Emergency nurse intelligent examination and training method and system based on multi-dimensional growth path

Abstract

The invention relates to the field of intelligent emergency nursing education, and provides an emergency nurse intelligent examination and training method and system based on a multi-dimensional growth path, and the method comprises the steps: firstly constructing a main knowledge subgraph which covers the core knowledge of an emergency process, critical disease monitoring, emergency drugs and the like, and the relation of the core knowledge; and auxiliary knowledge such as psychological intervention and doctor-patient communication is extracted to form independent cross-domain sub-graphs, and cross-graph weak association is established through semantic alignment. According to the nurse ability portrait, a learning path is planned in the main knowledge graph, cross-domain knowledge is introduced as required, dual standards of semantic relevancy and information entropy are utilized to realize redundancy filtering and dynamic fusion of important knowledge, and an enhanced node is generated through knowledge distillation to be embedded into the main graph. And then a graph neural network and a path planning algorithm are adopted to output a personalized growth test path considering both core and cross-domain knowledge.

Description

Technical Field

[0001] This invention relates to the fields of intelligent medical education, knowledge graphs and semantic recommendation technology, and in particular to an intelligent examination and training method and system for emergency nurses based on a multi-dimensional growth path. Background Technology

[0002] With the rapid development of the healthcare industry, the knowledge upgrading and multi-dimensional competence development of emergency nurses are becoming increasingly important. Existing intelligent training and assessment systems for emergency nurses mostly employ knowledge-point-linked course recommendations or simple domain-expansion path generation schemes. These systems typically construct a knowledge framework centered on emergency procedures, critical care, and emergency medication, supplemented by targeted training modules and assessment mechanisms, striving to improve nurses' comprehensive professional competence in multiple aspects, including professional skills, psychological intervention, and doctor-patient communication abilities.

[0003] The existing technology has the following main problems: First, most current systems are based on splicing or aggregating courses, making it difficult to accurately distinguish the actual relevance and value of auxiliary domain knowledge and the main learning content of emergency nursing. This easily introduces a large number of low-density, low-relevance information points, increasing the learning burden on nurses.

[0004] Second, there is a lack of systematic evaluation standards for information density and synergistic effects.

[0005] Third, the ability to express cross-domain content is limited. Existing splicing or aggregation recommendation methods cannot perform semantic compression and secondary abstraction of cross-domain knowledge points, making it difficult to grasp the core capability transfer elements. This results in lengthy expressions of integrated content that are difficult to accurately match the needs of the main emergency nursing system. Summary of the Invention

[0006] This application provides an intelligent examination and training method for emergency nurses based on a multi-dimensional growth path, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] This application provides an intelligent assessment and training method for emergency nurses based on a multi-dimensional growth path, specifically including: S1: Based on the emergency nursing professional knowledge base, a core knowledge subgraph is constructed. The core knowledge subgraph includes knowledge points on emergency procedures, critical care, and emergency drug use, as well as their logical relationships, serving as the main knowledge system for personalized path planning.

[0008] S2: Extract auxiliary knowledge points from fields related to psychological intervention, doctor-patient communication, and disaster medicine, construct an independent cross-domain auxiliary knowledge subgraph, and establish a weak association mapping between it and the core knowledge subgraph to represent potential knowledge synergy relationships.

[0009] S3: Based on the nurse's current competency profile, determine the main learning path in the core knowledge subgraph and identify the target nodes that need to be introduced into cross-domain knowledge as triggering conditions for cross-domain knowledge fusion.

[0010] S4: Perform text semantic analysis on candidate cross-domain knowledge points, calculate their semantic relevance score with the current main path target node, and combine the information entropy value of the knowledge point in the original domain to comprehensively evaluate the information density per unit content.

[0011] S5: Based on the information density assessment results, filter and mark cross-domain knowledge points with high redundancy, retain high-density knowledge points, and output the set of knowledge points to be distilled.

[0012] S6: Input the knowledge points to be distilled into the knowledge distillation engine, extract their core semantics into lightweight semantic vectors, and embed them into the corresponding positions in the core knowledge subgraph to form enhanced knowledge nodes.

[0013] S7: Based on the enhanced core knowledge subgraph, execute the path planning algorithm to generate personalized growth path recommendation results containing necessary cross-domain knowledge.

[0014] S8: After nurses complete the learning, collect the application cases they submit, and compare the degree of cross-domain knowledge application reflected in the cases using natural language processing technology to generate feedback signals.

[0015] S9: Based on the feedback signal, dynamically adjust the knowledge distillation model parameters and semantic density evaluation threshold to continuously optimize the quality of subsequent recommendations.

[0016] It also provides an intelligent examination and training method and system for emergency nurses based on a multi-dimensional growth path, and uses the above-mentioned intelligent examination and training method for emergency nurses based on a multi-dimensional growth path to examine and train nurses' growth.

[0017] The intelligent training and assessment system for emergency nurses based on a multi-dimensional growth path includes training and assessment modules. For nurses at different levels, the system precisely stratifies them based on their seniority, professional title, past assessment results, and core competencies demonstrated in actual clinical scenarios. For example, it divides nurses into several tiers with clear competency standards, such as novice nurses, competent backbone nurses, senior experts, and field authorities. Simultaneously, the system fully considers the different career development aspirations of individual nurses and the talent structure needs of departments, planning differentiated professional development paths for them, such as clinical emergency care expert paths, nursing management and teaching paths, and specialized disease-specific nursing research paths. Different training and assessment plans are specified for nurses in each tier and for each development path. The training content and corresponding requirements also differ, specifically in the depth, breadth, focus, and teaching methods of the training content, exhibiting systematic differences.

[0018] For novice nurses, the focus is on basic theory, standardized procedures, and proficiency in individual skills. For experienced nurses, the emphasis is on comprehensive judgment of complex cases, teamwork in critical care, and optimization of resuscitation procedures. Building upon this foundation, nurses choosing the clinical emergency specialist path will receive in-depth training in advanced life support, management of rare emergencies, and innovation in emergency techniques. For nurses focusing on nursing management and teaching, core modules such as team management, curriculum design, teaching methods, and quality improvement will be systematically integrated. For nurses dedicated to specialized disease research, in-depth content on research methodology, data analysis, critical literature review, and cutting-edge advancements in the specialty will be provided. For expert nurses, the focus is on developing their skills in handling difficult cases, clinical decision support, updating cutting-edge knowledge, teaching guidance, and departmental management. Therefore, different approaches are used to cultivate their growth through personalized training content within in-service education. This personalization is highly data-driven. The system continuously collects and analyzes nurses' operational data, learning behavior data, and multi-dimensional assessment data during simulation training and daily work, dynamically updating their individual competency profiles. Therefore, the training and assessment content may differ based on their varying performance, abilities, and professional development orientations. Even among core nurses, the assessment of those on the management route will focus on their organizational, coordination, and project implementation abilities, while the assessment of those on the emergency specialist route will focus more on their technical decision-making limits and responsiveness in highly simulated rescue situations.

[0019] The intelligent training and assessment system for emergency nurses based on a multi-dimensional growth path includes training management and examination management modules. The training management module manages the entire process from planning to implementation, including the release, execution tracking, progress monitoring, and feedback of training plans. The examination management module supports various forms of test paper generation, test execution, automatic grading, and score analysis, such as online theoretical exams and simulated skills assessments. It also features a courseware library, a vast and continuously updated digital resource center covering a complete set of teaching materials from basic nursing knowledge to cutting-edge emergency care techniques, including but not limited to text textbooks, video demonstrations, interactive simulated cases, virtual reality operation scenarios, and interpretations of the latest clinical guidelines. Through data analysis, the courseware library can generate training courseware and assessment questions tailored to different nurses. Specifically, the system performs deep semantic understanding and intelligent tagging of the massive resources in the courseware library. When the system determines that a nurse has weaknesses in a specific knowledge area or skill node, or when her career development path requires support from a specific knowledge system, it can automatically and accurately extract, combine, and even dynamically generate reinforcement training courseware and special assessment questions for the weak points and development direction from the courseware library. This achieves truly personalized and precise empowerment, greatly improving the efficiency and effectiveness of training, and building a data-driven, individual growth-centered intelligent education closed-loop ecosystem.

[0020] The intelligent assessment and training method and system for emergency nurses based on a multi-dimensional growth path provided in this application have the following beneficial effects: (1) Significantly improve the relevance of the recommended path, effectively avoid information redundancy and learning burden caused by non-core cross-domain content, and ensure that the learning content for emergency nurses is highly relevant and practical. Greatly improve the knowledge expression accuracy and information density of the personalized growth path, making the recommended content more compact and focused on key aspects of ability improvement.

[0021] (2) It achieves precise screening and organic integration of cross-domain knowledge, avoiding the chaotic content and knowledge transfer obstacles caused by traditional splicing recommendation. It dynamically filters low-density and low-relevance knowledge points, realizes proactive control over irrelevant content in the learning path, and enhances the intelligence level of path recommendation. The synergistic application of information density and knowledge distillation effectively compresses cross-domain knowledge, reduces ineffective learning time, and achieves a significant improvement in learning efficiency.

[0022] (3) The system possesses adaptive optimization capabilities, continuously adjusting model parameters and screening thresholds based on actual learning and case application effects to ensure the long-term evolution of the recommendation algorithm and its continuous adaptation to the needs of nurses' skill development. It expands the applicability of the growth path recommendation system, supporting not only emergency nursing but also smoothly introducing auxiliary knowledge from other medical or non-medical fields, greatly enhancing the system's openness and scalability. It effectively assists in the multidimensional improvement of nurses' comprehensive abilities, accelerating the synchronous development of professional skills and soft power, and enhancing rapid response capabilities in diverse scenarios such as actual treatment and communication collaboration. It significantly reduces resource waste and repetitive training, substantially lowering training costs and time investment for hospitals and training institutions. It comprehensively optimizes the recommendation experience, enhancing nurses' learning enthusiasm and path completion rate, providing solid data-driven support for the construction of an intelligent talent training system in the emergency medical industry. Attached Figure Description

[0023] Figure 1 This is the main flowchart of an intelligent examination and training method for emergency nurses based on a multi-dimensional growth path.

[0024] Figure 2 This is a sub-flowchart of an intelligent examination and training method for emergency nurses based on a multi-dimensional growth path.

