Dynamic cerebral apoplexy knowledge graph intelligent generation and maintenance system and method

Abstract

The invention discloses an intelligent generation and maintenance system and method for a dynamic cerebral apoplexy knowledge graph, and relates to the field of intelligent medical treatment, and the system comprises a data access and preprocessing module which is used for accessing and retrieving cerebral apoplexy related information from a multi-source data source; the knowledge extraction module based on LLM is used for acquiring candidate knowledge fragment streams; the knowledge graph structuring and filling module is used for integrating the candidate knowledge fragment flow into a formal knowledge graph structure stored in a graph database; the LLM-driven verification and refinement module is used for detecting the candidate knowledge fragment stream and the formal knowledge graph structure to obtain a detection result; and the dynamic updating and maintaining module is used for realizing dynamic fusion and version control of the formal knowledge graph structure. Through modular architecture design, a high-performance knowledge extraction and verification mechanism of a large language model and a dynamic processing capability oriented to continuous updating are fused, and high automation, standardization and intelligentization of a medical knowledge graph construction process are realized.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of smart healthcare, and more particularly, the present application relates to a dynamic stroke knowledge graph intelligent generation and maintenance system and method. BACKGROUND

[0002] Stroke is one of the leading causes of death and disability worldwide, with complex etiology, diverse diagnostic processes, and constantly updated treatment plans. The management process relies on the integration of a large amount of information from multiple sources of medical data such as literature, guidelines, electronic health records (EHR), and image reports. To effectively organize and utilize this information, knowledge graphs have become a key tool, which assists clinical decision-making and scientific research analysis by structuring relevant entities (such as subtypes, symptoms, risk factors, drugs, treatment methods, genes, etc.) and their relationships (such as "treatment", "cause", "diagnostic basis") in stroke.

[0003] However, the construction and maintenance of current stroke knowledge graphs still face many challenges: traditional manual construction is inefficient and difficult to expand; existing NLP (Natural Language Processing) / machine learning methods rely on a large amount of labeled data and are difficult to accurately extract complex relationships in medical language; the graph content is slow to update and lacks timeliness, and it is still difficult to extract knowledge uniformly from multiple heterogeneous data sources; in addition, automatic verification of extraction results also has bottlenecks.

[0004] In view of the above problems, there is an urgent need for a more efficient, scalable, and dynamically updated dynamic stroke knowledge graph intelligent generation and maintenance system and method. SUMMARY

[0005] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiments section. The summary section of the present application does not mean to attempt to limit the key features and necessary technical features of the claimed technical solutions, nor does it attempt to determine the protection scope of the claimed technical solutions.

[0006] In a first aspect, the present application proposes a dynamic stroke knowledge graph intelligent generation and maintenance system, comprising:

[0007] A data access and preprocessing module for accessing and retrieving stroke-related information from multiple data sources;

[0008] a LLM-based knowledge extraction module configured to perform named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information via a large language model to obtain a candidate knowledge snippet stream, wherein the candidate knowledge snippet stream comprises entities, relations, attributes, relevant source information, and initial confidence scores;

[0009] a knowledge graph structuring and populating module configured to integrate the candidate knowledge snippet stream into a formal knowledge graph structure stored in a graph database;

[0010] an LLM-driven verification and refinement module configured to detect the candidate knowledge snippet stream and the formal knowledge graph structure to obtain a detection result;

[0011] a dynamic updating and maintaining module configured to periodically monitor the multi-source data sources accessed by the data access and preprocessing module, identify newly published or updated content related to stroke, and trigger incremental processing procedures of the knowledge extraction module, the knowledge graph structuring and populating module, and the LLM-driven verification and refinement module to realize dynamic fusion and version control of the formal knowledge graph structure.

[0012] In a feasible implementation, the system further comprises:

[0013] a human-machine collaboration module configured to, in a case where the detection result includes low-confidence information and conflicting information, display the detection result to a domain expert to obtain feedback information of the domain expert, so as to optimize the formal knowledge graph structure.

[0014] In a second aspect, the present application further provides a dynamic stroke knowledge graph intelligent generation and maintenance method, which is used for the dynamic stroke knowledge graph intelligent generation and maintenance system of the first aspect, and comprises:

[0015] accessing and retrieving stroke-related information from multi-source data sources via the data access and preprocessing module;

[0016] performing named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information via a large language model by using the LLM-based knowledge extraction module to obtain a candidate knowledge snippet stream, wherein the candidate knowledge snippet stream comprises entities, relations, attributes, relevant source information, and initial confidence scores;

[0017] integrating the candidate knowledge snippet stream into a formal knowledge graph structure stored in a graph database via the knowledge graph structuring and populating module;

[0018] detecting the candidate knowledge snippet stream and the formal knowledge graph structure via the LLM-driven verification and refinement module to obtain a detection result;

[0019] The dynamic updating and maintaining module periodically monitors the multi-source data sources accessed by the data accessing and preprocessing module, identifies newly published or updated content related to stroke, and triggers the incremental processing procedures of the knowledge extraction module, the knowledge graph structuring and filling module, and the LLM-driven verification and refinement module, to realize dynamic fusion and version control of the formal knowledge graph structure.

[0020] In a feasible implementation, the knowledge extraction module based on the LLM utilizes a large language model to perform named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information to obtain a candidate knowledge fragment stream, including:

[0021] Multi-dimensional feature analysis is performed on the stroke-related information, and a prompt demand score and a context dependency score are calculated based on the specified task type;

[0022] Based on the demand score, the context dependency score, the text type information, and the specified task type, the most suitable prompt template ID and context extraction strategy are automatically selected through an internal rule engine query;

[0023] According to the context extraction strategy, the corresponding context information is extracted from the stroke-related information;

[0024] The dynamic prompt information is filled into the placeholders of the template according to the prompt template ID, the context information, and the stroke-related information to determine the dynamic prompt information;

[0025] The dynamic prompt information and the corresponding context information are subjected to named entity recognition, relation extraction, and attribute extraction tasks based on the target LLM model to obtain a candidate knowledge fragment stream.

[0026] In a feasible implementation, the target LLM model includes a zero-shot LLM model, a few-shot LLM model, and a fine-tuned LLM model.

