Kawasaki disease intelligent hospital guide method and system based on knowledge graph reasoning

Abstract

The invention discloses a Kawasaki disease intelligent hospital guide method and system based on knowledge graph reasoning. The system comprises a knowledge base construction layer, a question understanding layer, a core reasoning layer and an answer generation and evaluation layer. The knowledge base construction layer comprises an entity extraction module, a relationship extraction module and a knowledge storage module and is used for constructing a Kawasaki disease knowledge graph; according to the Kawasaki disease intelligent guide method and system based on knowledge graph reasoning, a structured Kawasaki disease knowledge graph which can be understood and reasoned by a machine is constructed to serve as a brain, the deep medical intention of a question sentence of a user can be understood, and the knowledge graph reasoning-based Kawasaki disease intelligent guide method and system based on knowledge graph reasoning have the advantages that the knowledge graph reasoning-based Kawasaki disease intelligent guide method and system are high in practicability. And a plurality of symptoms can be associated for comprehensive judgment like a doctor, so that accurate and credible answers with risk early warning capability are output, and the problems of inaccurate, incomplete and uncredible answers in the prior art are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent triage technology for Kawasaki disease, specifically to an intelligent triage system for Kawasaki disease based on knowledge graph reasoning. Background Technology

[0002] Kawasaki disease is an acute, febrile, exanthematous vasculitis syndrome that commonly affects infants and young children. It is a major cause of acquired heart disease in children. Its etiology is unknown, and its clinical manifestations are diverse and often atypical, making it easily confused with scarlet fever, measles, and viral infections. Users often use general medical Q&A robots or search engines to search for Kawasaki disease symptoms. However, some of these robots or search engines have the following drawbacks in the process of searching for Kawasaki disease symptoms: 1. Poor accuracy and authority of information, posing a risk of being misled. For example, the core mechanism of general search engines and Q&A platforms is keyword matching and link ranking (such as PageRank). They index the entire Internet space and cannot make medical-level identifications of the professionalism and accuracy of information sources. As a result, the results returned to users after the query are filled with a large number of unprofessional personal experience sharing, outdated medical information, commercial advertisements and even pseudoscientific content. For ordinary parents who lack professional judgment, they are easily misled and may take incorrect home care measures or delay the best time for medical treatment, thereby increasing the risk of serious complications such as coronary artery disease in children. 2. Lack of in-depth domain knowledge and inability to understand professional contexts. For example, existing technologies are trained on general corpora, and their vocabulary and semantic understanding models have not been deeply optimized for the vertical domain of Kawasaki disease. This results in medical question-answering robots or search engines being unable to accurately interpret professional terms and specific clinical manifestations such as "incomplete Kawasaki disease", "coronary artery Z-value", and "carpal redness and swelling". When users use such professional or colloquial descriptions, the system cannot correctly identify their medical entities, resulting in search failures or answer deviations. 3. Lack of clinical reasoning ability; answers are one-sided and lack relevance. For example, the current technology is essentially "information retrieval" rather than "intelligent question answering," lacking the structured representation and reasoning ability of the inherent logical relationship between symptoms, signs, and diseases. When a user inputs multiple symptoms such as "fever for five days, red eyes, rash," the system can only return a list of web pages related to the three independent keywords "fever," "red eyes," and "rash." It cannot perform clinical thinking like a doctor, connect these symptoms, infer the possibility of "Kawasaki disease," and proactively provide its diagnostic criteria and differential diagnosis information. 4. Weak interpretability and low user trust: The decision-making process of existing technologies (especially deep learning-based chatbots) is a "black box." Users cannot know the basis for the generation of answers. When the system returns a conclusion (such as "it may be Kawasaki disease"), there is no chain of medical evidence to support the conclusion (such as which diagnostic criteria were used and which diseases it was differentiated from). This makes it difficult for users, especially cautious doctors, to assess the credibility of the answers, thereby reducing their dependence on and trust in the system. To address this, we propose a knowledge graph-based intelligent triage method and system for Kawasaki disease. Summary of the Invention