[0025] Figure 3 This is another sub-flowchart of an intelligent examination and training method for emergency nurses based on a multi-dimensional growth path. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0028] like Figure 1 As shown, this application provides an intelligent assessment and training method for emergency nurses based on a multi-dimensional growth path, specifically including: S1: Based on the emergency nursing professional knowledge base, a core knowledge subgraph is constructed. The core knowledge subgraph includes knowledge points on emergency procedures, critical care, and emergency drug use, as well as their logical relationships, serving as the main knowledge system for personalized path planning.

[0029] S2: Extract auxiliary knowledge points from fields related to psychological intervention, doctor-patient communication, and disaster medicine, construct an independent cross-domain auxiliary knowledge subgraph, and establish a weak association mapping between it and the core knowledge subgraph to represent potential knowledge synergy relationships.

[0030] S3: Based on the nurse's current competency profile, determine the main learning path in the core knowledge subgraph and identify the target nodes that need to be introduced into cross-domain knowledge as triggering conditions for cross-domain knowledge fusion.

[0031] S4: Perform text semantic analysis on candidate cross-domain knowledge points, calculate their semantic relevance score with the current main path target node, and combine the information entropy value of the knowledge point in the original domain to comprehensively evaluate the information density per unit content.

[0032] S5: Based on the information density assessment results, filter and mark cross-domain knowledge points with high redundancy, retain high-density knowledge points, and output the set of knowledge points to be distilled.

[0033] S6: Input the knowledge points to be distilled into the knowledge distillation engine, extract their core semantics into lightweight semantic vectors, and embed them into the corresponding positions in the core knowledge subgraph to form enhanced knowledge nodes.

[0034] S7: Based on the enhanced core knowledge subgraph, execute the path planning algorithm to generate personalized growth path recommendation results containing necessary cross-domain knowledge.

[0035] S8: After nurses complete the learning, collect the application cases they submit, and compare the degree of cross-domain knowledge application reflected in the cases using natural language processing technology to generate feedback signals.

[0036] S9: Based on the feedback signal, dynamically adjust the knowledge distillation model parameters and semantic density evaluation threshold to continuously optimize the quality of subsequent recommendations.

[0037] Step S1: Based on the emergency nursing professional knowledge base, construct a core knowledge subgraph. This core knowledge subgraph includes knowledge points on emergency procedures, critical care, and emergency medication use, along with their logical relationships, serving as the main knowledge system for personalized path planning. Specifically, it includes: S1.1: Obtain structured knowledge points related to emergency procedures, critical care, and emergency drug use from the emergency nursing professional knowledge base, and use them as the original input data source for the construction of the core knowledge graph.

[0038] The basic theoretical component of the emergency nursing professional knowledge base relies on classic textbooks such as *Emergency Medicine* and *Critical Care Nursing*. These textbooks systematically explain the pathophysiological mechanisms, typical clinical manifestations, key points of differential diagnosis, and comprehensive nursing assessment and monitoring requirements for various acute and critical illnesses, ranging from cardiovascular and cerebrovascular accidents and severe trauma to acute poisoning. This forms the knowledge foundation for all clinical judgments. At the core operational level, internationally recognized guidelines such as *Basic Life Support*, *Advanced Cardiac Life Support*, and *Trauma Life Support* provide the gold standard for cardiopulmonary resuscitation, defibrillation, airway management, and initial trauma assessment and management. Furthermore, the internal *Emergency Nursing Operation Procedures* developed by various medical institutions based on these international guidelines further refine the steps, contraindications, and aseptic principles for each specific technique, such as arterial and venous puncture, gastric lavage, urinary catheterization, and hemostasis and bandaging, ensuring uniformity and safety in operations. Regarding medications and equipment, rapid reference tools such as the "Emergency Department Commonly Used Medications Handbook" or the drug database in the hospital information system clearly define the precise dosage, preparation method, infusion rate, and strategies for handling potential adverse reactions for each emergency medication. Simultaneously, every critical life support device, such as a ventilator, defibrillator, or gastric lavage machine, must be equipped with its official "Operating Manual" and follow strict "Medical Equipment Maintenance and Care Procedures" to ensure nurses can use it correctly and safely. For specific clinical pathways, documents such as the "Stroke Center Construction Guidelines," "Chest Pain Center Diagnosis and Treatment Standards," and "Severe Trauma Treatment Procedures" specify time windows and standardized collaborative processes for each stage, from patient admission and imaging examinations to thrombolysis or interventional therapy, greatly optimizing treatment efficiency. Furthermore, this knowledge base must include the "Hospital Infection Prevention and Control Guidelines," which details key aspects such as hand hygiene, isolation measures, and medical waste disposal to ensure patient and healthcare worker safety. Finally, laws and regulations such as the "Regulations on the Handling of Medical Accidents," the "Regulations on Nurses," and the "Basic Standards for Medical Record Writing" form the baseline and framework for the entire practice, clarifying nurses' legal responsibilities, patients' rights, and the legal requirements for medical documentation. All these texts together form a robust knowledge network, supporting emergency nursing work throughout the entire process, from rapid triage and emergency resuscitation to subsequent treatment and legal compliance.

[0039] Structured knowledge points related to emergency procedures, critical care, and emergency drug use are retrieved from the emergency nursing professional knowledge base and used as the original input data source for constructing the core knowledge graph.

[0040] A medical standard terminology mapping algorithm (parameter settings based on the Chinese Emergency Nursing Discipline Standard Code Table) is used to achieve unique code matching for emergency procedure entries in the original knowledge base and generate a standardized knowledge point list.

[0041] Furthermore, by using a data pattern parsing algorithm (parameter settings: field parsing template includes knowledge point name, definition, applicable scenario, and operation steps), the fields of critical care-related items are decomposed, and a structured attribute dataset is obtained.

[0042] Furthermore, through a multi-source data fusion algorithm (parameter setting: fusion weight based on data source quality score), the association mapping of drug usage, dosage, contraindications and other information from the emergency drug database with entries in the emergency nursing knowledge base is realized, and a unified drug knowledge point structure across databases is generated.

[0043] Furthermore, through a data consistency verification algorithm (parameter setting: consistency threshold is 0.95), cross-consistency verification is achieved between knowledge points related to emergency procedures, critical care, and emergency drug use, and a set of knowledge points that have passed the verification is generated.

[0044] Through the multi-stage data extraction and fusion processing described above, the original knowledge base information from the previous step is transformed into a structured, standardized, and consistency-verified set of knowledge points, thereby enabling the construction of high-quality original input data for the core knowledge graph.

[0045] For example, in an emergency nursing knowledge system, the dictionary size of the medical standard terminology mapping algorithm is set to 5000 entries. During encoding matching, a word order priority rule within the field is used to uniquely encode and convert emergency procedure items, resulting in 300 standardized procedure nodes. In critical care data pattern parsing, the field parsing template includes six categories of fields: name, definition, applicable scenarios, implementation conditions, operational details, and precautions. After parsing, a complete attribute dataset of 150 monitoring nodes is obtained. In the multi-source fusion process of emergency drugs, the quality score of the drug database is 0.92, and the fusion weight is set at 0.6 for the core library and 0.4 for the drug library, forming a structure of 50 drug knowledge points containing dosage, route of administration, and contraindications information. In the consistency verification stage, cross-validation is performed on the fused knowledge point set, and a consistency index is calculated.

[0046] The numerator represents the number of knowledge points that passed the consistency verification, the denominator represents the total number of verified knowledge points, and the consistency value is compared to a threshold of 0.95. 285 core knowledge points of emergency nursing that passed the verification are output as high-quality input for the subsequent construction of knowledge graph nodes, which achieves the technical effect of significantly improving the quality of the original data input.

[0047] S1.2: Perform semantic parsing and entity extraction on the structured knowledge point data to identify key knowledge point entities and their attribute information, so as to form a knowledge graph node set with semantic description capabilities.

[0048] Based on the structured knowledge point data related to emergency procedures, critical care, and emergency drug use obtained in step S1.1, a domain-adaptive word segmentation algorithm (parameters: including an emergency nursing professional terminology dictionary and a general medical lexicon) is used to achieve accurate word segmentation and terminology segmentation of the knowledge point description text data, so as to reduce noise interference in the subsequent semantic parsing process.

[0049] Furthermore, by using a named entity recognition algorithm (model parameters: BiLSTM-CRF structure, training corpus: Emergency Nursing Standard Manual), the entity types in the word segmentation results are identified, including disease entities, operation step entities, drug entities, and equipment entities, and a dataset with entity labels is output.

[0050] Furthermore, an attribute extraction algorithm (parameter: attribute parsing rule set based on dependency parsing) is used to extract relevant attribute information of the identified entities, including execution conditions, dosage parameters, monitoring indicators and time requirements, and form a set of entity-attribute key-value pairs.

[0051] Furthermore, through a semantic type mapping algorithm (parameter: concept matching rules based on the medical ontology), synonymous or near-synonymous entities from different sources are merged into a unified standardized semantic category to ensure the semantic consistency of the knowledge graph node set.

[0052] Furthermore, an entity unique identifier generation algorithm (parameter: Snowflake distributed ID generation strategy) is adopted to encode the globally unique identifier of the standardized entity nodes, and the attributes, categories and identifiers are assembled into a set of knowledge graph node objects with complete semantic descriptions.

[0053] Through the semantic parsing and entity extraction processing methods described above, the structured knowledge point data obtained in the previous step is transformed into a set of node semantic expressions, thereby achieving the completeness of node semantics and the effect of structured expression in the core knowledge subgraph.

[0054] For example, an emergency nursing knowledge base contains a structured description of the knowledge point "Cardiopulmonary Resuscitation - Chest Compressions - Compression Rate," with the text content being "Perform chest compressions in cases of cardiac arrest, maintaining a rate of 100 to 120 compressions per minute." The word segmentation process uses an emergency nursing terminology dictionary to divide the text into segments such as [cardiac arrest], [chest compressions], [rate], [maintain], and [100 to 120 compressions per minute]. The named entity recognition stage marks [cardiac arrest] as a disease entity, [chest compressions] as an operational step entity, and [100 to 120 compressions per minute] as a monitoring indicator attribute value. The attribute extraction stage identifies the attribute key "rate" and the attribute value range 100 to 120 compressions per minute. A semantic type mapping algorithm matches "chest compressions" with the corresponding standard concept of "CPR chest compressions" in the ontology database, merging them into standardized entities. The unique identifier generation stage generates an ID of 145893215478 for this node, ultimately forming a node object {ID:145893215478, Category: Operation Step, Attribute:{Frequency:100-120 times / min}}, which is then added to the node set. This process significantly improves the semantic clarity and attribute completeness of the node, providing accurate input for relation extraction in S1.3.