[0027] In a feasible implementation, the multi-dimensional feature analysis includes text type recognition, structural complexity evaluation, information density evaluation, and task characteristic analysis,

[0028] The multi-dimensional feature analysis of the stroke-related information and the calculation of the prompt demand score and the context dependency score based on the specified task type include:

[0029] The stroke-related information is subjected to text type recognition, structural complexity evaluation, information density evaluation, and task characteristic analysis, and the prompt demand score and the context dependency score are calculated based on the specified task type.

[0030] In an implementation, the detection operations in the detection of the candidate knowledge fragment stream and the formal knowledge graph structure by the LLM-driven verification and refinement module include consistency detection, fact detection, confidence score detection, and redundancy detection.

[0031] In an implementation, the method further includes:

[0032] In the case where the detection results include low-confidence information and conflicting information, the detection results are presented to a domain expert to obtain feedback information of the domain expert;

[0033] Based on the feedback information of the domain expert, the formal knowledge graph structure is optimized or the LLM and prompt information used by the extraction and verification module are refined.

[0034] In an implementation, the integration of the candidate knowledge fragment stream into the formal knowledge graph structure stored in the graph database by the knowledge graph structuring and filling module includes:

[0035] The candidate knowledge fragment stream is integrated into a triple form conforming to a predefined medical ontology structure by the knowledge graph structuring and filling module, and entity standardization and disambiguation operations are performed to obtain a structured result, where the entity standardization includes linking the entity to a corresponding concept entry in a standard medical terminology system;

[0036] The structured result is stored in the graph database to construct the formal knowledge graph.

[0037] In an implementation, the knowledge graph structure includes entity nodes, relationship edges of semantic relationships between entities, attribute sets associated with the entity nodes or relationship edges, and mapping link information between entities or relationships and a standard medical ontology system.

[0038] In summary, the present embodiment realizes high automation, standardization, and intelligentization of the medical knowledge graph construction process through modular architecture design, high-performance knowledge extraction and verification mechanism based on large language models, and dynamic processing capability for continuous updating. The system can significantly improve the accuracy, timeliness, and practicality of the stroke knowledge graph, providing a solid and reliable knowledge base for clinical decision support, scientific research discovery, and medical AI systems, and has significant technical advantages and broad application prospects in the field of medical artificial intelligence and knowledge engineering.

[0039] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following description, and will be understood by those skilled in the art upon reading and understanding the specification. BRIEF DESCRIPTION OF DRAWINGS

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0041] Figure 1 A schematic diagram illustrating the principle of a dynamic stroke knowledge graph intelligent generation and maintenance system provided in this application embodiment;

[0042] Figure 2 This is a flowchart illustrating a method for intelligent generation and maintenance of a dynamic stroke knowledge graph, as provided in an embodiment of this application.

[0043] Figure 3 This is a schematic diagram illustrating the principle of a method for intelligent generation and maintenance of a dynamic stroke knowledge graph, as provided in an embodiment of this application.

[0044] Figure 4 A schematic diagram illustrating the internal workflow of the verification and refinement module driven by LLM in a dynamic stroke knowledge graph intelligent generation and maintenance system provided in this application embodiment; Detailed Implementation

[0045] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0046] Please see Figure 1 This is a schematic diagram illustrating the principle of a dynamic stroke knowledge graph intelligent generation and maintenance system provided in this application embodiment, which may specifically include:

[0047] Firstly, this application proposes a dynamic intelligent generation and maintenance system for stroke knowledge graphs, including:

[0048] A data access and preprocessing module 10 is configured to access and retrieve stroke-related information from multiple heterogeneous data sources;

[0049] An LLM-based knowledge extraction module 20 is configured to perform named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information by using a large language model (LLM) to obtain a candidate knowledge fragment stream, wherein the candidate knowledge fragment stream includes entities, relations, attributes, related source information, and initial confidence scores.

[0050] A knowledge graph structuring and filling module 30 is configured to integrate the candidate knowledge fragment stream into a formal knowledge graph structure stored in a graph database.

[0051] An LLM-driven verification and refinement module 40 is configured to detect the candidate knowledge fragment stream and the formal knowledge graph structure to obtain a detection result.

[0052] A dynamic updating and maintenance module 50 is configured to periodically monitor the multiple heterogeneous data sources accessed by the data access and preprocessing module, identify newly published or updated content related to stroke, and trigger incremental processing procedures of the knowledge extraction module, the knowledge graph structuring and filling module, and the LLM-driven verification and refinement module to realize dynamic fusion and version control of the formal knowledge graph structure.

[0053] For example, the present application mainly includes five core functional modules, each of which works together to automatically extract, structure, verify the quality of, and dynamically update medical knowledge in the field of stroke. The functions of each module and their interactions are explained in detail below, and the overall operation logic of the system is described in continuous language:

[0054] The data access and preprocessing module is configured to access and retrieve stroke-related information from multiple heterogeneous data sources, including but not limited to medical literature databases (such as PubMed, Embase), clinical guidelines, electronic health records (EHR), medical textbooks, structured drug databases (such as DrugBank), and clinical trial registries (such as ClinicalTrials.gov). After obtaining the data through automated interfaces (API calls, crawlers, PDF parsing, etc.), the module performs preprocessing procedures including format standardization, noise text removal, sentence segmentation, and preliminary semantic filtering to generate well-structured intermediate text corpus with processing value.

[0055] Subsequently, these pre-processed stroke-related texts will be input into the LLM-based knowledge extraction module. This module utilizes a large language model (LLM) that has been prompted and optimized or fine-tuned, to perform advanced natural language processing tasks for medical scenarios. Specifically, this includes: named entity recognition (NER), which identifies entities such as stroke subtypes, symptoms, treatment drugs, diagnostic tests, genetic variations, etc.; relation extraction (RE), which identifies semantic relationships between entities, such as "treatment," "complication," "diagnostic basis," etc.; and attribute extraction, which obtains specific information associated with entities or relationships (such as drug dosage, action time limit, occurrence probability, etc.). The extraction results are uniformly organized into a stream of candidate knowledge fragments, each of which contains entities, relationships, attributes, and their corresponding data source information and initial confidence scores assigned by the LLM scoring mechanism.

[0056] The knowledge graph structuring and filling module receives this stream of candidate knowledge fragments and is responsible for integrating them into the formal stroke knowledge graph structure. Based on predefined medical ontologies, this module standardizes entities, including name normalization, semantic disambiguation, and linking to medical terminology libraries (such as SNOMED CT, ICD-10, RxNorm, etc.); and converts the extraction results into structured triple (subject, predicate, object) format, with additional metadata information including timestamps, source IDs, and confidence scores. Finally, these structured information is written into a graph database, forming a formal knowledge graph structure that can be used for reasoning and querying.

