Knowledge graph enhanced reasoning method and system for high-risk field medical decision

Abstract

The invention discloses a knowledge graph enhanced reasoning method and a knowledge graph enhanced reasoning system for high-risk field medical decision, which are used for strictly limiting the reasoning of a large language model within a high-quality knowledge graph range, so that the reliability is improved, and factual errors and'model illusion 'are fundamentally eliminated. Through an innovative constraint generation and traceability mechanism, each conclusion can be traced to a direct evidence in the knowledge graph, the interpretability is enhanced, and the complete transparency of the reasoning process is realized. The mixed query driven sub-graph retrieval method can deeply understand the intention of a user, accurately position and extract structured knowledge related to complex problems, improves the retrieval precision, and has a retrieval effect far better than that of traditional keyword or vector retrieval.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and is particularly applicable to knowledge graph-enhanced reasoning methods and systems for medical decision-making in high-risk fields. Background Technology

[0002] In recent years, artificial intelligence technologies, represented by Large Language Modeling (LLM), have shown broad application prospects in the medical field. These models can understand and generate natural language, assisting doctors in tasks such as medical record analysis and literature retrieval. However, existing technologies have the following technical shortcomings when applied to high-risk, high-reliability medical decision-making scenarios such as aviation and emergency rescue: 1. Factual Unreliability and "Model Illusion": Existing large language models are mainly based on statistical learning from massive amounts of text data, and the authenticity of the content they generate cannot be guaranteed. When dealing with medical issues that require precise factual evidence, the model may generate "illusionary" information that does not conform to the facts, which may lead to disastrous consequences in high-risk decision-making. 2. Lack of transparency in the reasoning process: The reasoning process of existing models is like a "black box," and users cannot understand the specific reasons and basis for reaching a certain conclusion. This lack of interpretability makes it difficult for domain experts to trust and adopt its output results, seriously hindering its application in serious medical scenarios. 3. Insufficient utilization of structured knowledge: Although knowledge graphs (KGs) can store accurate domain knowledge in a structured form, current technologies typically use them only as simple information query sources, failing to effectively utilize their rich semantic relationships and logical structures to guide and constrain the complex reasoning process of large language models. For example, traditional retrieval-enhanced generation (RAG) techniques often retrieve information from unstructured text, resulting in search results that frequently contain noise and fail to reveal the deep logic between knowledge points. Summary of the Invention

[0003] The purpose of this invention is to provide a knowledge graph-enhanced reasoning system and method for medical decision-making in high-risk fields, which solves the problem of deep integration between the language capabilities of large language models and the knowledge accuracy of knowledge graphs, and realizes reliable and interpretable intelligent medical decision-making, significantly improving the reliability, accuracy and interpretability of medical decision support. To achieve the above objectives, the knowledge graph-enhanced reasoning system for high-risk medical decision-making described in this invention includes a user interaction interface, a core engine, and a result output interface. The user interaction interface is used to receive natural language query commands from users and display text and image query results. The result output interface is used to integrate and render the query results processed by the core engine. The core engine includes a knowledge graph storage module, a hybrid query-driven subgraph retrieval module, and a constraint generation and tracing module. The knowledge graph storage module is used to store medical knowledge graphs in high-risk areas; The hybrid query-driven subgraph retrieval module extracts a knowledge subgraph from the high-risk domain medical knowledge graph based on the user's input natural language query requirements. The constraint generation and tracing module translates the knowledge subgraph into a format that the large language model can understand, and constrains the large language model to generate faithful and traceable query results. Furthermore, it can be deployed on a single server, a distributed server cluster, or a cloud platform.

[0004] Furthermore, the knowledge graph storage module adopts a graph database or a relational database that supports graph computation.

[0005] The knowledge graph-enhanced reasoning method for medical decision-making in high-risk domains described in this invention includes the following steps: Construct a knowledge graph; for specific high-risk areas, integrate multi-source heterogeneous data to construct a high-risk area medical decision-making knowledge graph containing domain entities, relationships, and attributes; Hybrid query-driven subgraph retrieval; receiving natural language query requests from users; parsing the query requests; identifying core entities and relational intents; based on the identification results, predicting relational paths connecting core entities in the knowledge graph of high-risk medical decision-making, and performing weighted graph traversal and pruning along the predicted paths; generating a knowledge subgraph highly relevant to the query requests; Constraint-based generation and tracing: The knowledge subgraph is linearized into structured text and combined with a preset prompt template to construct an input for a large language model, driving the large language model to reason based on the knowledge subgraph and generate a natural language answer containing machine-readable tracing identifiers; the tracing identifiers in the answer are mapped back to nodes and edges in the knowledge subgraph to form an interactive and visual evidence chain.