[0055] S1.3: Based on natural language processing technology, the relationship is extracted from the text describing knowledge points, and semantic relationships such as hierarchical, causal, and process order between knowledge points are identified in order to construct a set of directed edges in the knowledge graph.

[0056] Based on the set of knowledge graph nodes with semantic description capabilities output from the previous sub-step, a dual-channel natural language processing relation extraction algorithm (parameters: channel one is used for dependency parsing, and channel two is used for semantic role labeling) is adopted to realize multi-dimensional semantic relation recognition of knowledge point description text in the field of emergency nursing.

[0057] Furthermore, by using dependency parsing (parameters: Stanford Parser model, dependency arc score threshold set to 0.75), the backbone syntactic structure between entities in the text is parsed, and a list of basic modification, subject-verb, and object dependencies between entities is obtained.

[0058] Furthermore, through a semantic role labeling method (parameters: Bi-LSTM+CRF structure, semantic labeling system including hierarchical relationship, causal relationship, and process sequence relationship), the semantic function recognition of entities in sentences is realized, and a mapping table between entity pairs and role types is generated.

[0059] Furthermore, a relation fusion algorithm (parameters: based on rule priority: causality > hierarchical > process order) is used to cross-validate dependency syntax relations and semantic role relations, and obtain a set of validated relation candidates.

[0060] Furthermore, the credibility score w of the candidate relation is calculated using the relation weight calculation formula:

[0061] in, For the confidence score of syntactic analysis, Confidence score for semantic role labeling and These are the weighting coefficients for the two types of scores.

[0062] By using a relation weight filtering method, data below a preset weight threshold is removed, thereby outputting a set of highly reliable semantic relations.

[0063] By using a directed edge generation algorithm, the aforementioned set of relations is transformed into a set of directed edges in the core knowledge graph, ensuring that each edge is accompanied by a relation type label and a weight value, thereby realizing computable semantic associations between knowledge nodes.

[0064] For example, in a scenario involving the extraction of relationships between emergency medical procedures and critical care knowledge points, the input set of knowledge nodes includes semantic descriptions such as "establishing intravenous access," "ECG monitoring," and "drug administration." The dependency parsing channel parses the sentence: "When a patient experiences arrhythmia, intravenous access should be established immediately and ECG monitoring should be initiated for drug administration." The dependency arc score is 0.83, identifying a parallel procedural relationship between "establishing intravenous access" and "ECG monitoring." The semantic role labeling channel identifies a causal relationship where "drug administration" depends on "establishing intravenous access," with a role confidence score of 0.91. According to the formula, when... =0.6、 When the ratio is 0.4, the weighting of the relationship between intravenous access establishment and drug bolus injection is calculated as follows: The result is w=0.862, which exceeds the set threshold of 0.8, so it is retained as a directed edge representing a causal relationship. The final set of directed edges contains three valid edges, representing process sequence, parallel processes, and causal relationships, thus achieving high-precision construction of the semantic connectivity of the core knowledge graph.

[0065] S1.4: The knowledge graph node set and directed edge set are fused and modeled into a graph structure. Graph database technology is used to construct a core knowledge subgraph of emergency nursing with a topological structure to realize semantic connectivity modeling between knowledge points.

[0066] After extracting the knowledge graph node set and directed edge set, the input conditions include structured emergency rescue procedures, critical care, emergency drug use and other fields of node datasets and semantic relationship sets between nodes.

[0067] A graph fusion modeling method (parameters: node set N, edge set E, fusion mode: semantic connectivity priority) is adopted to integrate the node set and edge set in a unified graph data structure, realizing the corresponding binding of node entities and semantic relationships. Furthermore, a graph pattern matching algorithm (parameters: node attribute labels, relationship type constraints) is used to perform pattern verification on the fused graph structure according to the emergency nursing knowledge logic template, and node connection methods that do not conform to the preset structural specifications are corrected, resulting in a structurally normalized graph dataset.

[0068] Furthermore, through graph database construction algorithms, standardized graph datasets are written into the database system to achieve persistent storage of node entities, attribute fields, and relation edges, and to generate a unified graph index scheme to improve the efficiency of subsequent semantic queries.

[0069] S1.5: Based on the core knowledge subgraph of emergency nursing, execute the graph optimization algorithm to prune redundant paths and isolated nodes to improve the connectivity and interpretability of the knowledge graph, and ensure its stability and operability as the main knowledge system for personalized path planning.

[0070] Based on the emergency nursing core knowledge subgraph data obtained from graph structure fusion modeling, a node degree distribution analysis method based on graph traversal statistics was adopted (parameter: node degree threshold d). min Path length threshold l max This enables the identification of low-degree nodes and long-chain redundant path structures in the graph.

[0071] Furthermore, a path pruning algorithm based on shortest path coverage (parameter: coverage threshold c) is used. thr The algorithm uses path weights (w) to sort the weights of redundant paths and prune paths with coverage below a threshold to obtain a simplified set of connected paths.

[0072] Furthermore, an isolated node detection method is adopted (parameter: minimum number of nodes in the connected components n). min This allows for the identification of isolated nodes from the core knowledge subgraph that do not have a valid semantic connection with key nodes in the main graph, and the generation of an index list of isolated nodes.

[0073] Furthermore, a retention decision algorithm based on node centrality index (parameter: betweenness centrality b) is used. c eigenvector centrality e c This allows for the assessment of the importance of isolated nodes and the removal of nodes below the centrality threshold, resulting in a node set with optimized structure.

[0074] By using a connectivity optimization algorithm based on graph Laplacian spectrum analysis, the results of the previous step are transformed into topological data with enhanced stability, thereby achieving the expected technical effect of high connectivity and high interpretability of the emergency nursing core knowledge subgraph in personalized path planning.

[0075] For example, in actual system deployment, a node degree threshold d is set for the core knowledge subgraph that includes three knowledge domains: emergency response procedures, critical care, and emergency drug use. min 2. Path length threshold l max With a degree of 6, a node degree distribution analysis was performed, resulting in a list of 12 low-degree nodes. Shortest path coverage pruning was applied to the path set, with a coverage threshold c set. thr The path weights are set to 0.85, and the average semantic relevance is used. Eighteen redundant paths with low coverage are removed, retaining only the set of high-weight paths. For isolated node detection, n is set... min The value is 3, and 8 isolated nodes were detected; in the importance assessment, the betweenness centrality threshold b is set. c The eigenvector centrality threshold is 0.15. c The centrality was set to 0.12, and five low-centrality nodes were removed from the isolated nodes. Finally, graph Laplacian spectral analysis was performed, with the spectral gap g... thr With a value of 0.05, the connectivity index of the graph is significantly improved by optimizing the feature vector v, forming a highly connected emergency nursing core knowledge subgraph. In subsequent path planning tests, this results in a reduction in the average number of hops of learning path nodes and a significant improvement in path coverage.

[0076] Step S2: Extract auxiliary knowledge points from fields related to psychological intervention, doctor-patient communication, and disaster medicine, construct an independent cross-domain auxiliary knowledge subgraph, and establish a weak association mapping between it and the core knowledge subgraph to represent potential knowledge synergy relationships. Specifically, this includes: S2.1: Based on professional literature and teaching resources in non-core fields such as psychological intervention, doctor-patient communication, and disaster medicine, a cross-domain knowledge corpus is constructed to extract auxiliary knowledge points that meet the competency development needs of emergency nurses, which serve as the original input for constructing a cross-domain auxiliary knowledge subgraph.

[0077] Based on professional literature and teaching resources in non-core fields such as psychological intervention, doctor-patient communication, and disaster medicine, the types and format requirements of input data sources for cross-domain knowledge acquisition are determined. The input data includes three categories: structured text, semi-structured course outlines, and unstructured research reports.

[0078] By employing a literature crawling and teaching resource database interface call method (parameters: target domain keyword list, time range, and document type filtering conditions), we can achieve batch collection of knowledge data in a specified domain and generate raw text datasets.

[0079] Furthermore, through a domain classification algorithm (parameters: TF-IDF weight matrix, professional dictionary), the collected literature resources are classified by domain, resulting in three initial classification datasets: psychological intervention, doctor-patient communication, and disaster medicine.

[0080] Furthermore, a text quality assessment method (parameters: integrity scoring model, duplication rate detection threshold) is used to perform quality detection on the domain classification dataset and generate quality assessment results containing usability labels.

[0081] Furthermore, a domain-oriented knowledge requirement matching algorithm is used (parameters: emergency nurse competency development requirement vector, cosine similarity threshold). This allows for the screening of capability relevance in high-quality datasets and the generation of a candidate set of auxiliary knowledge that meets the needs of optimizing the comprehensive capabilities of emergency nurses.

[0082] By using a cross-domain corpus construction method, the auxiliary knowledge candidate set is transformed into a cross-domain knowledge corpus with a unified encoding format, thereby preparing the raw input for the construction of cross-domain auxiliary knowledge subgraphs.

[0083] For example, in the data preparation phase of an emergency nursing training platform, the latest five years of professional literature in the fields of psychological intervention, doctor-patient communication, and disaster medicine were selected. This article accesses the disaster medicine resource database to obtain instructional videos. Hours, Psychological Intervention Course Materials This paper uses a literature crawling program with keyword parameters set to "emergency + psychology", "emergency + communication", and "emergency + disaster", and a time range parameter of recent. In 2018, the document type filtering criteria selected journal articles and conference papers. The domain classification phase used a combination of the TF-IDF matrix and a psychological intervention dictionary. (Terms), Doctor-Patient Communication Dictionary () (terms) and the Dictionary of Disaster Medicine ( The document is categorized using a set of terms. Three domain scores are calculated for each document, and the highest score is used as the domain label for that document, resulting in the psychological intervention category. Articles, Doctor-Patient Communication Articles, Disaster Medicine In the quality assessment phase, the completeness scoring model sets the field coverage weight to [value missing]. Information content weight is The repetition rate detection threshold is set to Remove those with high repetition rates This is a literature review. In the competency-related screening process, the competency development needs vector for emergency nurses (including the weight of psychological counseling ability) was used. Communication and coordination skills weighting Weight of emergency decision-making ability Cosine similarity calculations were performed on the TF-IDF vectors of each document, retaining those with similarity scores no lower than [value missing]. Literature and resources to form psychological intervention categories Articles, Doctor-Patient Communication Articles, Disaster Medicine This cross-domain knowledge corpus is used for subsequent structured triple extraction. This construction process significantly improves the relevance and usability of the input data, providing high-quality raw corpus support for the accurate construction of cross-domain auxiliary knowledge subgraphs.

