured causal explanations, and multi-dimensional verification results, and simultaneously provide users with fragment-level citations, causal evidence diagram visualizations, and verification result explanations.
[0098] In one specific embodiment, the nursing education traceable, explainable, and verifiable large-scale intelligent teaching assistant system provided in this application, specifically applied to a text-input-based traceable nursing question-and-answer system, has the following workflow:
[0099] When a user raises the question of how to verify patient information before intravenous infusion, the system analyzes that the task belongs to pre-operation verification, identifies key entities including intravenous infusion, patient identity, medication and expiration date, and determines the risk level to be important. Then, it generates a standardized template of key questions for verifying patient information before intravenous infusion.
[0100] The system retrieved and recalled relevant evidence fragments from the knowledge base, namely the guideline with ID 00123 (version v3) and the hospital infection control guidelines with ID 00987 (version v2). After reordering, it obtained five candidate evidence sets with confidence levels of 0.92, 0.90, 0.88, 0.86, and 0.85, respectively.
[0101] During the answer generation process, the system bound the answer segment of verifying the patient's name and identification to the content in characters 112 to 156 of the guideline, with a support score of 0.92; bound the segment of verifying the medical order content and the name and concentration of the liquid to the content in characters 45 to 120 of the hospital infection control guidelines, with a support score of 0.90; and bound the segment of verifying the expiration date and appearance of the medication to the content in characters 200 to 260 of the guideline, with a support score of 0.88. As for the extended statement that it is recommended to communicate the purpose of the operation to the patient, since no corresponding evidence was found to support it, the system marked it with a warning according to the important scenario rules, indicating that there is currently no clear evidence to support this content.
[0102] During the causal evidence graph construction phase, the system uses the three types of verification points before infusion as claim nodes, sets evidence edges pointing to the three bound evidence fragments mentioned above, and associates the medical order consistency rule node and the validity period constraint node within the current time window through constraint edges, and generates a unique graph structure hash value a7f3bc9d1e2f4g5h6i7j8k9l0m;
[0103] After natural language reasoning consistency verification, the system achieved a high score of 0.95, indicating no logical conflicts. In a counterfactual perturbation test, changing the scenario to chemotherapy drug infusion, the logical consistency score was 0.82, exceeding the threshold requirement of 0.7. Therefore, no manual review process was required. The system fully recorded the deterministic replayable decoding information of this interaction, including the model version LLM-finetuned-v1, the knowledge snapshot Merkle root hash value d2e4f1a3b5c7d9e0f2g4, the random seed 123456, complete reference mapping data, and the graph structure hash value.
[0104] The final output presents the user with a complete result including the answer text, fragment-level citations, a causal evidence diagram visualization interface, and a consistent and counterfactually stable verification explanation.
[0105] In one specific embodiment, the nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant system provided in this application, specifically applied to the pre-catheterization assessment scenario, has the following workflow:
[0106] When a user asks what checks are needed before catheterization, the system analyzes and determines that the task is a pre-operation assessment, identifies key entities including urinary tract infection risk, allergy history, aseptic preparation, bladder filling and informed consent, and determines the risk level to be high risk.
[0107] The system retrieved and recalled relevant content from the knowledge base for nursing guidelines (ID 124, version v2) and textbooks (ID 201, version v5), and the confidence level of the obtained candidate evidence fragments was no less than 0.88.
[0108] In the answer generation and evidence binding phase, the segment assessing the risk and contraindications of urinary tract infection, such as urethral stricture, was bound to the content in characters 310 to 372 of the nursing guidelines, with a support score of 0.93. The segment confirming medical orders and indications, and verifying patient identity through double identification, was bound to the content in characters 45 to 100 of the textbook, with a support score of 0.91. The segment assessing bladder fullness through percussion or ultrasound examination, pain scoring, and psychological preparation was bound to the content in characters 120 to 210 of the textbook, with a support score of 0.89. The segment preparing sterile instruments, including catheters and lubricants, and performing hand hygiene was bound to the content in characters 400 to 468 of the nursing guidelines, with a support score of 0.90. Due to the high-risk scenario, the system strictly enforced the rule of prohibiting output without evidence, ensuring that there were no segments in the answer that were not bound to evidence.
[0109] When constructing the causal evidence graph, the system sets rule nodes containing key points of aseptic operation and time constraint nodes containing the requirement that the shelf life of disposable medical devices shall not exceed 24 months. The weight of the evidence edge is calculated by weighting the cross encoder score by 0.9 and the expert rule by 0.1.