[0057] To ensure that the constructed knowledge graph has logical consistency and factual accuracy, an LLM-driven verification and refinement module is introduced. This module performs automatic quality detection on the stream of candidate knowledge fragments and the filled formal graph, including the following functions: first, logical consistency check, to identify mutually exclusive information or contradictory statements in knowledge; second, fact checking, by prompting the LLM to consult background knowledge or cross-check sources to confirm the credibility of the extracted content; third, confidence scoring mechanism, to assign a credibility score to the triple based on content certainty, source authority, and cross-consistency; fourth, redundancy detection and integration, to identify and merge duplicate or similar knowledge items. Through this module, the system can continuously optimize the quality and stability of the knowledge graph structure.

[0058] The dynamic updating and maintaining module periodically monitors various medical data sources connected to the data access and preprocessing module to identify newly added or revised information about stroke. Once an update is found, the module automatically triggers the processing flow of the knowledge extraction module, the structured and filling module, and the verification and refining module. This processing method not only avoids repeated construction of the whole graph, but also efficiently integrates the newly added or changed medical knowledge into the knowledge graph, and retains the historical version for comparison and backtracking, thereby realizing the dynamic fusion and version control capabilities of the knowledge graph.

[0059] In summary, compared with the traditional method which relies entirely on manual construction, the present embodiment automatically integrates multi-source heterogeneous medical data through the data access and preprocessing module, and combines automatic interface and standardized cleaning process to greatly reduce manual intervention and improve the efficiency and consistency of data collection. The system supports parallel access to multiple types of information sources such as literature databases (e.g. PubMed), guidelines, EHR, clinical trial databases, etc., making the data basis of the knowledge graph more comprehensive. Second, compared with the problem of insufficient medical terminology expression ability and cumbersome updating of traditional NLP / machine learning models, the present embodiment innovatively introduces a large language model (LLM) based on prompt optimization or fine-tuning for knowledge extraction. Through tasks such as named entity recognition (NER), relation extraction (RE), and attribute extraction, the system can accurately identify complex medical terminology, implicit semantic relationships, and their related attributes in the field of stroke, thereby constructing more detailed and semantically rich knowledge fragments. At the same time, by constructing a candidate knowledge fragment stream containing confidence scores, subsequent quality assessment becomes more operational. Third, compared with the limitations of traditional knowledge graph structure being static and difficult to adapt to new knowledge updates, the present embodiment provides a dynamic evolution-oriented graph structuring and filling module, which combines medical ontology for semantic disambiguation and standardized entity linking (e.g. to SNOMED CT, ICD-10, etc.), not only ensuring the consistency and scalability of the graph structure, but also laying a unified semantic foundation for subsequent multi-graph fusion and cross-platform applications. In addition, the present embodiment introduces an LLM-driven verification and refinement module, breaking through the bottleneck of relying on manual verification of automatically extracted results. This module can perform consistency detection, fact checking, confidence scoring, and redundancy merging on candidate knowledge and graph content, improving the accuracy, logic, and stability of the overall graph, and significantly reducing the risk of spreading false knowledge. This module also provides technical support for realizing continuous optimization and closed-loop feedback of graph quality. Finally, unlike the traditional way of updating the knowledge graph with a long cycle and manual-driven process, the present embodiment provides a dynamic updating and maintenance module that can automatically monitor content changes in external data sources, trigger incremental knowledge extraction, structuring, and verification processes, and implement version management and historical rollback of the knowledge graph. This not only ensures the timeliness of the graph content, but also enhances the system's ability to quickly respond to events such as changes in clinical guidelines and updates to drug indications. In summary, the present embodiment realizes the high automation, standardization, and intelligence of the medical knowledge graph construction process through modular architecture design, high-performance knowledge extraction and verification mechanisms based on large language models, and dynamic processing capabilities for continuous updating. The system can greatly improve the accuracy, timeliness, and practicality of the stroke knowledge graph, providing a solid and reliable knowledge foundation for clinical decision support, scientific research discovery, and medical AI systems, and has significant technical advantages and broad application prospects in the fields of medical artificial intelligence and knowledge engineering.

[0060] In an implementation, the system further comprises:

[0061] The human-computer collaboration module 60 is used to display the detection results to a domain expert to obtain feedback information of the domain expert to optimize the formal knowledge graph structure in the case that the detection results include low confidence information and conflict information.

[0062] For example, to further improve the accuracy and reliability of the knowledge graph, the system further includes a human-computer collaboration module. The module is mainly used to process low-confidence information or conflicting knowledge fragments marked by the system in the automatic knowledge extraction and verification process. By introducing the participation of human experts, manual supplement, correction and quality control of the graph content are realized, so that the formal knowledge graph structure generated finally is optimized.

[0063] Specifically, after the LLM-driven verification and refinement module completes automatic detection, if it is found that the confidence score of some candidate knowledge fragments is low (such as inconsistent sources, ambiguous language expression, weak information certainty) or there is obvious logical conflict (such as a drug being labeled as both treatment and contraindication), the system will pass these fragments together with related context information (such as original text fragments, data sources, extraction logic, etc.) to the human-computer collaboration module.

[0064] The human-computer collaboration module provides an interactive interface for domain experts (such as neurologists, evidence-based medicine researchers, drug evaluation experts, etc.) to manually review suspicious information. Based on their professional knowledge and clinical experience, experts can judge the validity and rationality of the information and give corresponding feedback opinions, such as confirmation, negation, suggestion for modification or alternative representation, etc. After receiving the feedback information, the system can automatically update the review results to the formal knowledge graph structure, and record the audit trail of the expert's operation for subsequent version management and knowledge traceability analysis.

[0065] In addition, expert feedback information can also be used to further optimize the knowledge extraction model or prompt template in the system. For example, when a certain type of information is frequently modified or negated by experts, the system can automatically adjust the task weight, context strategy or fine-tuning examples in the MDPAC (Multi-Dimensional Prompt and Adaptive Context algorithm) algorithm based on these feedbacks to improve the performance of the corresponding task when processing similar texts in the future.