[0084] S2.2: Perform entity recognition and relation extraction processing on the text content in the cross-domain knowledge corpus. Identify key knowledge entities based on the named entity recognition algorithm and extract semantic relationships between entities by combining dependency parsing to construct a structured cross-domain knowledge triple set.

[0085] For structured and unstructured text content in cross-domain knowledge corpora, a named entity recognition algorithm (parameter settings: domain-specific dictionary, entity category set including event, skill, terminology, etc.) is used to achieve accurate location and boundary division of auxiliary knowledge point entities.

[0086] Furthermore, by using a Conditional Random Field (CRF) sequence labeling model (feature templates include part-of-speech tags, context window words, and glyph features), entity labeling in cross-domain corpora is assigned, and multi-domain knowledge point labeled sequence data is obtained.

[0087] Furthermore, through dependency parsing (based on the enhanced Universal Dependencies grammar framework, with parameters set to include subject-verb, verb-object, modification, and conditional relations), we construct syntactic dependency trees between entities in cross-domain corpora and extract direct and indirect dependency path data between entity nodes.

[0088] Furthermore, by combining dependency path data with the results of the Semantic Role Labeling (SRL) algorithm, the semantic association types between entities are identified, including causal relationships, temporal relationships, conditional triggering relationships, and synergistic relationships, and a set of semantic relationship identifiers is generated.

[0089] A relation extraction strategy combining rule matching and word vector similarity calculation is adopted to map the named entity recognition results to the semantic relation identifier set, generating a cross-domain knowledge triple set. Each triple consists of (main entity, relation, object entity), and attribute labels are attached to both the entity and relation sides for semantic weight initialization in the subsequent graph construction stage.

[0090] Through the multi-level dependency processing of the above algorithm chain, the entity recognition result of the previous step is transformed into a cross-domain knowledge triple structure with semantic computability, realizing the structured expression of auxiliary knowledge points and the explicit representation of their relationships, providing high-precision input data for the subsequent construction of cross-domain auxiliary knowledge subgraphs.

[0091] For example, in the knowledge extraction scenario of psychological intervention and doctor-patient communication, the input corpus contains 1000 emergency nurse interview records and training material excerpts. The named entity recognition algorithm is set to eight entity categories, including "psychological assessment methods," "patient emotional state," and "communication skills." The corpus is labeled using a CRF model (window size 5, with both part-of-speech and glyph features enabled), achieving an average F1 score of 0.92. In the dependency parsing stage, a neural network-based UD parser is used to construct a dependency tree for each sentence and extract the verb-object relationship path. The parameters are set to the optimal model weights selected through cross-validation, obtaining the dependency path length and connection type between each pair of entities. In the semantic role labeling stage, the causal relationship between "psychological assessment methods" and "patient emotional state," as well as the synergistic relationship between "communication skills" and "anxiety relief," are identified. Using this batch of data as input, during the relationship extraction process, a threshold calculated using word vector cosine similarity is used. As a condition for confirming relationships, valid relationship pairs are selected, ultimately generating approximately 4,800 cross-domain knowledge triples, such as: (psychological assessment methods, relief, patient emotional state) and (communication skills, collaboration, anxiety relief). This structured triple set can significantly improve the semantic connectivity and accuracy of the cross-domain auxiliary knowledge subgraph during subsequent graph construction.

[0092] S2.3: Based on the cross-domain knowledge triple set, a cross-domain auxiliary knowledge subgraph is constructed using a graph neural network model. The cross-domain auxiliary knowledge subgraph contains auxiliary knowledge point nodes and semantic association edges between them, forming an auxiliary knowledge expression structure independent of the core knowledge system.

[0093] Based on the input data of cross-domain knowledge triple sets, a graph neural network construction method is adopted (parameters: node initial vectors are derived from cross-domain knowledge entity embeddings, and edge weights are derived from triple relation confidence) to realize the generation and structured construction of the initial representation of nodes in cross-domain auxiliary knowledge subgraphs.

[0094] Furthermore, by using the adjacency matrix normalization method (parameter: adopting a symmetric normalization strategy to prevent weight shifts caused by degree differences), the local topological features between cross-domain knowledge nodes are balanced, and normalized adjacency matrix data is obtained.

[0095] Furthermore, a graph convolutional network algorithm with a message passing mechanism (parameters: 2 convolutional layers, ReLU activation function) is used to aggregate and update the features of neighboring nodes, generating node representations containing first-order and second-order semantic neighborhood information.

[0096] Furthermore, through the graph neural network extension module of the attention mechanism (parameter: attention weight calculation adopts the fusion of cosine similarity and weight normalization), the weighted adjustment of the node representation of edge pairs of different relation types is realized, and an optimized node embedding vector containing relation type weights is generated.

[0097] Furthermore, by using a node vector clustering analysis method (parameter: cosine distance clustering threshold set to 0.75), cross-domain knowledge nodes with high semantic similarity are merged into the same auxiliary knowledge cluster to reduce graph structure redundancy and enhance semantic coherence.

[0098] By constructing and optimizing the graph neural network as described above, the cross-domain knowledge triplet set result from the previous step is transformed into cross-domain auxiliary knowledge subgraph data with strong semantic connectivity and stable structure, thus achieving the technical effect of an auxiliary knowledge expression structure independent of the core knowledge system.

[0099] For example, in the implementation process, the input cross-domain knowledge triple set includes 120 nodes from the psychological intervention domain, 85 nodes from the doctor-patient communication domain, and 60 nodes from the disaster medicine domain, with a total of 1500 triple relations and an average confidence level of 0.82. The initial node vectors are generated into 768-dimensional feature vectors using BERT encoding, and the adjacency matrix is ​​symmetrically normalized so that all diagonal elements are 1. A two-layer graph convolutional network is used, with the first layer weight matrix having a size of [missing information]. The size of the second layer weight matrix is The activation function is ReLU. In the attention mechanism's relation type weight calculation, cosine similarity ranges from 0.65 to 0.98, with a maximum weight of 0.31 after Softmax normalization. Cluster analysis divides nodes with similarity above 0.75 into 42 auxiliary knowledge clusters, with an average of 6 nodes per cluster. The total number of nodes in the final cross-domain auxiliary knowledge subgraph is reduced from 265 to 252, and the number of edge connections is optimized from 1500 to 1320, verifying that the subgraph significantly improves connectivity and conforms to semantic collaboration logic.

[0100] S2.4: Perform semantic alignment processing on the key knowledge point nodes in the core knowledge subgraph and the auxiliary knowledge point nodes in the cross-domain auxiliary knowledge subgraph. Calculate the semantic similarity between nodes based on the cross-modal semantic embedding model, identify knowledge node pairs with potential collaborative relationships, and establish cross-graph mapping relationships.

[0101] For the key knowledge point nodes in the core knowledge subgraph and the auxiliary knowledge point nodes in the cross-domain auxiliary knowledge subgraph, a cross-modal semantic embedding model (parameters: text semantic dimension 768, graph structure embedding dimension 128, vector normalization method is L2 paradigm) is used to generate high-precision semantic vectors, so as to realize the unified semantic representation of knowledge nodes in different domains.

[0102] Furthermore, by using a cross-domain vector space alignment algorithm (parameters: alignment matrix dimension 896×896, optimization method is to minimize mean square similarity error), the coordinates of core knowledge nodes and auxiliary knowledge nodes in the shared embedding space are aligned, and a cross-domain node alignment vector set is obtained.

[0103] Furthermore, a node relevance calculation method based on cosine similarity is adopted (parameter: similarity threshold set to 0.75) to perform relevance calculation on each pair of core nodes and auxiliary nodes in the alignment vector set, and to generate a preliminary collaborative relationship scoring matrix for node pairs. The formula is as follows:

[0104] in For the semantic vectors of core knowledge nodes, Semantic vectors for cross-domain auxiliary knowledge nodes.

[0105] Furthermore, a semantic noise suppression and weight reconstruction algorithm (parameters: noise suppression coefficient of 0.2, weight reconstruction iteration number of 50) is used to denoise the scoring matrix and correct the relevance weights, and generate a set of potential collaborative relationship node pairs with high confidence.

[0106] The cross-graph mapping generation module transforms the set of high-confidence node pairs into a cross-domain mapping relationship data structure, achieving a semantic-level weak association mapping effect between the core knowledge subgraph and the cross-domain auxiliary knowledge subgraph.

[0107] For example, in the core knowledge subgraph of emergency nursing, the node "Response procedure for sudden cardiac arrest in critically ill patients" is selected to generate a semantic vector with a length of 768. After normalization, it is stored at index position #51 in the shared embedding space. In the knowledge subgraph of psychological intervention assistance, the node "Psychological comforting strategies for acute patients and their families" is selected to generate a semantic vector with a length of 768. It is also normalized and stored at index position #212. The alignment matrix obtained by training a cross-domain vector space alignment algorithm projects the two vectors onto a unified coordinate system, and the cosine similarity value is calculated. Higher than the set threshold Therefore, it enters the preliminary collaborative relationship scoring matrix. After noise suppression processing, the weight coefficients of this correspondence are reduced from... Revised to The set of node pairs with the highest confidence level is retained. Finally, in the cross-graph mapping generation module, this mapping relationship is transformed into a weakly related mapping entry, which is used to trigger the cross-domain introduction of psychological intervention knowledge in the subsequent path fusion stage, thereby improving the comprehensive handling capabilities of emergency nurses at the emergency scene.

[0108] S2.5: Based on the cross-graph semantic similarity results, construct a weak association mapping matrix. The weak association mapping matrix represents the potential knowledge collaboration relationship between the core knowledge subgraph and the cross-domain auxiliary knowledge subgraph, which is used to trigger the cross-domain knowledge fusion mechanism in the subsequent path planning process.

[0109] Based on the cross-graph semantic similarity result matrix data, a sparse matrix construction algorithm (parameters: similarity threshold τ, density limit δ) is used to initially generate a weak association mapping matrix between the core knowledge subgraph and the cross-domain auxiliary knowledge subgraph.

[0110] Furthermore, by using a weighted adjacency matrix combination method (parameters: weight coefficient α comes from semantic similarity, weight coefficient β comes from node importance index), the elements of the weak association mapping matrix are fused and weighted, and a weighted weak association matrix data structure reflecting the cross-domain collaboration strength is obtained.

[0111] Furthermore, by using matrix normalization methods (parameters: column normalization, row normalization), the dimensions of the mapping relationship weights are unified, and a normalized weak association mapping matrix is ​​generated for the comparability verification of association weights between different knowledge subgraphs.