[0003] The technical problem this invention aims to solve is to overcome the shortcomings of existing methods and provide an intelligent triage method and system for Kawasaki disease based on knowledge graph reasoning. This method and system construct a structured, machine-understandable, and reasonable Kawasaki disease knowledge graph as its core, and implement a closed-loop intelligent processing flow around it, including question understanding, graph reasoning, answer generation, and evaluation. This solution can understand the deep medical intent of user questions and, like a doctor, can associate multiple symptoms to make a comprehensive judgment, thereby outputting accurate, reliable answers with risk warning capabilities. This fundamentally solves the problems of inaccurate, incomplete, and unreliable answers in existing technologies and can effectively address the problems in the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a Kawasaki disease intelligent triage method and system based on knowledge graph reasoning, wherein the system includes a knowledge base construction layer, a question comprehension layer, a core reasoning layer, and an answer generation and evaluation layer; Knowledge base construction layer: It includes an entity extraction module, a relation extraction module, and a knowledge storage module, which are used to construct the Kawasaki disease knowledge graph; Question Comprehension Layer: Used to identify entities and nodes within the Kawasaki disease consultation information input by the user; Core reasoning layer: used to query, reason about, and output the symptom entities within the Kawasaki disease consultation information input by the user; Answer generation and evaluation layer: This layer refines the inference results and assesses the credibility and risk of Kawasaki disease. The method and system construct a structured, machine-understandable, and reasonable Kawasaki disease knowledge graph as the brain, and implement a closed-loop intelligent processing flow around it, including question understanding, graph reasoning, answer generation, and evaluation. This solution can understand the deep medical intent of user questions and, like a doctor, can associate multiple symptoms to make a comprehensive judgment, thereby outputting accurate, credible answers with risk warning capabilities, fundamentally solving the problems of inaccurate, incomplete, and unreliable answers in existing technologies.

[0005] Furthermore, the entity extraction module uses a pre-trained model such as BERT to perform medical entity recognition in text, the relation extraction module uses a pre-trained model such as BERT to identify the relationships between entities, and the knowledge storage module stores the extracted entities and relationships into a graph database to form a structured Kawasaki disease knowledge graph. The structured Kawasaki disease knowledge graph is constructed through information extraction.

[0006] Furthermore, the question understanding layer includes a named entity recognition module and an intent recognition module, which perform in-depth understanding and recognition of the natural language questions input by the user.

[0007] Furthermore, the named entity recognition module uses a finely tuned BERT model on medical text to perform sequence annotation on user questions, identify key medical entities, and identify key medical entities in the entity and relation data of the knowledge storage module.

[0008] Furthermore, the intent recognition module uses a BERT-based text classification model to classify questions into preset intent categories and recognize intent categories within the entity and relation data of the knowledge storage module.

[0009] Furthermore, the core reasoning layer includes a graph query module, a multi-hop reasoning module, and a candidate answer generation module. The graph query module uses the reasoning engine to find common disease nodes between entities in the knowledge graph, starting from the identified entities. The multi-hop reasoning module performs reasoning on multiple paths. The candidate answer generation module summarizes the disease nodes, their confidence levels, and reasoning paths for the engine, queries the graph database with the output results of the question comprehension layer, and performs in-depth reasoning, enabling the system to make a comprehensive judgment by associating multiple symptoms like a doctor.

[0010] Furthermore, the answer generation and evaluation layer includes an answer refinement and generation module and a credibility and risk assessment module. The answer refinement and generation module uses a fine-tuned text generation model, taking the data for generating candidate answer outputs, the reasoning path of multi-hop reasoning, and relevant original text fragments stored in the knowledge storage as inputs, to automatically generate natural language answers, and to refine and assess the credibility and risk of the Kawasaki disease intelligent triage output results based on knowledge graph reasoning.

[0011] Furthermore, the credibility and risk assessment module includes a credibility assessment submodule, a risk assessment submodule, and an interpretability output submodule, which assess the credibility, risk, and interpretability of the Kawasaki disease intelligent triage output results based on knowledge graph reasoning.

[0012] This invention also provides a triage method for a knowledge graph-based intelligent triage system for Kawasaki disease, comprising the following steps: Step 1: Construct a precise, authoritative, and verifiable knowledge graph of Kawasaki disease through the knowledge base construction layer; Step Two: Through the question comprehension layer and the core reasoning layer, a qualitative leap is achieved from information retrieval to clinical auxiliary reasoning; Step 3: Through the answer generation and evaluation layer, different levels of risk warnings are automatically triggered based on the answer content, and a quantitative credibility assessment is given. Finally, a comprehensive answer output mechanism that combines information, security and interpretability is generated, effectively avoiding medical delays caused by information asymmetry.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: The intelligent triage method and system for Kawasaki disease based on knowledge graph reasoning has the following advantages: 1. The Kawasaki disease intelligent triage method and system based on knowledge graph reasoning, through the knowledge base construction layer (constructing a multimodal Kawasaki disease knowledge graph), fundamentally limits the knowledge source to authoritative medical textbooks, clinical guidelines, and desensitized medical records. This means that any answer generated by the system has underlying data that has been recognized by the medical community. For example, when answering "diagnostic criteria for Kawasaki disease", the system directly maps and outputs the main criteria such as "persistent fever" and "rash" clearly defined in the guidelines, rather than personal experiences from online forums, thereby eliminating erroneous and misleading information at the source.