[0006] Furthermore, the process of parsing the query request specifically involves identifying the core entities in the user's natural language query instructions based on a named entity recognition component for a specific high-risk domain; the entity linking component ultimately selects the unique entity ID with the highest confidence from the medical decision-making knowledge graph in the high-risk domain as the link target; and using dependency parsing and keyword technology, the syntactic structure of the user's query input text is analyzed to identify the user's core relational intent.

[0007] Furthermore, the prediction of relational paths connecting core entities in the knowledge graph of medical decision-making in high-risk areas based on the identification results specifically includes, based on the identified core entities and relational intentions, outputting one or more relational paths that are logically most likely to connect the core entities, using rule templates or learning models.

[0008] Furthermore, the rule templates are a series of high-frequency, medically logical reasoning path templates predefined by experts in high-risk domains; the learning model is a lightweight classification or ranking model formed through training, which has a large number of existing real paths and can perform probabilistic ranking of all possible paths according to the type of the input entity.

[0009] Furthermore, weighted graph traversal and pruning are performed along the predicted path; generating a knowledge subgraph highly relevant to the query requirements specifically includes executing a biased graph traversal algorithm, prioritizing nodes located on the predicted path when selecting the next node to be expanded; dynamically assigning a comprehensive weight to each node during the traversal process; sorting all candidate nodes according to the comprehensive weight, and retaining only the top few candidate nodes with the highest comprehensive weight to form a knowledge subgraph.

[0010] Furthermore, the comprehensive weight , The semantic similarity score between the text description of node n and the user query Q is calculated. Let n be the centrality score of node n in the entire high-risk domain medical decision-making knowledge graph; The path matching score for node n; , , This is a hyperparameter and can be set based on experience.

[0011] Furthermore, the preset prompt template includes linearized structured text of the knowledge subgraph and the user's query requirements.

[0012] The advantages of this invention lie in its strict constraint of the reasoning of large language models within a high-quality knowledge graph, thereby improving reliability and fundamentally eliminating factual errors and "model illusions." Through innovative constraint generation and tracing mechanisms, each conclusion can be traced back to its direct evidence within the knowledge graph, enhancing interpretability and achieving complete transparency in the reasoning process. The hybrid query-driven subgraph retrieval method deeply understands user intent, accurately locates and extracts structured knowledge related to complex problems, improves retrieval accuracy, and achieves retrieval results far superior to traditional keyword or vector retrieval. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall architecture of the knowledge graph-enhanced reasoning system for medical decision-making in high-risk fields according to the present invention.

[0014] Figure 2 This is a flowchart illustrating the knowledge graph-enhanced reasoning method for medical decision-making in high-risk fields, as described in this invention.

[0015] Figure 3This is a schematic diagram of the output interface and interactive evidence chain of the knowledge graph-enhanced reasoning system for high-risk medical decision-making according to the present invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] The knowledge graph-enhanced reasoning system for high-risk medical decision-making described in this invention can be deployed on a single server, a distributed server cluster, or a cloud platform. For example... Figure 1 As shown, it includes a user interaction interface, a core engine, and a result output interface. The user interaction interface can be a webpage or an application interface, used to receive the user's natural language query commands and ultimately display the query results to the user in a graphic and textual format.

[0018] The core engine is the inventive core of the knowledge graph-enhanced reasoning system for high-risk medical decision-making described in this invention. It is the operating platform of the knowledge graph-enhanced reasoning method for high-risk medical decision-making described in this invention, and includes a knowledge graph storage module, a hybrid query-driven subgraph retrieval module, and a constraint generation and tracing module.