[0112] Furthermore, a threshold filtering algorithm is adopted (parameter: the filtering threshold σ is set according to the cross-domain fusion triggering condition) to mark and remove weakly correlated edges below the threshold in the normalized matrix, and generate a sparse weakly correlated mapping matrix to reduce redundant cross-domain mapping edges.

[0113] By using a graph index optimization algorithm, the sparse weak association mapping matrix from the previous step is transformed into a cross-domain knowledge fusion index that can be quickly accessed and triggered in the path planning algorithm, thereby enabling a precise triggering mechanism for cross-domain fusion in the subsequent path planning stage.

[0114] like Figure 2 As shown, step S3 involves determining the main learning path in the core knowledge subgraph based on the nurse's current competency profile, and identifying target nodes that require the introduction of cross-domain knowledge as triggering conditions for cross-domain knowledge fusion. Specifically, this includes: S3.1: Obtain nurse competency profile data, which includes multi-dimensional competency indicators such as nurses' mastery of basic nursing knowledge, clinical emergency handling ability, and communication and psychological intervention ability. Based on the competency profile data, determine the nurse's current knowledge level status to generate an individual nurse competency status vector.

[0115] Based on multi-dimensional competency indicators such as nurses' mastery of basic nursing knowledge, clinical emergency response capabilities, and communication and psychological intervention abilities, an input feature matrix is ​​constructed as the original data source for competency profiles.

[0116] The feature standardization method (parameters: mean 0, variance 1) is used to normalize each dimension of the input feature matrix to achieve consistency of different indicator dimensions.

[0117] Furthermore, based on the principal component analysis algorithm (parameter: the dimension of the covariance matrix equals the number of capability indicators), comprehensive capability features are extracted and the feature dimension is reduced to improve the stability of subsequent vector calculations and generate a dimensionality-reduced capability feature vector.

[0118] Furthermore, a weighted feature fusion method (weighting based on the importance coefficient of each competency indicator in the requirements of the nursing position) is adopted to linearly weight and combine the dimensionality-reduced competency feature vector to form a comprehensive competency score vector.

[0119] Furthermore, the modulus of the comprehensive competency score vector is standardized to a unit length using a vector normalization method to generate an individual nurse competency status vector, which is then prepared for input into the path matching algorithm for determining the main learning path.

[0120] Through the above normalization, dimensionality reduction and weighted fusion processing, the multidimensional capability indicators are transformed into a structured vector of individual nurse capabilities that can be used for graph search, thereby realizing a quantitative expression of the nurse's current knowledge level.

[0121] S3.2: Based on the nurse's individual ability state vector, the path matching algorithm is executed to search for the optimal main learning path in the core knowledge subgraph. The core knowledge subgraph includes knowledge points on emergency procedures, critical care, and emergency drug use, as well as their logical connections, to output a main learning path sequence that matches the nurse's current ability level and has progressive potential.

[0122] Based on the individual nurse's ability state vector, a multi-feature similarity path matching algorithm (parameters: ability state vector dimension = 12, total number of nodes in the core knowledge subgraph = 450) is used to calculate the matching degree between each node in the core knowledge subgraph and the ability state.

[0123] Furthermore, by using a graph search optimization method (parameters: the heuristic function is the weighted sum of the node capability gaps, and the weight ratios are: basic nursing 0.4, emergency treatment 0.4, and psychological communication 0.2), the matching degree data is dynamically updated in the graph structure, and the cumulative matching score of each possible path is obtained.

[0124] Furthermore, a heuristic A-search algorithm (parameters: evaluation function f(n) = g(n) + h(n), where g(n) is the cost of the already traveled path, and h(n) is the predicted value of the remaining gap based on the capability state vector) is used to prune the search space, reducing low-matching paths that do not match the current capability state and generating a set of remaining candidate paths. Further, a path connectivity evaluation model (parameters: connectivity threshold = 0.75, path length limit = 15 nodes) is used to verify the logical continuity of knowledge points in the candidate path set and eliminate paths with knowledge gaps, obtaining excellent candidate paths after connectivity verification. Further, an advancement evaluation algorithm is used to rank candidate paths on the capability improvement gradient, selecting the optimal main learning path sequence that matches the current level and has room for continuous improvement. Through a path matching algorithm combining multi-feature similarity and heuristic graph search, the mapping relationship between the capability state vector and the core knowledge subgraph nodes is transformed into an optimal path sequence, achieving accurate matching and advancement guarantee of the main learning path. For example, in a competency profile of an emergency room nurse, the basic nursing knowledge mastery is 0.65, emergency response capability is 0.72, and communication and psychological intervention capability is 0.50, generating a 12-dimensional competency state vector covering sub-dimensions such as basic medical skills, nursing process execution, critical care, and medication guidelines. The core knowledge subgraph consists of 450 nodes, with node semantics including emergency procedure steps (node ​​depth maximum 12), critical care techniques, and emergency drug use guidelines. A multi-feature similarity path matching algorithm is used, with a matching threshold of 0.6, to calculate the similarity between each node and the competency state. For example, the similarity of the emergency procedure node "airway opening" is 0.78, and the similarity of the node "intravenous administration" is 0.70. An A search is performed, with the heuristic function h(n) being the normalized value of the sum of squared competency gaps. Paths longer than 15 and with connectivity less than 0.75 are pruned. Five candidate pathways were retained after connectivity assessment, with pathways having an advancement assessment index (P) in the range of 0.15–0.25 deemed suitable for advanced learning. The final selected pathway sequence was {airway opening → artificial respiration → chest compressions → intravenous drug administration → rapid blood pressure assessment → key points of psychological reassurance}. This sequence covers continuous skills from basic first aid to psychological intervention, significantly improving comprehensive emergency response capabilities.

[0125] S3.3: Perform knowledge gap analysis on each target node in the main learning path sequence to identify the gap between the current nurse's competence and the competence required by the target node, and generate a node-level knowledge gap vector to quantitatively assess whether each target node needs to be supplemented with cross-domain knowledge.

[0126] S3.4: Based on the node-level knowledge gap vector and combined with the preset knowledge fusion trigger threshold, determine whether cross-domain knowledge is introduced into each target node, generate a cross-domain knowledge introduction trigger mark set, and determine the target node set that needs to be fused with cross-domain knowledge.

[0127] Based on node-level knowledge gap vector data, a threshold determination method (parameter: preset knowledge fusion trigger threshold τ) is used to determine the cross-domain knowledge introduction requirements of each target node in the main learning path.

[0128] Furthermore, based on the comparison algorithm, the gap score of each target node is compared with the trigger threshold one by one, and a binary judgment result set is obtained.

[0129] Furthermore, through a set generation algorithm (parameter: r) i The judgment result set aggregates the node indices with positive values ​​in the judgment result set to generate a cross-domain knowledge introduction trigger mark set M.

[0130] Furthermore, using a semantic mapping retrieval algorithm (parameters: M set, weak association mapping matrix of core knowledge subgraph and cross-domain auxiliary knowledge subgraph), the node correspondence between the trigger tag set and the cross-domain auxiliary knowledge is matched, and the target node set T that needs to be fused with cross-domain knowledge is obtained.

[0131] By using a cross-domain knowledge fusion trigger judgment method, node-level gap vectors are transformed into cross-domain knowledge trigger marker data, enabling accurate identification of target nodes that need to be supplemented across domains in personalized learning paths.

[0132] For example, in the process of constructing a personalized pathway for emergency nurses, the input node-level knowledge gap vector is [0.15, 0.42, 0.37, 0.61], and the preset knowledge fusion trigger threshold τ is 0.4. A comparison algorithm is used to score the gap at the second node. Compared with τ, the result is a positive judgment; the gap score of the third node. Compared with τ, the result is a negative judgment; the gap score of the fourth node. Compared with τ, the result is a positive determination, and r is finally obtained. i The result set is [0,1,0,1]. A set of cross-domain knowledge introduction trigger markers, M={2,4}, is formed based on a set generation algorithm. Combining this with a weak association mapping matrix, semantic mapping retrieval is performed on set M to obtain the target node set T={psychological intervention skills nodes, disaster medical emergency coordination nodes} that needs to be integrated with cross-domain knowledge. This set will serve as the input condition for subsequent cross-domain knowledge screening and distillation processes, significantly improving the accuracy of the recommended path's response to cross-domain knowledge requirements.

[0133] S3.5: Based on the cross-domain knowledge introduction trigger mark set, construct a cross-domain knowledge fusion request signal and use it as the input condition for subsequent cross-domain knowledge screening and distillation processes, so as to achieve accurate triggering and dynamic adaptation of cross-domain knowledge introduction in the personalized path recommendation process.

[0134] Based on cross-domain knowledge, a set of trigger markers is introduced, and a signal encoding method (parameters: trigger marker set, node identifier, path sequence index) is used to achieve the initial format encoding of the cross-domain knowledge fusion request signal, so as to ensure that the subsequent processing link can accurately interpret the trigger conditions.

[0135] Furthermore, by using an event-driven mapping algorithm (parameters: trigger tag set, weak association mapping matrix), the cross-graph adjacency node mapping of the request signal is realized, and the semantic coupling sequence data of the request signal between the core knowledge subgraph and the cross-domain auxiliary knowledge subgraph is obtained.

[0136] Furthermore, a semantic compression coding method (parameters: coupled sequence data, node semantic vector) is adopted to achieve data compression and feature preservation of the request signal, and to generate a lightweight fused request signal packet containing the node semantic instruction set.

[0137] Furthermore, by utilizing a priority scheduling algorithm (parameters: node gap score, fusion node influence coefficient), the execution order of each fusion instruction within the signal packet is optimized, and a weighted serialized signal packet structure is generated.

[0138] By using protocol adaptation processing, the serialized signal packets from the previous step are transformed into a set of input parameters for the cross-domain knowledge filtering and distillation process, enabling precise triggering and dynamic adaptation of cross-domain knowledge introduction during personalized path recommendation.

[0139] like Figure 3 As shown, step S4 involves performing textual semantic analysis on candidate cross-domain knowledge points, calculating their semantic relevance score to the current main path target node, and combining this with the information entropy value of the knowledge point in the original domain to comprehensively evaluate the information density per unit of content. Specifically, this includes: S4.1: Preprocess the text content of candidate cross-domain knowledge points, including word segmentation, stop word removal, and stemming, to generate standardized text semantic representation vectors.