[0014] 2. The Kawasaki disease intelligent triage method and system based on knowledge graph reasoning, through the collaborative work of the question understanding layer and the core reasoning layer, enables the system to handle complex clinical symptoms. Specifically, the named entity recognition module of the question understanding layer can extract multiple symptom entities (such as "fever", "red eyes", "rash") from the question, and the graph reasoning engine of the core reasoning layer treats these entities as clues and searches for "disease" nodes in the knowledge network that can explain all these clues at the same time. This process simulates the differential diagnosis thinking of doctors: instead of simply listing the information of "fever" and "red eyes", it calculates that "Kawasaki disease" is one of the common causes that can cause these two symptoms at the same time and actively gives this comprehensive judgment. This is a functional leap that keyword matching technology cannot achieve at all.

[0015] 3. The Kawasaki disease intelligent triage method and system based on knowledge graph reasoning, through the answer generation and evaluation layer, mandates that the output content must include its reasoning path and credibility score. For example, when suggesting "suspected Kawasaki disease", the system will indicate which symptom entities in the question are based on this conclusion and give a 95% confidence level (this confidence level is derived from the strength of the evidence chain supporting the conclusion in the knowledge graph). This interpretability mechanism allows users (especially primary care physicians) to review the system's logic, thereby building trust in AI and making them more willing to adopt its suggestions in real-world scenarios.

[0016] 4. A dedicated risk assessment module is integrated into the answer generation and evaluation layer. This module has built-in risk rules defined by medical experts (such as "fever ≥5 days" + "any other major symptoms" = high risk). When the results in the answer generation and evaluation layer trigger these rules, the system no longer passively returns information, but actively intervenes and inserts a strong warning at the beginning of the answer. This design is equivalent to equipping each user with a tireless "triage nurse" who can provide crucial safety warnings beyond the user's inquiry itself in emergency situations that have a significant impact on life and health, effectively avoiding delays in medical treatment caused by information asymmetry. Attached Figure Description