[0066] In summary, the human-computer collaboration module not only makes up for the possible misjudgment of the completely automatic processing flow in the face of complex medical semantics, but also builds a continuous learning and iterative optimization mechanism. Through the closed-loop interaction between experts and the system, the reliability and usability of the system output knowledge graph are further enhanced, which is especially suitable for the auxiliary diagnosis and clinical decision-making scenarios with extremely high requirements for medical knowledge accuracy.

[0067] In a second aspect, the present application also provides a dynamic stroke knowledge graph intelligent generation and maintenance method, which is used for the dynamic stroke knowledge graph intelligent generation and maintenance system of the first aspect, as shown in Figure 2 and Figure 3 , a flowchart of a dynamic stroke knowledge graph intelligent generation and maintenance method provided by an embodiment of the present application is shown in Figure 2 , and a principle diagram of a dynamic stroke knowledge graph intelligent generation and maintenance method provided by an embodiment of the present application is shown in Figure 3 The method comprises the following steps:

[0068] S210, accessing and retrieving stroke-related information from multiple data sources through the data access and preprocessing module described above;

[0069] S220, using the knowledge extraction module based on LLM to perform named entity recognition, relationship extraction and attribute extraction tasks on the stroke-related information through a large language model to obtain a candidate knowledge fragment stream, wherein the candidate knowledge fragment stream includes entities, relationships, attributes, related source information and initial confidence scores;

[0070] S230, integrating the candidate knowledge fragment stream into the formal knowledge graph structure stored in the graph database through the knowledge graph structuring and filling module;

[0071] S240, detecting the candidate knowledge fragment stream and the formal knowledge graph structure through the LLM-driven verification and refinement module to obtain a detection result;

[0072] S250, periodically monitoring the multiple data sources accessed by the data access and preprocessing module through the dynamic updating and maintaining module, identifying newly published or updated content about stroke, and triggering the incremental processing flow of the knowledge extraction module, the knowledge graph structuring and filling module, and the LLM-driven verification and refinement module to realize the dynamic fusion and version control of the formal knowledge graph structure.

[0073] Exemplarily, the data access and preprocessing module is used to access and retrieve stroke-related information from multiple heterogeneous data sources, including but not limited to medical literature databases (such as PubMed, Embase), clinical guidelines, electronic health records (EHR), medical textbooks, structured drug databases (such as DrugBank), and clinical trial registries (such as ClinicalTrials.gov). After the module obtains data through automated interfaces (API calls, crawlers, PDF parsing, etc.), it performs a preprocessing process including format standardization, noise text removal, sentence segmentation, and preliminary semantic screening to generate well-structured and valuable intermediate text corpus.

[0074] Subsequently, these preprocessed stroke-related texts will be input into the LLM-based knowledge extraction module. This module uses a large language model (LLM) that has been prompted, optimized, or fine-tuned to perform advanced natural language processing tasks for medical scenarios. Specifically, it includes: named entity recognition (NER) for identifying stroke subtypes, symptoms, treatment drugs, diagnostic tests, genetic variations, etc.; relationship extraction (RE) for identifying semantic relationships between entities, such as "treatment", "complication", "diagnostic basis", etc.; and attribute extraction for obtaining specific information associated with entities or relationships (such as drug dosage, action time limit, occurrence probability, etc.). The extraction results are organized into a candidate knowledge fragment stream, where each fragment contains entities, relationships, attributes, and their corresponding data source information and initial confidence scores assigned by the LLM scoring mechanism.

[0075] The knowledge graph structuring and filling module receives these candidate knowledge fragment streams and is responsible for integrating them into the formal stroke knowledge graph structure. Based on predefined medical ontology, this module performs standardization processing on entities, including name normalization, semantic disambiguation, and linking to medical terminology libraries (such as SNOMED CT, ICD-10, RxNorm, etc.); and converts the extraction results into structured triple (subject, predicate, object) format, with additional metadata information including timestamp, source ID, confidence, etc. Finally, these structured information is written into a graph database to form a formal knowledge graph structure that can be used for reasoning and querying.

[0076] To ensure the constructed knowledge graph has logical consistency and factual accuracy, an LLM-driven verification and refinement module is introduced. This module conducts automatic quality detection on the candidate knowledge fragment stream and the filled formal graph, mainly including the following functions: first, logical consistency check, used to identify mutually exclusive information or contradictory statements in knowledge; second, fact checking, by prompting LLM to consult background knowledge or cross-compare sources to confirm the credibility of the extracted content; third, confidence scoring mechanism, which gives the triple a credibility score determined by content certainty, source authority and cross-consistency; fourth, redundancy detection and integration, identifying and fusing repeated or similar knowledge items. Through this module, the system can continuously optimize the quality and stability of the knowledge graph structure.

[0077] The dynamic updating and maintenance module regularly monitors various medical data sources connected by the data access and preprocessing module to identify new or revised information about stroke. Once an update is found, the module automatically triggers the above-mentioned knowledge extraction module, structured and filling module, verification and refinement module processing flow. This processing method not only avoids repeated construction of the whole graph, but also efficiently integrates the new or changed medical knowledge into the knowledge graph, and retains the historical version for comparison and backtracking, thereby realizing the dynamic fusion and version control ability of the knowledge graph.