[0140] S4.2: Embed the standardized text semantic representation vector based on the BERT semantic encoding model to obtain a high-dimensional semantic feature vector of cross-domain knowledge points.

[0141] S4.3: Extract the semantic feature vector of the current main path target node and perform cosine similarity calculation with the high-dimensional semantic feature vector of the cross-domain knowledge point to obtain the semantic relevance score between the two.

[0142] Given the high-dimensional semantic feature vectors of candidate cross-domain knowledge points and the target node information of the main learning path, a node semantic feature vector extraction algorithm (parameters: node description text, semantic encoding model type, vector dimension setting) is used to achieve high-dimensional semantic vectorization of the text description of the target nodes of the main path. Furthermore, through a vector normalization method (parameters: normalization type selection, range setting), a basis for comparing each semantic feature vector under the same dimension is established, resulting in a standardized set of semantic feature vectors for the target nodes of the main path.

[0143] The cosine similarity calculation method (parameters: vector dimension consistency check, value range setting) is used to calculate the semantic relevance between the main path target node and candidate cross-domain knowledge points, and obtain the original similarity score matrix. The formula is as follows:

[0144] in, The standardized semantic feature vector of the main path target node. This is a standardized semantic feature vector for candidate cross-domain knowledge points. The vector dot product of the two. and These are the Euclidean norms of the vectors. Furthermore, a vector norm normalization algorithm (parameters: L2 norm constraint, numerical stability adjustment) is used to stabilize the similarity calculation results and generate a list of semantic relevance scores between nodes.

[0145] A relevance-weighted smoothing algorithm (parameters: smoothing coefficient α, weight distribution function) is employed to smoothly adjust the similarity score list, reducing the impact of noise features on relevance calculation and obtaining a smoothed semantic relevance evaluation index set. Furthermore, a threshold filtering method (parameters: relevance threshold setting, filtering strategy type) is used to initially filter the semantic relevance scores, eliminating low-relevance knowledge points and obtaining a set of highly relevant cross-domain knowledge points.

[0146] Through the above calculation and screening process, the high-dimensional semantic feature matching results are transformed into a high-precision semantic relevance score between cross-domain knowledge points and main path target nodes, thereby improving the accuracy of knowledge point selection in subsequent information entropy-based evaluation.

[0147] For example, in the emergency nursing career path recommendation scenario, the main path target node "emergency treatment of acute heart failure" is encoded into a 768-dimensional semantic vector using the BERT-base model, and then normalized to form a standardized vector v using Z-score normalization. In the same scenario, the candidate cross-domain knowledge point "psychological comfort during patient crisis" is encoded into a 768-dimensional semantic vector w, and undergoes the same normalization process. In the cosine similarity calculation, the dot product of the node vectors is 0.354, with norms of 1.000 and 1.005 respectively. Substituting these values ​​into the formula:

[0148] The relevance score was approximately 0.352. After adjustment using a smoothing algorithm (α=0.1), the score became 0.345, which is higher than the preset relevance threshold of 0.3, thus entering the high-relevance cross-domain knowledge point identification set. This result can be preferentially retained for path enhancement when subsequently fused with the information entropy value, significantly improving the targeting and applicability of cross-domain recommended content.

[0149] S4.4: Based on the information entropy calculation model, the context distribution of the cross-domain knowledge point in its respective auxiliary knowledge subgraph is modeled to obtain the information entropy value of the knowledge point, which reflects its knowledge density in the original domain.

[0150] For the context distribution data of cross-domain knowledge points in the auxiliary knowledge subgraph, the context window statistical method (parameter: window radius of 3 nodes) is used to model the frequency distribution of the occurrence of semantic nodes in the neighborhood of the knowledge point.

[0151] Furthermore, by using a node probability normalization algorithm (parameter: Laplace smoothing coefficient set to 0.01), the probability of candidate cross-domain knowledge points and their context nodes is standardized, and a probability distribution vector is obtained.

[0152] Furthermore, an information entropy calculation model (parameter: discrete probability distribution form) is adopted to realize the quantitative calculation of the information entropy of knowledge points.

[0153] S4.5: Perform weighted fusion processing on the semantic relevance score and information entropy value to generate a comprehensive information density evaluation index for cross-domain knowledge points, which is used for subsequent identification and filtering of redundant content.

[0154] The semantic relevance score obtained in step S4.3 and the information entropy value obtained in step S4.4 are combined using a weighted fusion algorithm (parameter setting: relevance weight). Information entropy weight This enables the calculation of the comprehensive information density index of cross-domain knowledge points.

[0155] Furthermore, by using a linear normalization method (parameter: normalization range set to 0 to 1), the relevance scores and information entropy values ​​of different dimensions are uniformly processed, and the mapped standardized input data pairs are obtained. .

[0156] Furthermore, the comprehensive information density assessment value is calculated using a weighted summation formula, defined as follows:

[0157] in, As a comprehensive information density evaluation index, This is a standardized value for semantic relevance. The information entropy is the standardized value. and These are weighting coefficients that are dynamically set based on the system optimization strategy.

[0158] Furthermore, an adaptive weight adjustment algorithm is used (parameter updates are based on user feedback matching degree of recently recommended paths and knowledge mastery effect) to achieve... Dynamic optimization is performed to generate a comprehensive information density evaluation result with optimal discrimination in recommendation scenarios.

[0159] By using a comprehensive weighted processing method, the semantic relevance and information entropy analysis results from the previous step are transformed into a quantifiable and comparable comprehensive information density score, thereby achieving the expected technical effect of cross-domain knowledge point redundancy identification and content refinement.

[0160] For example, in the scenario of optimizing the career development path for emergency nurses, semantic relevance weights are set. The information entropy weight is 0.6. The relevance score of candidate knowledge points in a certain psychological intervention field is 0.4 after normalization. The information entropy value is 0.85. The value is 0.65. Applying the weighted fusion formula:

[0161] get =0.51 + 0.26 = 0.77. The system uses a dynamic evaluation model to... and Periodic adjustments were made, and it was observed that when the weight ratio was optimized to 0.55:0.45, the user learning completion rate of this knowledge point in path recommendation was significantly improved. This indicates that the information density evaluation result under this weight configuration is more in line with the actual learning needs and effectively avoids the introduction of knowledge points with low relevance and high entropy.

[0162] Step S5: Based on the information density assessment results, highly redundant cross-domain knowledge points are filtered and marked, high-density knowledge points are retained, and a set of knowledge points to be distilled is output. Specifically, this includes: S5.1: Normalize the semantic density evaluation results of cross-domain knowledge points to eliminate evaluation bias between different knowledge sources and obtain a density score vector with unified dimensions, which serves as the basic input for subsequent filtering operations.

[0163] S5.2: Based on the preset semantic density threshold range, the normalized density score vector is binarized and labeled. Cross-domain knowledge points below the threshold are labeled as redundant knowledge points, forming a set of redundant knowledge point labels to control the content redundancy in the recommendation path.

[0164] S5.3: Perform semantic association tracing on the knowledge points in the redundant knowledge point tag set, identify their original semantic nodes in the cross-domain auxiliary knowledge subgraph, and evaluate their potential synergistic impact on the main knowledge system based on the node importance index, generating a potential retention candidate set.

[0165] S5.4: Based on the weighted fusion value of the synergistic impact assessment results and density scores, a secondary screening is performed on the knowledge points in the potential retention candidate set. Knowledge points with synergistic impact below the dynamic threshold are eliminated, and high-density knowledge points with semantic fusion potential are retained to form a preliminary set of knowledge points to be distilled.

[0166] For the set of knowledge points in the potential candidate set, a weighted fusion calculation method is used (parameter: synergistic impact assessment value vector W). c semantic density score vector W d The fusion of weighting coefficients α and β enables quantitative processing for comprehensive screening of multiple indicators.

[0167] Furthermore, the comprehensive evaluation value of each knowledge point is calculated using a weighted fusion formula, as follows:

[0168] in, For comprehensive evaluation, The weighting coefficients for the synergistic impact assessment value. The weighting coefficients for the semantic density evaluation value. The metric for the impact of node collaboration. Score the semantic density of nodes.

[0169] Furthermore, based on the dynamic threshold calculation algorithm (parameters: statistical distribution of knowledge point importance, mean μ and standard deviation σ of historical fusion evaluation), a dynamic retention threshold is generated, as shown in the following formula:

[0170] in, For dynamic filtering thresholds, This is the historical average. The standard deviation of the historical comprehensive evaluation. This is the threshold adjustment coefficient.

[0171] Furthermore, a secondary screening algorithm is executed to remove knowledge points whose comprehensive evaluation value is lower than the dynamic threshold, and to mark knowledge points that are higher than or equal to the threshold and have semantic fusion potential as preliminary retained knowledge points, so as to form a preliminary set of knowledge points to be distilled.

[0172] Furthermore, using a semantic fusion potential assessment model (parameters: cross-domain semantic alignment vector similarity matrix and main knowledge system adjacency relation matrix), the semantic fit between each initially retained node and the embedding position of the main graph is calculated. Nodes with insufficient fit are removed, while nodes with high fit are retained to ensure the effectiveness of subsequent distillation fusion.

[0173] Through the aforementioned weighted fusion and dynamic threshold filtering algorithm, the potential candidate set to be retained in the previous step is transformed into a preliminary set of knowledge points to be distilled with optimized structure, thereby enhancing the synergistic fusion effect with the main knowledge system while ensuring high semantic density.

[0174] For example, in an emergency nurse career development path planning scenario, assume the potential candidate set contains 40 cross-domain knowledge points, each with a synergistic impact assessment value and a semantic density score. Let the weighting coefficient α for the synergistic impact assessment value be 0.6, and the weighting coefficient β for the semantic density score be 0.4. In actual calculations, if a certain knowledge point has a synergistic impact assessment value of 0.85 and a semantic density score of 0.75, the comprehensive evaluation value is calculated as follows:

[0175] This value was obtained through calculation. In the historical comprehensive evaluation data of this batch of knowledge points, the mean μ is 0.78, the standard deviation σ is 0.05, and the threshold adjustment coefficient k is taken as 1.0. Therefore, the dynamic threshold formula is as follows:

[0176] Its value is This knowledge point was removed because its overall evaluation score of 0.81 was below the threshold of 0.83. The other knowledge point had a synergistic impact evaluation score of 0.90 and a semantic density score of 0.88; its overall evaluation score was calculated as follows:

[0177] Its value is The value exceeded the threshold of 0.83, and the semantic fusion potential model evaluation fit was 0.92. Therefore, this knowledge point was included in the initial distillation set. The actual effect of this process is that approximately 18 cross-domain knowledge points with high fusion potential are retained from the initial 40 candidate knowledge points, significantly improving the accuracy and knowledge relevance of the recommendation path.