[0017] Figure 1 This is a structural block diagram of the Kawasaki disease intelligent triage system based on knowledge graph reasoning of the present invention; Figure 2 This is a sub-block diagram of the credibility and risk assessment module of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-2 This embodiment provides a technical solution: a Kawasaki disease intelligent triage method and system based on knowledge graph reasoning. The system includes a knowledge base construction layer, a question comprehension layer, a core reasoning layer, and an answer generation and evaluation layer. The knowledge base construction layer includes an entity extraction module, a relation extraction module, and a knowledge storage module, used to construct the Kawasaki disease knowledge graph. The entity extraction module uses pre-trained models like BERT to recognize medical entities in text. The relation extraction module uses pre-trained models like BERT to identify relationships between entities. The knowledge storage module stores the extracted entities and relationships into a graph database, forming a structured Kawasaki disease knowledge graph. The Kawasaki disease knowledge graph includes refined definitions of entities and attributes: Symptom entities: These not only have points but also attributes. For example, the entity 'fever' has attributes such as 'duration' and 'fever peak'. In the knowledge graph, 'fever greater than 5 days' is represented as '(symptom entity: fever) - [attribute: duration > 5 days]'. This allows the system to distinguish between 'fever for 5 days' and 'fever for 3 days' during reasoning, as the former has a very high weight in Kawasaki disease diagnosis. Special sign entities: Entities such as 'carpal redness and swelling' and 'coronary artery Z-score', which are uncommon or have very low weight in general medical knowledge bases, are introduced. The `Z-value` is defined as an entity with a `numerical range` attribute, such as `(Z-value)-[attribute: range ≥ 2.5]` indicating coronary artery dilation. First, information is extracted from authoritative medical sources (such as clinical guidelines, medical textbooks, and anonymized electronic medical records) using natural language processing (NLP) technology. During this process, the entity extraction module uses a pre-trained model, such as BERT, to identify medical entities in the text, such as "Kawasaki disease" (disease), "strawberry tongue" (symptom), and "gamma globulin" (drug). These entities constitute "nodes" in the knowledge graph. Subsequently, the relation extraction module uses a pre-trained model, such as BERT... The model identifies relationships between entities, such as "Kawasaki disease" - HasSymptom -> "fever" and "Kawasaki disease" - TreatWith -> "gamma globulin". These relationships are then used as "edges" connecting nodes to achieve relationship extraction. Subsequently, the "entities" and "relationships" extracted from the entity extraction module and the relationship extraction module are stored in its own graph database (such as Neo4j) through the knowledge storage module to form a structured "Kawasaki disease knowledge graph", thereby achieving knowledge storage. In addition, the knowledge storage module uses embedding technology to associate unstructured data such as medical image features and laboratory indicators with corresponding entities to form a multimodal knowledge representation. Question Understanding Layer: This layer identifies entities and nodes within user-inputted Kawasaki disease consultation information. It includes a named entity recognition module and an intent recognition module. The entity recognition and relation extraction technology selection uses a BERT model further pre-trained on medical corpora (such as MIMIC-III, Chinese medical papers) as a foundation, followed by task-specific fine-tuning. This is based on the superior performance of BERT and other Transformer architectures in understanding contextual semantics. Medical pre-training enabled it to master medical terminology, and fine-tuning for Kawasaki disease medical records and guidelines allowed it to accurately identify specific expressions in areas such as "Z-score" and "Kawasaki scar." The named entity recognition module uses a finely tuned BERT model on medical text to perform sequence labeling on user questions and identify key medical entities. The intent recognition module uses a base... Based on the BERT text classification model, questions are categorized into preset intent categories. When a user inputs a natural language question (such as "My child has had a fever for five days and his eyes are red") into the named entity recognition module through an external input module to perform a symptom query, the named entity recognition module uses a BERT model finely tuned on medical text to perform sequence labeling on the user's question and identify key medical entities. For example, it accurately identifies the two symptom entities "fever" and "red eyes" from the question. At the same time, the intent recognition module uses a text classification model (also based on BERT) to classify the entire question into preset intent categories, such as "diagnosis consultation", "treatment plan consultation", "cause inquiry", etc. In this example, the question is identified as the intent of "diagnosis consultation", and the result is then input into the graph query module as structured data consisting of {intent: diagnosis consultation, entity: [fever, red eyes]}. The core reasoning layer is used to query, reason about, and output the symptom entities within the Kawasaki disease consultation information input by the user. The core reasoning layer includes a graph query module, a multi-hop reasoning module, and a candidate answer generation module. The graph query module uses the reasoning engine to simulate a graph structure for differential diagnosis, starting from the identified entities. This graph not only contains the relationship `[Kawasaki disease] - [Has Symptom] -> [Fever]`, but also constructs rich `[Differential Diagnosis]` relationships. For example, both `[Scarlet Fever]` and `[Kawasaki disease]` can be identified through `[Differential Diagnosis]`. The system connects to common symptoms such as strawberry tongue and rash. The reasoning process is as follows: when a user inputs "fever, strawberry tongue, rash," the system starts from these three nodes and searches for diseases that are commonly connected. It finds scarlet fever and Kawasaki disease as candidates. At this point, the pre-set "weights" or "strength of evidence" in the graph come into play. If the user also inputs "scarred and swollen skin" (a