[0078] In summary, compared with the traditional method which relies entirely on manual construction, the present embodiment can greatly reduce manual intervention and improve the efficiency and consistency of data collection by automatically integrating multi-source heterogeneous medical data through the data access and preprocessing module, combined with automatic interface and standardized cleaning process. The system supports parallel access to multiple types of information sources such as literature databases (e.g. PubMed), guidelines, EHR, clinical trial databases, etc., making the data basis of the knowledge graph more comprehensive. Second, compared with the problem of insufficient medical terminology expression ability and cumbersome updating of traditional NLP / machine learning models, the present embodiment innovatively introduces a large language model (LLM) based on prompt optimization or fine-tuning for knowledge extraction. Through tasks such as named entity recognition (NER), relation extraction (RE), and attribute extraction, the system can accurately identify complex medical terminology, implicit semantic relationships, and their related attributes in the field of stroke, thereby constructing more detailed and semantically rich knowledge fragments. At the same time, by constructing a candidate knowledge fragment stream containing confidence scores, subsequent quality assessment becomes more operational. Third, compared with the limitations of traditional knowledge graph structure being static and difficult to adapt to new knowledge updates, the present embodiment provides a dynamic evolution-oriented graph structuring and filling module, combined with medical ontology for semantic disambiguation and standardized entity linking (e.g. to SNOMED CT, ICD-10, etc.), which not only guarantees the consistency and scalability of the graph structure, but also lays a unified semantic foundation for subsequent multi-graph fusion and cross-platform applications. In addition, the present embodiment introduces an LLM-driven verification and refinement module, breaking through the bottleneck of relying on manual verification of automatically extracted results. This module can perform consistency detection, fact checking, confidence scoring, and redundancy merging on candidate knowledge and graph content, improving the accuracy, logic, and stability of the overall graph, and significantly reducing the risk of spreading false knowledge. This module also provides technical support for realizing continuous optimization and closed-loop feedback of graph quality. Finally, unlike the traditional way of updating the knowledge graph with a long cycle and manual-driven process, the present embodiment provides a dynamic updating and maintenance module that can automatically monitor content changes in external data sources, trigger incremental knowledge extraction, structuring, and verification processes, and implement version management and historical rollback of the knowledge graph. This not only ensures the timeliness of the graph content, but also enhances the system's ability to quickly respond to events such as changes in clinical guidelines and updates to drug indications. In summary, the present embodiment realizes the high automation, standardization, and intelligence of the medical knowledge graph construction process through modular architecture design, high-performance knowledge extraction and verification mechanisms based on large language models, and dynamic processing capabilities for continuous updating. The system can greatly improve the accuracy, timeliness, and practicality of the stroke knowledge graph, providing a solid and reliable knowledge foundation for clinical decision support, scientific research discovery, and medical AI systems, and has significant technical advantages and broad application prospects in the fields of medical artificial intelligence and knowledge engineering.

[0079] In one possible implementation, the above utilizes the above LLM-based knowledge extraction module to perform named entity recognition, relation extraction, and attribute extraction tasks on the above stroke-related information through a large language model to obtain a candidate knowledge fragment stream, including:

[0080] The above stroke-related information is subjected to multi-dimensional feature analysis, and prompt demand scores and context dependency scores are calculated based on the specified task type;

[0081] Based on the demand score, the context dependency score, the text type information, and the specified task type, an internal rule engine is queried to automatically select the most suitable prompt template ID and context extraction strategy;

[0082] According to the context extraction strategy, the corresponding context information is extracted from the above stroke-related information;

[0083] According to the prompt template ID, the context information, and the stroke-related information, the placeholders in the template are filled to determine the dynamic prompt information;

[0084] The above dynamic prompt information and the corresponding context information are subjected to named entity recognition, relation extraction, and attribute extraction tasks based on the target LLM model to obtain a candidate knowledge fragment stream.

[0085] For example, in the LLM-based knowledge extraction module, by combining multi-dimensional feature analysis and dynamic prompt construction mechanisms, intelligent understanding and structured knowledge extraction of stroke-related information are completed to generate a candidate knowledge fragment stream.

[0086] First, the system receives pre-processed stroke-related information text and performs multi-dimensional feature analysis thereon. Based on the analysis results and the currently specified task type (e.g., named entity recognition NER, relation extraction RE, or attribute extraction Attribute), two key score indicators are calculated respectively:

[0087] Prompt demand score (PDS): measures the degree of demand for complex prompt instructions for the task;

[0088] Context dependency score (CDS): measures the range of context information required for task completion.

[0089] Subsequently, the system submits the above scoring results (PDS and CDS), text type information, and task type as joint inputs to an internal rule engine for querying and matching. The rule engine determines the most suitable prompt template ID (prompt_template_id) and context extraction strategy (context_strategy) based on a pre-set knowledge base or empirical rules for the current task. For example, when the task is "relation extraction", the text is guideline content, and the context dependency is high, the system may select the "Guideline_Detailed_RE" template and the "Paragraph_Summary" strategy.

[0090] After determining the extraction strategy, the system extracts the corresponding context information from the original stroke-related information according to the selected strategy. For example, if the strategy is "Sentence_±1", the current target sentence and one sentence before and after it are extracted; if it is "Document_Keywords", the semantically related auxiliary text segments are obtained through keyword search.

[0091] Next, the system dynamically fills the above context information, stroke-related information original text, and task description into the placeholders of the selected prompt template ID to form the final dynamic prompt information (dynamic prompt). This dynamic prompt not only contains the clearly defined task target, but also combines the appropriate context range to help the LLM better understand the semantic boundaries and extraction target.

[0092] Finally, the system submits the dynamic prompt information and context as input to the target LLM model (such as GPT-4, BioGPT, or fine-tuned version of PubMedBERT) to perform the specified extraction task. Based on the understanding of the instructions, the model completes the identification and classification of target entities (such as stroke subtypes, drug names), relationships (such as "treatment", "contraindication", "complication"), and attributes (such as dose value, time window, risk ratio, etc.) in the stroke text, and outputs a stream of candidate knowledge fragments, each containing structured knowledge elements, their source information, and initial confidence scores.

[0093] Specifically, it can include, for example: stroke-specific named entity recognition (NER) (e.g., identifying stroke subtypes, symptoms, drugs, doses, genes, anatomical locations, risk factors, diagnostic tests, outcomes). Relation extraction (RE) to identify complex relationships between entities (e.g., 'Alteplase' treats 'Acute Ischemic Stroke' within '4.5 hours', 'Atrial Fibrillation' is a risk factor for 'Cardioembolic Stroke', 'CYP2C19 Polymorphism' affects 'Clopidogrel' metabolism). Attribute extraction to capture specific details related to entities or relationships (e.g., dose value, time window, prevalence, research confidence level).

[0094] In summary, the embodiment introduces a prompt generation and context scheduling mechanism based on the MDPAC algorithm, enabling the LLM to have stronger task adaptation and semantic perception capabilities when processing medical text. Compared to static prompts or fixed input structures, this method can dynamically adjust prompt instructions and context ranges for different tasks and different text structures, significantly improving the accuracy, versatility, and explainability of knowledge extraction, providing a high-quality data foundation for subsequent knowledge graph construction and verification.

[0095] In a feasible implementation, the target LLM model described above includes a zero-shot LLM model, a few-shot LLM model, and a fine-tuned LLM model.

[0096] For example, according to the task complexity and available resources (or preset configuration), the module can select different target LLM model execution modes.

[0097] Zero-shot LLM model: For general or simple tasks, directly use dynamic_prompt to call the pre-trained LLM.