[0178] S5.5: Perform knowledge consistency verification on the initial set of knowledge points to be distilled, detect whether there are semantic conflicts or logical contradictions in the main knowledge system, and output the final set of knowledge points to be distilled after semantic verification, as the input data of the knowledge distillation engine.

[0179] The initial set of knowledge points to be distilled is input into the consistency verification module. A logical reasoning rule matching algorithm (parameters: core knowledge subgraph structure description, knowledge point semantic vector features) is used to realize the semantic context positioning of the knowledge points to be distilled in the main knowledge system and the binding of the corresponding logical rules.

[0180] Furthermore, a conflict detection algorithm (parameters: semantic similarity threshold, logical relationship constraints) is used to identify semantic conflicts between the target node and its neighboring nodes, and a conflict marker set is obtained.

[0181] Furthermore, through the conflict resolution algorithm (parameters: rule priority matrix, node weight coefficient), the knowledge points in the conflict marker set are compared with the constraint rules of the main knowledge system to generate a set of candidate knowledge points after resolution.

[0182] Furthermore, a semantic consistency scoring model (parameters: node semantic vector, adjacency weight matrix) is used to calculate the semantic consistency between the resolved candidate knowledge points and the relevant node set within the core knowledge subgraph, and generate a consistency scoring vector.

[0183] Furthermore, a scoring filtering algorithm (parameter: consistency score threshold) is used to remove knowledge points whose consistency score vectors are lower than the threshold, thus obtaining the final set of knowledge points to be distilled.

[0184] By using consistency detection and filtering, the initial set from the previous step is transformed into a set of knowledge points that has been semantically verified and is logically coherent, thereby optimizing the stability and reliability of the content before it is input into the knowledge distillation engine.

[0185] For example, in the core knowledge subgraph of emergency nursing, the initial set of knowledge points to be distilled consists of 40 cross-domain knowledge points, each with a 512-dimensional semantic vector feature. The logical reasoning rule matching algorithm embeds the knowledge points into the main graph and locates them in their corresponding context. The conflict detection algorithm uses cosine similarity... (i.e., 0.6) is the conflict threshold. Seven nodes were identified as having semantic conflicts with their neighboring nodes, generating a conflict marker set. The conflict resolution algorithm uses a rule priority matrix with a dimension of 10×10 and weight coefficients ranging from 0.1 to 0.9, compressing the conflict set to three nodes after processing. The semantic consistency scoring model calculates the average consistency score of each node with its neighboring group based on the weight matrix, and sets a scoring filtering threshold of [value missing]. (i.e., 0.8), nodes with scores below 0.8 are removed. The final output set to be distilled contains 35 knowledge points. The verification results show that this set did not trigger any logical contradictions or semantic conflicts during the subsequent distillation and embedding process, significantly improving the fusion stability of the main knowledge system.

[0186] Step S6: Input the knowledge points to be distilled into the knowledge distillation engine, extract their core semantics into lightweight semantic vectors, and embed them into the corresponding positions in the core knowledge subgraph to form enhanced knowledge nodes. Specifically, this includes: S6.1: Perform text semantic vectorization processing on the cross-domain knowledge point set after redundancy filtering, and generate a high-dimensional semantic embedding representation based on the BERT model to extract its core semantic features.

[0187] S6.2: Based on the attention mechanism, perform semantic distillation on the high-dimensional semantic embedding representation, extract key semantic components and compress them into a low-dimensional semantic vector space to generate a lightweight semantic vector representation.

[0188] S6.3: Input the lightweight semantic vector into the knowledge distillation engine and perform semantic alignment and knowledge transfer operations to construct a semantic representation form compatible with the core knowledge subgraph.

[0189] Based on the lightweight semantic vectors compressed by semantic distillation, a cross-domain semantic alignment algorithm (parameters: core knowledge subgraph node embedding vector, cross-domain knowledge lightweight semantic vector) is used to realize the construction of a unified mapping relationship between the cross-domain semantic space and the core knowledge semantic space.

[0190] Furthermore, by using the orthogonal projection matrix solution method (parameters: cross-domain semantic embedding matrix, core subgraph semantic embedding matrix), the mapping matrix of the two semantic spaces is calculated, and the representation of the cross-domain lightweight vector in the core semantic coordinate system is obtained.

[0191] S6.4: Based on the semantic context of the target node in the core knowledge subgraph, calculate the semantic similarity between the lightweight semantic vector and its neighboring nodes to determine its embedding position in the main graph.

[0192] S6.5: Embed the lightweight semantic vector into the corresponding position in the core knowledge subgraph to generate an enhanced knowledge node, and update the semantic association weight and logical relationship of the node to form an enhanced master knowledge system that integrates cross-domain knowledge.

[0193] When a lightweight semantic vector after semantic distillation and its embedding position information are received as input, a graph node update algorithm (parameters: embedding position index, semantic vector feature dimension, and adjacent node list) is used to write the lightweight semantic vector into the target node data storage area of ​​the core knowledge subgraph, thereby realizing the data assignment of the enhanced knowledge node.

[0194] Furthermore, through the semantic weight redistribution algorithm (parameters: semantic similarity matrix of adjacent nodes, weight factor of current node), the association weights of the target node and its connected edges are updated based on the similarity calculation results of the lightweight semantic vector and adjacent nodes, and a node-edge weight update table is obtained.

[0195] Furthermore, by using a logical relationship reconstruction algorithm (parameters: node update table, original logical dependency set), the hierarchical relationship, causal chain and process order of the target node in the core knowledge subgraph are adjusted, so that the newly added or enhanced cross-domain semantic elements are integrated into the existing knowledge structure and a logical relationship update set is generated.

[0196] Furthermore, based on the topology consistency check algorithm (parameters: updated node weight table, updated logical relationship set), the connectivity and acyclicity of the updated subgraph are checked to ensure that the newly embedded enhanced nodes will not cause knowledge structure breaks or logical closure defects, and the topology state verification results are output.

[0197] Furthermore, an incremental storage synchronization mechanism (parameters: core knowledge subgraph database handle, update timestamp) is adopted to write the updated semantic association weights and logical relationships into the graph database persistent storage, ensuring that the updates of the enhanced master knowledge system are traceable and version managed.

[0198] Step S7: Based on the enhanced core knowledge subgraph, execute a path planning algorithm to generate personalized growth path recommendation results containing necessary cross-domain knowledge. Specifically, this includes: S7.1: Perform topological modeling of the node and edge relationships in the enhanced core knowledge subgraph, and generate node embedding representations based on graph neural network algorithms to extract the potential logical dependencies between knowledge nodes.

[0199] After receiving the enhanced core knowledge subgraph output by step S6, the node set contains enhanced knowledge nodes that have undergone knowledge distillation, semantic refinement, and fusion of cross-domain knowledge, and the edge set contains updated semantic association weights and logical relationships.

[0200] A topological structure analysis method (parameters: node set, edge set, edge weight) is used to realize the topological structure modeling of the enhanced core knowledge subgraph and generate the adjacency matrix and degree matrix representing the connectivity between nodes.

[0201] Furthermore, by using a normalization method (parameters: adjacency matrix, degree matrix), the graph structure is preprocessed to standardize before being input into the graph neural network, and a standardized adjacency matrix is ​​obtained that can be used for graph embedding operations.

[0202] Furthermore, the graph convolutional network (GCN) algorithm (parameters: normalized adjacency matrix, node feature vector matrix, number of convolutional layers, activation function type) is used to achieve multi-layer information aggregation of node features in the topological neighborhood and generate intermediate embedding representation matrices for each node.

[0203] Furthermore, the graph attention network (GAT) algorithm (parameters: intermediate embedding representation matrix, attention coefficient calculation function LeakyReLU, number of attention heads) is used to achieve semantic relevance weighted aggregation based on the weights between neighboring nodes, and the final embedding representation vector of each node is obtained.

[0204] By using the node embedding representation analysis method, the embedding vectors from the previous step are converted into embedding feature data that represent the potential logical dependencies of knowledge nodes, thus providing support for S7.2 path priority calculation and main path skeleton construction.

[0205] S7.2: Based on the nurse's competency profile and learning objectives, calculate the priority weight of the target knowledge nodes, and use Dijkstra's shortest path algorithm to determine the basic learning path within the main knowledge system as the initial skeleton for path planning.

[0206] S7.3: Perform semantic expansion analysis on the enhanced knowledge nodes on the main path. Based on the semantic vector similarity generated by knowledge distillation, identify and insert highly relevant cross-domain knowledge nodes to enhance the knowledge breadth of the path.

[0207] S7.4: Dynamically evaluate the path structure after inserting cross-domain nodes, and adjust the order of path nodes based on the learning duration prediction model and the knowledge absorption rate model to optimize the learning pace and cognitive load distribution.

[0208] S7.5: The optimized path structure is serialized and encoded to generate an executable recommended path sequence, and the learning resource links and learning objective descriptions of the path nodes are encapsulated to form a personalized growth path recommendation result for users.

[0209] A serialization encoding method (parameters: unique node identifier, logical path order, and dependency weights between nodes) is used to linearize the optimized path structure, transforming the complex graph path structure into an ordered sequence index set for execution and storage.

[0210] Furthermore, by encapsulating the path node attribute set algorithm (parameters: node knowledge domain label, cross-domain fusion status bit, learning priority coefficient), the attribute information of each serialized node is structurally mapped, and the path node attribute matrix is ​​obtained for subsequent resource binding and target description generation.

[0211] Furthermore, based on the resource link generation mechanism (parameters: unique node identifier, resource library index table, access protocol parameters), an interface call with the learning resource management module is implemented, binding the node attribute matrix with the corresponding teaching resource URL one by one, and generating a path resource link set to ensure that the learning content of the node can be directly accessed.

[0212] Furthermore, by using a learning objective description generation algorithm (parameters: node knowledge point semantic vector, target ability profile, output document template), the ability improvement goals, application scenarios, and assessment standards corresponding to the nodes are transformed into structured explanatory text, and a set of learning objective descriptions is generated, so that users can clearly understand the learning expectations of each node when executing the path.

[0213] By combining path serialization encoding with attribute resource binding, the path structure optimization results from the previous step are transformed into executable recommended path sequences and node resource target mappings, achieving a highly available and personalized growth path recommendation effect for nurse users.