very specific but not sensitive sign), because this symptom has a very high connection weight with Kawasaki disease but no connection with scarlet fever, the system will significantly increase the confidence of Kawasaki disease and may indicate in the answer that "Kawasaki disease should be given high suspicion." The differential diagnosis of Kawasaki disease, stating that "although the symptoms overlap with scarlet fever, 'carbuncles and swelling' are characteristic manifestations of Kawasaki disease," directly simulates the clinical thinking process of a doctor. It involves searching for common disease nodes between entities in the knowledge graph, a multi-hop reasoning module reasoning along multiple paths, and a candidate answer generation module that aggregates disease nodes, their confidence levels, and reasoning paths. The graph query module, starting with identified entities (such as "fever" and "red eyes"), searches the knowledge graph for common disease nodes connected to these entities. Subsequently, the multi-hop reasoning module deduces the disease nodes connected to these entities. During this process, the multi-hop reasoning module, through the engine... Instead of simply finding direct connections to disease nodes, it performs "multi-hop" queries. For example, it finds "Kawasaki disease" from "fever," and then finds other related symptoms such as "rash" and "strawberry tongue" from "Kawasaki disease" through the HasSymptom relation. Simultaneously, it also finds "Kawasaki disease" from "red eyes." When multiple paths point to "Kawasaki disease," the weight (confidence) of that node is increased. The candidate answer generation module then uses the engine to aggregate all high-weight disease nodes, their confidence levels, and inference paths (i.e., all nodes and edges traversed from the question entity to the answer disease), forming a structured candidate answer, for example: {Disease: Kawasaki disease, Confidence: 0.}95. Path: [Fever -> Kawasaki Disease], [Red Eyes -> Kawasaki Disease]}. In this process, multiple isolated symptom / sign entities identified in the user query are used as joint query conditions. Graph queries and multi-hop reasoning are performed in the knowledge graph to find the disease node that can most reasonably explain all these symptoms and calculate its confidence. The specific algorithm and implementation path achieve a qualitative leap from information retrieval to clinical auxiliary reasoning. Knowledge graph reasoning and confidence calculation: The algorithm adopts the PathRanking Algorithm (PRA). Alternatively, in a Graph Neural Network (GNN) inference model, confidence is calculated as the confidence score of the candidate answer Disease(D) for the input symptom set (S), which can be quantified as follows: `Confidence(D|S) = λ1(number of paths linking D and S / total number of paths) + λ2*(average weight of the path) + λ3*(prior probability of node D itself)` where λ1, λ2, and λ3 are adjustable weight coefficients, and the path weights can be pre-set by medical experts or obtained through data learning. Answer Generation and Evaluation Layer: This layer refines the inference results and assesses the credibility and risk of Kawasaki disease. It includes an answer refinement and generation module and a credibility and risk assessment module. The answer refinement and generation module uses a fine-tuned text generation model, taking the data used to generate candidate answers, the inference path of multi-hop reasoning, and relevant original text fragments stored in the knowledge store as input to automatically generate natural language answers. The credibility and risk assessment module includes a credibility assessment submodule, a risk assessment submodule, and an interpretability output submodule. The core of the risk assessment submodule is a combination of a rule-based expert system and a list of high-risk entities. The rules are not generated out of thin air, but are implemented by translating the latest written clinical guidelines (such as the American Heart Association (AHA) Guidelines for the Diagnosis of Kawasaki Disease) and expert consensus into a series of IF-THEN logical statements. Example of rule source and transformation: Source: The AHA guidelines state that fever ≥5 days accompanied by at least 4 of the following 5 major symptoms (bilateral conjunctival hyperemia, lip changes, polymorphic rash, limb changes, non-suppurative cervical lymphadenopathy) is sufficient for a diagnosis of typical Kawasaki disease. Rule transformation: IF identifies entity [fever] and its attribute `[duration] ≥5 days` AND `other identified symptom entities belong to `[major]` If the symptom list is greater than or equal to 4, then THEN triggers a "high-risk" warning, outputting the message: "Warning! The child's symptoms highly meet the diagnostic criteria for typical Kawasaki disease, indicating a high risk of coronary artery disease. Please immediately seek medical attention at a qualified pediatric cardiovascular department of a hospital!" Subsequently, the answer refinement and generation module uses a finely tuned text generation model (such as T5) as input, taking the candidate answers output by the candidate answer generation module, the reasoning path, and relevant guideline excerpts stored in the knowledge graph as input to generate a fluent and easy-to-understand natural language answer, thus facilitating user comprehension. For example: "Persistent fever for more than 5 days accompanied by bilateral red eyes is a typical Kawasaki disease..." One of the main diagnostic criteria for Kawasaki disease is the generative function of the answer. Simultaneously, the credibility assessment submodule calculates a comprehensive credibility score (e.g., 95%) based on the confidence level calculated by the candidate answer generation module and the completeness of the evidence chain in the knowledge graph. The risk assessment submodule uses a pre-set risk rule base. When the risk assessment submodule identifies a specific high-risk entity (e.g., "fever lasting more than 5 days," "coronary artery dilation") or combination, an alarm is triggered. In this example, since "fever lasting more than 5 days" is a key warning signal for Kawasaki disease, the system automatically adds a strong risk warning before the final answer, such as "Warning: This symptom combination is highly suggestive of Kawasaki disease; please seek medical attention immediately!" Finally, through the interpretability output submodule, the system integrates the generated answer, risk warning, credibility score, and key reasoning evidence (e.g., "Based on: AHA guidelines") and outputs it to the user.