[0110] The system performs natural language reasoning consistency verification, cross-source conflict detection, and counterfactual stability testing on the answers and causal evidence graphs. If the logical consistency score is below 0.7, manual review is triggered. During the verification process, the system simulates and generates content containing statements about diagnostic catheterization not requiring informed consent. Since this content lacks supporting evidence, it is automatically intercepted, and a high-risk scenario warning is issued requiring mandatory evidence binding. Subsequently, a counterfactual perturbation test is performed. After setting the patient as a woman in late pregnancy, the logical consistency score of the system-generated answer is 0.85, exceeding the set threshold. Afterward, all deterministic replayable decoding records of interactions are synchronously stored. Teachers can subsequently query and retrieve the complete reference mapping and causal evidence graph by combining the assessment scenario ID and timestamp to analyze students' knowledge mastery.
[0111] Finally, the system outputs to the user the answer text without warning labels, fragment-level citations that can be jumped to the original text, a visualization of the causal evidence diagram highlighting the causal relationship between aseptic preparation and contraindication assessment, and a verification description that is without missing evidence, stable in counterfactual terms, and meets the requirements of high-risk scenarios.
[0112] In summary, the traceable and explainable intelligent teaching assistant system for nursing education provided in this application can achieve the following significant and quantifiable technical effects:
[0113] The system achieves sentence-by-sentence source tracing at the answer fragment level, with a source tracing accuracy of no less than 99%. At the same time, with the help of deterministic replayable decoding records, the result reproducibility rate reaches 100% under the same input and knowledge snapshot conditions, which fully meets the strict auditing requirements of medical education.
[0114] Structured causal evidence diagrams enable machine-automated verification of reasoning logic, improving verification efficiency by 80% compared to traditional free text interpretation. Instructors can use this diagram to directly locate errors such as evidence conflicts or missing rules in reasoning, effectively reducing the cost of troubleshooting.
[0115] By using a mandatory evidence binding mechanism to prevent the generation of unfounded conclusions, and combining it with a multi-dimensional verification strategy, the error output rate in nursing scenarios is reduced by more than 95%, significantly improving the applicability of the system in high-risk educational scenarios.
[0116] The fusion retrieval strategy improves inference speed by 2 to 3 times compared to traditional retrieval-enhanced generation schemes, while supporting multimodal input and flexible deployment methods to ensure that the system can meet the diverse equipment needs in the field of nursing education.
[0117] Thirdly, this application provides a traceable, explainable, and verifiable large-scale intelligent teaching assistant device for nursing education, on which the above-mentioned traceable, explainable, and verifiable large-scale intelligent teaching assistant system for nursing education runs. The traceable, explainable, and verifiable large-scale intelligent teaching assistant device for nursing education can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.
[0118] In this embodiment of the application, the large-scale intelligent teaching assistant device for traceable, explainable and verifiable nursing education may include a processor, a memory, a communication interface and a communication bus.
[0119] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.
[0120] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces enable interconnection of devices within the large-scale intelligent teaching aid for traceable, interpretable, and verifiable nursing education, as well as interfaces for interconnection between the large-scale intelligent teaching aid and other devices (such as other computing devices or user devices). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user devices can be displays, keyboards, etc.
[0121] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.
[0122] The processor can be a general-purpose processor, which can call the nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant program stored in memory and execute the nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant program is called can be referred to the various embodiments of the nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant method of this application, and will not be repeated here.
[0123] Fourthly, embodiments of this application also provide a readable storage medium.
[0124] This application stores a traceable, interpretable, and verifiable large-scale intelligent teaching assistant program for nursing education on a readable storage medium, wherein when the traceable, interpretable, and verifiable large-scale intelligent teaching assistant program for nursing education is executed by a processor, it implements the steps of the traceable, interpretable, and verifiable large-scale intelligent teaching assistant method for nursing education as described above.
[0125] The method implemented when the large-scale intelligent teaching assistant program for traceable, interpretable, and verifiable nursing education is executed can be referred to in the various embodiments of the large-scale intelligent teaching assistant method for traceable, interpretable, and verifiable nursing education in this application, and will not be repeated here.
[0126] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A large-scale intelligent teaching assistant system for traceable, explainable, and verifiable nursing education, characterized in that, include: The input capture and semantic parsing module is used to receive and parse the user's multimodal input, generate standardized question templates, and obtain the risk level information of the questions; The multi-source knowledge retrieval and version tracing module is connected to the input capture and semantic parsing module. It is used to perform retrieval based on the standardized question template and output a set of candidate evidence with source fragment fingerprints for the retrieved knowledge fragments. An executable reference and segment binding module is connected to the input capture and semantic parsing module and the multi-source knowledge retrieval and versioning tracing module. It is used to divide the answer into the smallest semantic unit according to the risk level information, and bind each smallest semantic unit to the source segment fingerprint in the candidate evidence set to generate the answer text and reference mapping relationship. The causal evidence graph and structured interpretation generation module is connected to the executable reference and segment binding module, and is used to construct a causal evidence graph and generate a corresponding causal evidence graph identifier based on the answer text and reference mapping relationship; The evidence consistency and counterfactual stability verification module is connected to the executable reference and segment binding module and the causal evidence graph and structured interpretation generation module. It is used to perform logical consistency verification and counterfactual stability testing on the answer text and the causal evidence graph, and output the verification results. The replayable decoding and evidence storage module is connected to the executable reference and segment-by-segment binding module, the causal evidence graph and structured interpretation generation module, and the evidence consistency and counterfactual stability verification module. It is used to store the replayable decoding information, the answer text, the causal evidence graph, and the verification result. The replayable decoding information includes at least the reference mapping relationship and the causal evidence graph identifier.
2. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described in claim 1, characterized in that, The forced binding in the executable reference and fragment binding module is specifically as follows: By using pointer network technology, the smallest semantic unit in the answer text is associated with one or more source fragment fingerprints, and the support strength score of each association is recorded in the reference map; Gating processing is configured differently based on the risk level information received from the input capture and semantic parsing module: for high-risk scenarios, outputting answer fragments without binding evidence is prohibited; for important scenarios, answer fragments without binding evidence are displayed with warning labels; for ordinary scenarios, answer fragments without binding evidence are displayed with visual degradation.
3. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described in claim 1, characterized in that, In the multi-source knowledge retrieval and version tracing module, the data structure of the source fragment fingerprint includes document identifier, version number, content hash value, and character offset.
4. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described in claim 1, characterized in that, In the causal evidence graph and structured interpretation generation module, the node types of the causal evidence graph include at least claim nodes, evidence nodes, rule nodes, and constraint nodes, and the edge types include at least support edges, rebuttal edges, causal edges, subordinate edges, and time-limited edges. The structured description of the causal evidence graph is generated through a controlled syntax based on the nursing domain ontology, and a corresponding graph structure hash value is generated.
5. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described in claim 1, characterized in that, The counterfactual stability test in the evidence consistency and counterfactual stability verification module specifically includes: Construct counterfactual perturbation inputs covering at least one of the key nursing dimensions: gender, pregnancy status, age, liver and kidney function, and allergy history; Resubmit the counterfactual perturbation input to the input capture and semantic parsing module, and trigger chained processing in subsequent modules; Calculate the logical consistency score between the original output and the perturbed output. If the score is lower than the preset threshold, automatically trigger the reduction of the result confidence level or start the manual review process.
6. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education according to claim 1, characterized in that, The replayable decoding information recorded in the replayable decoding and evidence storage module includes at least the large model version identifier used, the Merklegen hash value of the knowledge base snapshot, the random seed that controls the randomness of generation, the complete reference mapping data, and the graph structure hash value of the causal evidence graph.
7. The large-scale intelligent teaching assistant system for traceable, interpretable, and verifiable nursing education as described in claim 1, characterized in that, The evidence consistency verification module in the evidence consistency and counterfactual stability verification module is implemented using a natural language reasoning model. It detects logical conflicts between evidence from different sources by calculating the cosine similarity of sentence embedding vectors. When the similarity is lower than a set threshold, it is determined that there is a contradiction.
8. The nursing education traceable, interpretable, and verifiable large-scale intelligent teaching assistant system according to any one of claims 1 to 7, characterized in that, The system is applicable to nursing education, skills assessment, patient safety training, pharmaceutical education, and emergency care training scenarios.
9. A large-scale intelligent teaching assistant method for traceable, interpretable, and verifiable nursing education, characterized in that, The application of the nursing education traceable, explainable, and verifiable large-scale intelligent teaching assistant system as described in any one of claims 1-8 includes the following steps: The system parses user multimodal input, identifies entities and intents, determines risk levels, and generates standardized question templates. Based on the standardized question template, a set of candidate evidence is retrieved from the version-managed nursing knowledge base using a fusion retrieval strategy; By using pointer network technology, each smallest semantic unit is simultaneously bound to the fingerprint of a knowledge fragment in the candidate evidence set during the answer generation process, generating a reference mapping that records the supporting relationship, thereby determining the binding evidence that supports the answer; Based on the relationship between the answer and the binding evidence, a structured causal evidence graph is constructed; Perform evidence consistency verification and counterfactual stability testing on the answers and causal evidence diagrams, generate verification results including confidence scores and risk indicators, and record deterministic replayable decoding information; It outputs the final answer by integrating and binding evidence, structured causal explanations, and multi-dimensional verification results, and simultaneously provides users with fragment-level citations, causal evidence diagram visualizations, and verification result explanations.
10. The large-scale intelligent teaching assistant method for traceable, interpretable, and verifiable nursing education according to claim 9, characterized in that, Multimodal input includes text, speech, and images; The smallest semantic unit is bound through a pointer network; Evidence consistency verification is performed using a pre-trained natural language reasoning model.