[0214] For example, in the emergency nurse career development path recommendation scenario, the optimized path contains 12 knowledge nodes. Each node is uniquely identified using UUID encoding. The logical order of the path is determined by the shortest path index output by the Dijkstra algorithm, and the dependency weights are set based on node embedding similarity values ​​calculated using a graph neural network. The node attribute set includes knowledge domain tags (emergency care, psychological intervention, disaster medicine, etc.), cross-domain fusion status bits (core / fusion), and learning priority coefficients (normalized based on competency profiles, ranging from 0.1 to 1.0). These attributes are mapped to a 12×5 matrix. During resource binding, the learning resource management module API is called to quickly retrieve matching teaching resource URLs from the resource library index table, forming 12 resource links. The access protocol parameters use HTTPS and token authentication. The learning objective description generation algorithm uses BERT to vectorize node semantics, then performs cosine similarity calculation with the current nurse's competency profile, selecting the target competency description template with the highest matching degree to generate structured description text containing improvement goals, recommended application scenarios, and assessment standards. Each description has an average length of approximately 300 words. The final output recommended path sequence is a JSON structure containing node UUIDs, logical order, attribute tags, resource links, and learning objective descriptions. Nurses can directly access the resource learning by clicking on the system interface, achieving immediate executability of the path and a significant improvement in expected ability enhancement.

[0215] Step S8: After the nurses complete the learning, their submitted application cases are collected, and the degree of cross-domain knowledge application reflected in the cases is compared using natural language processing technology to generate feedback signals. Specifically, this includes: S8.1: Preprocess the application case texts submitted by nurses after completing the learning process, including word segmentation, stop word removal, and word form restoration, to generate standardized text corpus as input for subsequent semantic analysis.

[0216] S8.2: Based on the pre-trained cross-domain BERT model, the standardized text corpus is semantically encoded to extract cross-domain knowledge semantic feature vectors such as emergency care, psychological intervention, and doctor-patient communication involved in the text, forming a cross-domain knowledge representation matrix.

[0217] S8.3: Perform cosine similarity calculation on the semantic vectors of the corresponding knowledge nodes in the cross-domain knowledge representation matrix and the core knowledge subgraph to identify the cross-domain knowledge points actually applied in the case and their correlation strength, and generate a knowledge point application popularity distribution map.

[0218] S8.4: Based on the knowledge point application popularity distribution map, compare it with the cross-domain knowledge nodes in the original recommendation path, calculate the actual application coverage and semantic matching degree of the recommended knowledge points, so as to evaluate the practicality and effectiveness of cross-domain knowledge recommendation.

[0219] S8.5: Generate a feedback signal based on the coverage and matching degree metrics. The feedback signal includes suggestions for adjusting the parameters of the knowledge distillation model and optimization directions for the semantic density evaluation threshold, which are used to drive the adaptive updating of parameters and optimization of recommendation strategies in subsequent path recommendation models.

[0220] Step S9: Based on the feedback signal, dynamically adjust the parameters of the knowledge distillation model and the semantic density evaluation threshold to continuously optimize the quality of subsequent recommendations. Specifically, this includes: S8.1: Preprocess the application case texts submitted by nurses after completing the learning process, including word segmentation, stop word removal, part-of-speech tagging, and entity recognition, to extract structured semantic units as input data for subsequent knowledge application evaluation.

[0221] S8.2: Based on the BERT semantic model, the extracted structured semantic units are vectorized to generate high-dimensional semantic feature vectors of the case text, which serve as semantic input for evaluating the degree of cross-domain knowledge application.

[0222] S8.3: Perform cosine similarity calculation between the semantic feature vector of the case and the semantic vector of the corresponding enhanced knowledge node in the core knowledge subgraph to obtain the matching score of each knowledge point in actual application, so as to evaluate the application effect of cross-domain knowledge in clinical practice.

[0223] S8.4: Based on the comparison results between the matching score and the preset threshold, identify cross-domain knowledge points that have not been effectively mastered or applied, and generate a list of weak knowledge mastery points as the core content of the feedback signal.

[0224] S8.5: Based on the list of weak knowledge mastery, the attention weight parameters in the knowledge distillation model are updated with gradients, and the feature preservation strategy in the cross-domain knowledge semantic extraction process is optimized to improve the practicality and transferability of subsequent recommended content.

[0225] S8.6: By combining the distribution characteristics of knowledge gaps with historical data of semantic density assessment, dynamically adjust the information entropy weight coefficient and relevance threshold in the semantic density assessment module to enhance the system's accuracy in recognizing redundant content.

[0226] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0227] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0228] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent assessment and training method for emergency nurses based on a multi-dimensional growth path, specifically including: S1: Based on the emergency nursing professional knowledge base, construct a core knowledge subgraph. The core knowledge subgraph includes knowledge points on emergency procedures, critical care, and emergency drug use, as well as their logical relationships, serving as the main knowledge system for personalized path planning. S2: Extract auxiliary knowledge points from fields related to psychological intervention, doctor-patient communication, and disaster medicine, construct an independent cross-domain auxiliary knowledge subgraph, and establish a correlation mapping between it and the core knowledge subgraph; S3: Based on the nurse's current competency profile, determine the main learning path in the core knowledge subgraph and identify the target nodes that need to be introduced into cross-domain knowledge as triggering conditions for cross-domain knowledge fusion. S4: Perform text semantic analysis on candidate cross-domain knowledge points, calculate their semantic relevance score with the current main path target node, and combine the information entropy value of the knowledge point in the original domain to comprehensively evaluate the information density per unit content. S5: Based on the information density assessment results, filter and mark cross-domain knowledge points with high redundancy, retain high-density knowledge points, and output the set of knowledge points to be distilled; S6: Input the knowledge points to be distilled into the knowledge distillation engine, extract their core semantics into lightweight semantic vectors, and embed them into the corresponding positions in the core knowledge subgraph to form enhanced knowledge nodes; S7: Based on the enhanced core knowledge subgraph from step S6, perform path planning to generate personalized growth path recommendation results containing necessary cross-domain knowledge.

2. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 1, characterized in that, Step S7 is followed by: S8: After the nurses complete the learning, collect the application cases they submit, and compare the degree of cross-domain knowledge application reflected in the cases through natural language processing technology to generate feedback signals. S9: Based on the feedback signal, dynamically adjust the knowledge distillation model parameters and semantic density evaluation threshold to continuously optimize the quality of subsequent recommendations.

3. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 1, characterized in that, Step S2 specifically includes: Based on professional literature and teaching resources in non-core areas such as psychological intervention, doctor-patient communication, and disaster medicine, a cross-domain knowledge corpus is constructed to extract auxiliary knowledge points that meet the competency development needs of emergency nurses, which serve as the original input for constructing a cross-domain auxiliary knowledge subgraph. Entity recognition and relation extraction are performed on the text content in the cross-domain knowledge corpus to identify key knowledge entities. Semantic relationships between entities are extracted by combining dependency parsing to construct a structured set of cross-domain knowledge triples. Based on the cross-domain knowledge triple set, a cross-domain auxiliary knowledge subgraph is constructed, which includes auxiliary knowledge point nodes and semantic association edges between them. Semantic alignment is performed on the key knowledge point nodes in the core knowledge subgraph and the auxiliary knowledge point nodes in the cross-domain auxiliary knowledge subgraph. The semantic similarity between nodes is calculated based on the cross-modal semantic embedding model to identify knowledge node pairs with potential collaborative relationships. Based on the cross-graph semantic similarity results, a weak association mapping matrix is ​​constructed.

4. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 3, characterized in that, A graph neural network model is used to construct a cross-domain auxiliary knowledge subgraph.

5. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 1, characterized in that, The competency profile data in S3 includes multi-dimensional competency indicators such as nurses' mastery of basic nursing knowledge, clinical emergency response capabilities, and communication and psychological intervention abilities.

6. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 5, characterized in that, The identification of target nodes that require the introduction of cross-domain knowledge specifically includes: performing knowledge gap analysis on each target node in the main learning path sequence, identifying the gap between the current nurse's ability status and the ability required by the target node, generating a node-level knowledge gap vector, and determining whether to introduce cross-domain knowledge based on the node-level knowledge gap vector and a preset knowledge fusion trigger threshold.

7. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 1, characterized in that, Step S4 specifically includes: The text content of candidate cross-domain knowledge points is preprocessed, including word segmentation, stop word removal and stemming, to generate standardized text semantic representation vectors; The standardized text semantic representation vector is embedded to obtain a high-dimensional semantic feature vector of cross-domain knowledge points; The semantic feature vector of the current main path target node is extracted and cosine similarity is calculated with the high-dimensional semantic feature vector of the cross-domain knowledge point to obtain the semantic relevance score between the two. The cross-domain knowledge point is modeled in the context distribution of its respective auxiliary knowledge subgraph based on the information entropy calculation model, so as to obtain the information entropy value of the knowledge point, which reflects its knowledge density in the original domain. The semantic relevance score and information entropy value are weighted and fused to generate a comprehensive information density evaluation index for cross-domain knowledge points.

8. The intelligent examination and training method for emergency nurses based on a multi-dimensional growth path according to claim 7, characterized in that, Step S5 specifically includes: The semantic density evaluation results of cross-domain knowledge points are normalized to eliminate evaluation bias between different knowledge sources and obtain a density score vector with unified dimensions. Based on the preset semantic density threshold range, the normalized density score vector is binarized and labeled, and cross-domain knowledge points below the threshold are labeled as redundant knowledge points, forming a set of redundant knowledge point labels. Semantic association tracing is performed on the knowledge points in the redundant knowledge point tag set to identify their original semantic nodes in the cross-domain auxiliary knowledge subgraph, and their potential synergistic impact on the main knowledge system is evaluated based on the node importance index to generate a potential retention candidate set. Based on the weighted fusion value of the synergistic impact assessment results and density scores, the knowledge points in the potential retention candidate set are screened a second time. Knowledge points with synergistic impact below the dynamic threshold are removed, and high-density knowledge points with semantic fusion potential are retained to form a preliminary set of knowledge points to be distilled. Perform a knowledge consistency check on the initial set of knowledge points to be distilled, check whether there are semantic conflicts or logical contradictions in the main knowledge system, and output the final set of knowledge points to be distilled after semantic verification.

9. An intelligent examination and training system for emergency nurses based on a multidimensional growth path, which uses the intelligent examination and training method for emergency nurses based on a multidimensional growth path as described in any one of claims 1-8 to examine and train nurses' growth.

Images