[0020] The triage method of the Kawasaki disease intelligent triage system based on knowledge graph reasoning includes the following steps: Step 1: Construct a precise, authoritative, and verifiable knowledge graph of Kawasaki disease through the knowledge base construction layer; Step Two: Through the question comprehension layer and the core reasoning layer, a qualitative leap is achieved from information retrieval to clinical auxiliary reasoning; Step 3: Through the answer generation and evaluation layer, different levels of risk warnings are automatically triggered based on the answer content, and a quantitative credibility assessment is given. Finally, a comprehensive answer output mechanism that combines information, security and interpretability is generated, effectively avoiding medical delays caused by information asymmetry.

[0021] The working principle of the Kawasaki disease intelligent triage method and system based on knowledge graph reasoning provided by this invention is as follows: First, information is extracted from authoritative medical sources (such as clinical guidelines, medical textbooks, and desensitized electronic medical records) using natural language processing technology. During this process, the entity extraction module uses a pre-trained model, such as BERT, to identify medical entities in the text, such as "Kawasaki disease" (disease), "strawberry tongue" (symptom), and "gamma globulin" (drug). These entities constitute "nodes" in the knowledge graph. Subsequently, the relationship extraction module uses a pre-trained model, such as BERT, to identify the relationships between entities, such as "Kawasaki disease" (disease), "strawberry tongue" (symptom), and "gamma globulin" (drug). The relationships "disease" - HasSymptom -> "fever", "Kawasaki disease" - TreatWith -> "gamma globulin" are used as "edges" connecting nodes to achieve relation extraction. Subsequently, the "entities" and "relationships" extracted from the entity extraction and relation extraction modules are stored in a graph database (such as Neo4j) through a knowledge storage module, forming a structured "Kawasaki disease knowledge graph," thus achieving knowledge storage. Furthermore, the knowledge storage module uses embedding technology to associate unstructured data such as medical image features and laboratory indicators with corresponding entities, forming a multimodal knowledge representation. When a user inputs named entities through an external input module... When the entity recognition module inputs a natural language question (such as "My child has had a fever for five days and his eyes are red") to perform a symptom query, the named entity recognition module uses a BERT model fine-tuned on medical text to perform sequence labeling on the user's question, identifying key medical entities. For example, it accurately identifies the two symptom entities "fever" and "red eyes" from the question. Simultaneously, the intent recognition module uses a text classification model (also based on BERT) to classify the entire question into a preset intent category, such as "diagnosis consultation," "treatment plan consultation," or "cause inquiry." In this example, the question is identified as having a "diagnosis consultation" intent, and the result is displayed accordingly. The structured data, consisting of {Intent: Diagnostic Consultation, Entities: [Fever, Red Eyes]}, is input into the graph query module. The graph query module uses an inference engine to start with the identified entities (such as "fever" and "red eyes") and searches the knowledge graph for common disease nodes connected to these entities. Then, a multi-hop inference module deduces the disease nodes connected to these entities. During this process, the multi-hop inference module does not only search for direct connections between disease nodes but also performs multi-hop queries. For example, it finds "Kawasaki disease" from "fever" and then uses the HasSymptom relation from "Kawasaki disease" to find other related symptoms such as "rash" and "strawberry tongue."Simultaneously, it also identifies "Kawasaki disease" from "red eyes." When multiple paths point to "Kawasaki disease," the weight (confidence) of that node is increased. The candidate answer generation module then aggregates all high-weight disease nodes and their confidence levels and reasoning paths (i.e., all nodes and edges traversed from the question entity to the answer disease) through the engine, forming structured candidate answers, such as: {Disease: Kawasaki disease, Confidence: 0.95, Path: [Fever -> Kawasaki disease], [Red eyes -> Kawasaki disease]}. In this process, multiple isolated symptom / sign entities identified in the user query are used as joint query conditions. Graph queries and multi-hop reasoning are performed in the knowledge graph to find the disease node that most reasonably explains all these symptoms and calculate its confidence. The specific algorithm and implementation path achieve a qualitative leap from information retrieval to clinical auxiliary reasoning. Subsequently, the answer refinement and generation module uses a finely tuned text generation model (such as T5). The system takes the candidate answers, reasoning paths, and relevant guideline excerpts stored in the knowledge graph as input from the candidate answer generation module to generate a fluent and easy-to-understand natural language answer, facilitating user comprehension. For example, "Persistent fever for more than 5 days accompanied by bilateral red eyes is one of the main diagnostic criteria for typical Kawasaki disease." Simultaneously, the credibility assessment submodule, based on the confidence level calculated by the candidate answer generation module and the completeness of the evidence chain in the knowledge graph, provides a comprehensive credibility score (e.g., 95%). The risk assessment submodule uses a pre-set risk rule base. When the risk assessment submodule identifies a specific high-risk entity (e.g., "fever longer than 5 days," "coronary artery dilation") or combination, it triggers an alarm. In this example, since "fever longer than 5 days" is a key warning signal for Kawasaki disease, the system automatically adds a strong risk warning before the final answer, such as "Warning: This symptom combination is highly suggestive of Kawasaki disease; please seek medical attention immediately!" Finally, the interpretability output submodule outputs the system's generated answer, risk warning, credibility score, and key reasoning evidence (e.g., "Based on: AHA guidelines") to the user.