[0098] Few-shot LLM model: If the template library contains examples for specific task types, the algorithm can dynamically add a small number of relevant examples to dynamic_prompt to guide the LLM to better understand the task, and then call the LLM. Which examples to choose can also be part of the MDPAC strategy.

[0099] Fine-tuned model calling LLM model: If there is a fine-tuned LLM model for specific, high-difficulty entity / relationship types, and the current task matches it, the algorithm can choose to call this fine-tuned model and use the dynamic_prompt optimized for it (which may be different from the prompt for the general model).

[0100] In a feasible implementation, the multi-dimensional feature analysis described above includes text type recognition, structural complexity evaluation, information density evaluation, and task characteristic analysis,

[0101] The multi-dimensional feature analysis of the stroke-related information described above and the calculation of the prompt demand score and the context dependence score based on the specified task type include:

[0102] The text type recognition, the structural complexity evaluation, the information density evaluation, and the task characteristic analysis of the stroke-related information described above and the calculation of the prompt demand score and the context dependence score based on the specified task type.

[0103] Exemplarily, in a feasible implementation, the system, in the LLM-based knowledge extraction module, completes intelligent understanding and structured knowledge extraction of stroke-related information by combining multi-dimensional feature analysis and dynamic prompt construction mechanism, thereby generating a candidate knowledge fragment stream.

[0104] Specifically includes:

[0105] Sub-step 1: Text type recognition: Recognize the text type (such as 'literature abstract', 'clinical guideline','medical record') by source metadata source_metadata or lightweight classifier / LLM.

[0106] Sub-step 2: Complexity Assessment: Calculate the average sentence length, long sentence proportion, nested structure indicator density, medical term density, etc. of the text, and generate a complexity score CS.

[0107] Sub-step 3: Density Assessment: Calculate the entity mention density, numerical / measurement unit density, etc. of the text, and generate an information density score DS.

[0108] Sub-step 4: Task Specificity Analysis: Assign a basic weight according to the current extraction task task_type (NER, RE, Attribute).

[0109] Sub-step 5: Calculate the comprehensive score:

[0110] Prompt demand score (PDS): PDS = w1*CS + w2*DS + w3*TaskWeight_Prompt(task_type). This score reflects the degree of demand for generating detailed and accurate prompts.

[0111] Context dependence score (CDS): CDS = w4*CS + w5*(1-DS) + w6*TaskWeight_Context(task_type). This score reflects the degree of dependence of the task on more extensive context information. (Weights w1-w6 and TaskWeight_Prompt(task_type) are configurable parameters).

[0112] In summary, the embodiment introduces a prompt generation and context scheduling mechanism based on the MDPAC algorithm, enabling the LLM to have stronger task adaptation and semantic perception capabilities when processing medical text. Compared to static prompts or fixed input structures, this method can dynamically adjust the prompt instructions and context range for different tasks and different text structures, significantly improving the accuracy, versatility, and explainability of knowledge extraction, providing a high-quality data foundation for subsequent knowledge graph construction and verification.

[0113] As shown in Figure 4 , Figure 4 An internal workflow schematic diagram of an LLM-driven verification and refinement module in a dynamic stroke knowledge graph intelligent generation and maintenance system provided by the embodiment of the application.

[0114] In a possible implementation, the detection operation in the detection of the candidate knowledge fragment stream and the formal knowledge graph structure by the LLM-driven verification and refinement module includes consistency detection, fact detection, confidence score detection, and redundancy detection.

[0115] Illustratively, the LLM-driven verification and refinement module uses a special verification LLM model to improve the quality and reliability of the knowledge graph. It acts on the extracted fragments or the filled graph.

[0116] Consistency detection is used to identify logical contradictions or semantic conflicts between knowledge fragments or with existing content in the graph. The system prompts the LLM to evaluate the compatibility between statements: the LLM identifies conflicts. An example prompt: “Given the following statements extracted from the knowledge graph: (1) ‘Drug X is recommended for condition Y.’ (2) ‘Drug X is contraindicated for condition Y.’ Is there a contradiction? Please explain.”

[0117] Fact detection: used to prompt the LLM to verify statements based on its knowledge or provided context. Example prompt: “Verify the following statement based on current medical knowledge: ‘Aspirin is used for stroke primary prevention in low-risk individuals.’ Provide supporting evidence or reasons.”

[0118] Confidence score detection: the LLM analyzes the source text and extraction context to assign a confidence score. Example prompt: “Assign a confidence score (0-1) to the extracted relationship: ‘[subject] [predicate] [object]’ based on the certainty expressed in the source sentence: ‘[sentence text]’.”

[0119] Redundancy detection: used to identify duplicate or mergable knowledge.

[0120] In a possible implementation, the method further includes:

[0121] In the case of including low confidence information and conflicting information in the above detection results, the above detection results are displayed to the domain expert to obtain feedback information of the domain expert;

[0122] The formal knowledge graph structure is optimized or the LLM and prompt information used by the extraction and verification module are refined based on the feedback information of the domain expert.

[0123] For example, after the LLM-driven verification and refinement module completes consistency detection, fact checking, confidence scoring, and redundancy detection, the system will filter the results and automatically identify those segments with confidence scores below a preset threshold or with obvious conflicts with existing content in the graph. These marked candidate knowledge usually belongs to content with fuzzy semantic boundaries, uncertain expressions, or strong semantic ambiguity, which is difficult to directly confirm its effectiveness through an automatic model.

[0124] For these complex or uncertain detection results, the system will automatically integrate and display them in a human-machine collaborative interface. The interface provides clear knowledge segment display, context reference, conflict item highlighting, confidence visualization, and other functions to facilitate the domain expert to quickly understand the background and judgment basis. The domain expert can evaluate the knowledge segment based on his own professional knowledge and clinical experience, and give corresponding feedback information such as "confirm", "deny", "suggest modification", or "observe" and the like.

[0125] After the system receives the expert feedback, it performs subsequent optimization operations, which include two types of paths:

[0126] On the one hand, if the expert makes a clear decision on the validity of a certain segment (for example, confirming that the knowledge is correct or should be deleted), the system will directly update or correct the formal knowledge graph structure based on the feedback, ensuring that the graph content remains accurate, unified, and logically self-consistent. The system will also record the audit information and version information of this modification while updating the graph, supporting subsequent knowledge tracking and backtracking.