[0022] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A Kawasaki disease intelligent triage system based on knowledge graph reasoning, characterized in that: The system includes a knowledge base construction layer, a question comprehension layer, a core reasoning layer, and an answer generation and evaluation layer; Knowledge base construction layer: It includes an entity extraction module, a relation extraction module, and a knowledge storage module, which are used to construct the Kawasaki disease knowledge graph; Question Comprehension Layer: Used to identify entities and nodes within the Kawasaki disease consultation information input by the user; Core reasoning layer: used to query, reason about, and output the symptom entities within the Kawasaki disease consultation information input by the user; Answer generation and evaluation layer: used to refine the reasoning results and assess the credibility and risk of Kawasaki disease.

2. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 1, characterized in that: The entity extraction module uses a pre-trained model such as BERT to recognize medical entities in text, the relation extraction module uses a pre-trained model such as BERT to identify the relationships between entities, and the knowledge storage module stores the extracted entities and relationships into a graph database to form a structured Kawasaki disease knowledge graph.

3. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 1, characterized in that: The question understanding layer includes a named entity recognition module and an intent recognition module.

4. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 3, characterized in that: The named entity recognition module uses a finely tuned BERT model on medical text to perform sequence annotation on user questions and identify key medical entities within them.

5. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 3, characterized in that: The intent recognition module uses a BERT-based text classification model to classify questions into preset intent categories.

6. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 1, characterized in that: The core reasoning layer includes a graph query module, a multi-hop reasoning module, and a candidate answer generation module. The graph query module uses the reasoning engine to find common disease nodes between entities in the knowledge graph, starting from the identified entities. The multi-hop reasoning module performs reasoning on multiple paths. The candidate answer generation module is for the engine to summarize disease nodes, their confidence levels, and reasoning paths.

7. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 6, characterized in that: The answer generation and evaluation layer includes an answer refinement and generation module and a credibility and risk assessment module. The answer refinement and generation module uses a fine-tuned text generation model, taking the data of generated candidate answers, the reasoning path of multi-hop reasoning, and relevant original text fragments stored in the knowledge storage as input, to automatically generate natural language answers.

8. The Kawasaki disease intelligent triage system based on knowledge graph reasoning according to claim 7, characterized in that: The credibility and risk assessment module includes a credibility assessment submodule, a risk assessment submodule, and an interpretability output submodule.

9. A triage method for a Kawasaki disease intelligent triage system based on knowledge graph reasoning, characterized in that: The intelligent triage system for Kawasaki disease based on knowledge graph reasoning as described in any one of claims 1-8 includes the following steps: Step 1: Construct a precise, authoritative, and verifiable knowledge graph of Kawasaki disease through the knowledge base construction layer; Step Two: Through the question comprehension layer and the core reasoning layer, a qualitative leap is achieved from information retrieval to clinical auxiliary reasoning; Step 3: Through the answer generation and evaluation layer, different levels of risk warnings are automatically triggered based on the answer content, and a quantitative credibility assessment is given. Finally, a comprehensive answer output mechanism that combines information, security and interpretability is generated, effectively avoiding medical delays caused by information asymmetry.

Images