[0127] On the other hand, the system will also optimize the knowledge extraction and verification process based on the expert feedback information. For example, when multiple experts repeatedly make consistent corrections on a certain type of extraction error, the system can modify the prompt template, update the rule weights in the MDPAC algorithm, or fine-tune the LLM model used in the extraction / verification stage, thereby improving the adaptability and extraction ability of the model in similar tasks, achieving continuous iteration and self-enhancement.

[0128] In summary, the embodiment introduces a human-computer collaboration link, establishes an efficient feedback channel between system automatic processing capability and expert judgment, so that low-confidence or conflicting knowledge can be confirmed or corrected through manual review, not only improving the overall quality and clinical applicability of the stroke knowledge graph, but also promoting the progressive optimization of system prompt generation strategy and language model, forming a closed-loop knowledge evolution mechanism of knowledge extraction-verification-correction-enhancement.

[0129] In a feasible implementation, the integration of the candidate knowledge fragment stream into the formal knowledge graph structure stored in the graph database by the knowledge graph structuring and filling module includes:

[0130] The knowledge graph structuring and filling module integrates the candidate knowledge fragment stream into a triple form conforming to the pre-defined medical ontology structure, and performs entity standardization and disambiguation operations to obtain a structured result, wherein the entity standardization includes linking the entity to the corresponding concept entry in the standard medical terminology system;

[0131] The structured result is stored in the graph database to construct the formal knowledge graph.

[0132] For example, the system first receives a candidate knowledge fragment stream output from the LLM-based knowledge extraction module, each fragment containing entities, relationships, attributes, source information, and initial confidence scores. In the knowledge graph structuring stage, the system organizes these fragments according to the pre-defined medical ontology structure. The ontology structure includes common entity types in the field of stroke (such as "stroke subtype", "symptom", "treatment drug", "diagnostic method", "risk factor", etc.), relationship types (such as "cause", "treatment", "contraindication", "indication", "complication", etc.), and attribute items (such as "dose", "time window", "prevalence", "confidence level", etc.).

[0133] Subsequently, the system converts the extracted fragments into standardized triple form, i.e., the structure of (subject, predicate, object). For example, a candidate knowledge fragment "Alteplase can be used for acute ischemic stroke treatment" will be converted into a triple: (Alteplase, treatment, acute ischemic stroke).

[0134] While generating triples, the system performs entity standardization and disambiguation operations. For each identified entity, the system aligns it with the corresponding concept entry in the standard medical terminology system through entity linking mechanism. Common terminology systems include:

[0135] SNOMED CT: for disease names, anatomical sites, etc.

[0136] ICD-10: for disease classification and coding;

[0137] RxNorm: Used in relation to drug ingredients and generic brand names;

[0138] LOINC: Used for laboratory testing and diagnostic tests;

[0139] UMLS: A unified medical terminology mapping across semantic networks.

[0140] For example, terms such as "cerebral infarction," "ischemic stroke," and "cerebrovascular accident" are expressed differently in text, but the system can standardize and link them to the same SNOMED CT concept ID through context analysis and semantic matching, thereby eliminating semantic redundancy and ambiguity.

[0141] After entity standardization, the system packages the triples and their attribute information (such as extraction source, timestamp, and confidence score) into a structured result and writes it to a graph database. Commonly used graph databases include Neo4j, JanusGraph, or RDF triple storage platforms such as Blazegraph. Each node represents an entity, each edge represents a relation, and both edges and nodes can be attached with attributes and metadata to support subsequent inference, querying, and version control.

[0142] Through the above process, this embodiment achieves automated conversion from candidate fragments to formal atlases, ensuring both the professional standardization of medical terminology and improving the logical consistency and scalability of the atlas structure. This module not only supports the use of subsequent inference engines and diagnostic decision-making systems but also provides a semantically unified foundation for version updates and atlas fusion, making it a crucial link in the entire system.

[0143] In one feasible implementation, the knowledge graph structure includes entity nodes, relation edges representing semantic relationships between entities, attribute sets associated with the entity nodes or relation edges, and mapping link information between entities or relationships and the standard medical ontology system.

[0144] For example, in one feasible implementation, to achieve structured representation and efficient querying of stroke medical knowledge, the aforementioned knowledge graph structure is designed as a data graph model with complete semantic representation capabilities. Specifically, it includes entity nodes, relation edges representing semantic relationships between entities, sets of associated attributes, and mapping link information between entities or relationships and the standard medical ontology system. This structure is uniformly generated by the knowledge graph structuring and filling module during construction and storage, and is continuously maintained in the graph database.

[0145] Firstly, Entity Nodes are the basic units in the graph, representing important concept objects related to stroke, such as disease types (e.g., "Cerebral Hemorrhage", "Acute Ischemic Stroke"), symptoms (e.g., "Hemiplegia", "Speech Disorder"), therapeutic drugs (e.g., "Alteplase", "Clopidogrel"), interventions (e.g., "Intravenous Thrombolysis", "Mechanical Thrombectomy"), clinical examinations, risk factors (e.g., "Hypertension", "Atrial Fibrillation"), etc. Each Entity Node exists in the graph database as a unique identifier, containing attributes such as its original term, standard name, language synonyms, text extraction source, and timestamp.

[0146] Secondly, Relation Edges are used to connect Entity Nodes, representing specific semantic relationships between them. These relationship types are derived from predefined medical ontology structures, such as "Treatment", "Cause", "Complication of", "Contraindication", "Risk Factor", "Recommended for", etc. For example, the following Relation Edge can be established in the graph: ("Alteplase") — "Treatment" → ("Acute Ischemic Stroke"). Relation Edges not only identify semantic connections but also can be attached with attributes such as weight, source, confidence, etc.

[0147] Further, each Entity Node or Relation Edge can be associated with one or more attribute sets, describing specific information related to the node or edge. For example, for the drug entity "Clopidogrel", its attributes may include "Dose Range: 75mg-300mg", "Metabolic Pathway: CYP2C19", "Adverse Reaction Risk Level: Moderate"; while the Relation Edge "Treatment" can be accompanied by attributes such as "Indication Description", "Guideline Support Level", "Time Window Limitation", etc. These attributes constitute the extended semantic information part of the graph, making it have stronger expression ability in actual medical applications.

[0148] Finally, to enhance the portability and consistency of medical terminology of the graph, the system establishes mapping link information (Ontology Mapping) between each entity or relation and the standard medical ontology system. This mapping process includes Entity Linking and semantic normalization, supporting the connection of graph content to multiple mainstream standard systems, such as: SNOMED CT (Systematized Nomenclature of Medicine Clinical Terms) - Disease and Clinical Terminology; ICD-10 (International Classification of Diseases) - International Disease Coding; RxNorm - Drug and Generic Name; LOINC - Laboratory and Diagnostic Test Terminology; UMLS - Unified Medical Language System.

[0149] For example, the node "cerebral infarction" can be linked to the concept number 230690007 in SNOMED CT, corresponding to the standard name "Cerebral infarction". This linkage not only helps to realize the docking of the knowledge graph with external systems (such as electronic medical record systems, clinical decision support platforms, data sharing interfaces, etc.), but also provides a unified semantic basis for graph reasoning.

[0150] In summary, the knowledge graph structure constructed in this embodiment has rigorous semantic representation ability, and is composed of entity nodes, relationship edges, attribute sets and ontology mapping, forming a standardized, expandable and semantically consistent cerebral apoplexy medical knowledge network. This structure not only supports the visual understanding and editing of human experts, but also meets the calling needs of machine reasoning and automatic question answering systems, and is the key foundation for the system to realize efficient intelligent auxiliary diagnosis and knowledge management.

[0151] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A dynamic stroke knowledge graph intelligent generation and maintenance system, characterized in that, include: The data access and preprocessing module is used to access and retrieve stroke-related information from multiple data sources; The knowledge extraction module based on LLM is used to perform named entity recognition, relation extraction and attribute extraction tasks on the stroke-related information through a large language model to obtain a candidate knowledge fragment stream, wherein the candidate knowledge fragment stream includes entities, relations, attributes, relevant source information and initial confidence scores; The knowledge graph structuring and filling module is used to integrate the candidate knowledge fragment stream into the formal knowledge graph structure stored in the graph database; The LLM-driven verification and refinement module is used to detect the candidate knowledge fragment stream and the formal knowledge graph structure to obtain detection results; The dynamic update and maintenance module is used to periodically monitor the multi-source data sources accessed by the data access and preprocessing module, identify newly released or updated content about stroke, and trigger the incremental processing flow of the knowledge extraction module, the knowledge graph structuring and filling module, and the LLM-driven verification and refinement module, so as to realize the dynamic fusion and version control of the formal knowledge graph structure.

2. The intelligent generation and maintenance system for dynamic stroke knowledge graphs according to claim 1, characterized in that, Also includes: The human-machine collaboration module is used to display the detection results to domain experts to obtain feedback from the domain experts when the detection results include low-confidence information and conflicting information, so as to optimize the formal knowledge graph structure.

3. A method for intelligent generation and maintenance of a dynamic stroke knowledge graph, used in the intelligent generation and maintenance system for a dynamic stroke knowledge graph as described in claim 1 or 2, characterized in that, include: The data access and preprocessing module accesses and retrieves stroke-related information from multiple data sources. The LLM-based knowledge extraction module is used to perform named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information through a large language model to obtain a candidate knowledge fragment stream, wherein the candidate knowledge fragment stream includes entities, relations, attributes, relevant source information, and initial confidence scores; The candidate knowledge fragment stream is integrated into the formal knowledge graph structure stored in the graph database through the knowledge graph structuring and filling module. The LLM-driven verification and refinement module is used to detect the candidate knowledge fragment stream and the formal knowledge graph structure to obtain detection results. The dynamic update and maintenance module periodically monitors the multi-source data sources accessed by the data access and preprocessing module, identifies newly released or updated content related to stroke, and triggers the incremental processing flow of the knowledge extraction module, the knowledge graph structuring and filling module, and the LLM-driven verification and refinement module to achieve dynamic fusion and version control of the formal knowledge graph structure.

4. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 3, characterized in that, The step of utilizing the LLM-based knowledge extraction module to perform named entity recognition, relation extraction, and attribute extraction tasks on the stroke-related information through a large language model to obtain a candidate knowledge fragment stream includes: Multidimensional feature analysis was performed on the stroke-related information, and prompting demand score and context dependency score were calculated based on the specified task type. Based on the demand score, the context dependency score, the text type information, and the specified task type, the internal rule engine queries are used to automatically select the most suitable prompt template ID and context extraction strategy. According to the context extraction strategy, the corresponding context information is extracted from the stroke-related information; The prompt template ID, the context information, and the stroke-related information are used to fill in the placeholders in the template to determine the dynamic prompt information; The dynamic prompt information and the corresponding context information are used to perform named entity recognition, relation extraction and attribute extraction tasks based on the target LLM model to obtain a stream of candidate knowledge fragments.

5. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 4, characterized in that, The target LLM model includes zero-sample LLM model, few-sample LLM model, and fine-tuned LLM model.

6. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 4, characterized in that, The multidimensional feature analysis includes text type recognition, structural complexity assessment, information density assessment, and task characteristic analysis. The step of performing multidimensional feature analysis on the stroke-related information and calculating prompt demand score and context dependency score based on a specified task type includes: The stroke-related information is subjected to text type recognition, structural complexity assessment, information density assessment, and task characteristic analysis, and prompting demand score and context dependency score are calculated based on the specified task type.

7. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 3, characterized in that, The detection operations performed by the LLM-driven verification and refinement module to detect the candidate knowledge fragment stream and the formal knowledge graph structure include consistency detection, fact detection, confidence score detection, and redundancy detection.

8. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 3, characterized in that, The method further includes: If the detection results include low-confidence information and conflicting information, the detection results are presented to domain experts to obtain their feedback. The formal knowledge graph structure is optimized or the LLM and prompt information used by the extraction and verification modules are refined based on feedback from the domain experts.

9. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 3, characterized in that, The process of integrating the candidate knowledge fragment stream into the formal knowledge graph structure stored in the graph database through the knowledge graph structuring and filling module includes: The knowledge graph structuring and filling module integrates the candidate knowledge fragment stream into a triplet form that conforms to a predefined medical ontology structure, and performs entity standardization and disambiguation operations to obtain structured results. The entity standardization includes linking entities to corresponding concept entries in the standard medical terminology system. The structured results are stored in the graph database to construct the formal knowledge graph.

10. The method for intelligent generation and maintenance of dynamic stroke knowledge graph according to claim 9, characterized in that, The knowledge graph structure includes entity nodes, relation edges representing semantic relationships between entities, attribute sets associated with the entity nodes or relation edges, and mapping link information between entities or relationships and the standard medical ontology system.

Images