[0019] The knowledge graph storage module uses a graph database (such as Neo4j) or a relational database that supports graph computing to store high-risk domain medical knowledge graphs. The high-risk domain medical knowledge graph is pre-constructed by integrating multi-source heterogeneous data, such as electronic medical records, medical literature, expert questionnaires, and physiological monitoring data, using knowledge graph construction methods. This involves a series of steps including entity recognition, relation extraction, entity linking, and knowledge fusion. The specific construction methods are well-known to those skilled in the art and will not be elaborated upon here.

[0020] The hybrid query-driven subgraph retrieval module precisely extracts a small, refined knowledge subgraph from the vast medical knowledge graph of high-risk areas based on the user's input natural language query requirements.

[0021] The constraint generation and traceability module is responsible for translating a small, refined knowledge subgraph from the hybrid query-driven subgraph retrieval module into a format that the large language model can understand, and constraining the large language model to generate faithful and traceable query results.

[0022] The output interface is responsible for integrating and rendering the faithful and traceable query results processed by the core engine, forming a natural language answer with traceability identifiers and a visual chain of evidence, which is then displayed to the user through the user interaction interface.

[0023] like Figure 2 As shown, the knowledge graph-enhanced reasoning method for medical decision-making in high-risk domains described in this invention includes the following steps: Constructing a knowledge graph. For specific high-risk domains, integrate multi-source heterogeneous data, such as electronic medical records, medical literature, expert questionnaires, and physiological monitoring data, to construct a knowledge graph for high-risk domain medical decision-making that includes domain entities, relationships, and attributes. The specific construction method is well-known to those skilled in the art and will not be elaborated further.

[0024] Hybrid query-driven subgraph retrieval. It receives natural language query requests from users; parses the query requests; identifies core entities and relational intents; based on the identification results, predicts relational paths connecting core entities in the knowledge graph of high-risk medical decision-making, and performs weighted graph traversal and pruning along the predicted paths; finally, it generates a knowledge subgraph highly relevant to the query request. The purpose of hybrid query-driven subgraph retrieval is to prepare sufficient and highly relevant reliable "raw materials" for answering user queries. Specifically, this includes the following components.

[0025] 1. Analyze the query requirements to identify the core entities and relational intents.

[0026] First, a NER (Named Entity Recognition) component based on a specific high-risk domain identifies the core entities in the user's natural language query. Then, an entity linking component ultimately selects the unique entity ID with the highest confidence from the high-risk domain's medical decision-making knowledge graph as the link target. Specifically: After receiving the core entities identified by the named entity recognition component, the entity linking component searches for the core entities and their aliases within the high-risk domain medical decision-making knowledge graph, finding a set of potential candidate entity IDs. Using a pre-trained language model, high-dimensional semantic feature encoding is performed on the entity mention context from the named entity recognition component and the authoritative descriptions of the candidate entities from the high-risk domain medical decision-making knowledge graph, respectively, and the deep semantic similarity between the two encoded vectors is calculated. Simultaneously, the high-risk domain medical decision-making knowledge graph reasoning submodule is activated to analyze the network association strength between the candidate entities and other identified entities in the text. Combining the above semantic signals (deep semantic similarity) and graph structure signals (network association strength), a learnable ranking model is used for fusion decision-making. Finally, each core entity is selected as the unique entity ID with the highest confidence in the high-risk domain medical decision-making knowledge graph, thus determining the core entity object in the user's query text and locating that core entity object to a specific node in the high-risk domain medical decision-making knowledge graph.

[0027] Next, the relationships between the core entities in the user's query are analyzed to identify relational intent. Dependency parsing and keyword techniques are used to analyze the syntactic structure of the user's query input text, identifying the user's core relational intent, such as exploring potential causal relationships or risk relationships. The identified relational intent is then mapped to a predefined relational type within a high-risk domain medical decision-making knowledge graph.

[0028] 2. Based on the identification results, predict the relationship paths connecting core entities in the knowledge graph of medical decision-making in high-risk areas. Specifically, based on the identified core entities and relationship intentions, and using rule templates or learning models, output one or more logically most likely relationship paths connecting the aforementioned entities. The rule templates are a series of high-frequency, medically logical reasoning path templates predefined by experts in high-risk domains. The learning model, on the other hand, is a lightweight classification or ranking model formed through training, which already contains a large number of real-world paths and can probabilistically rank all possible paths based on the type of the input entity.

[0029] 3. Perform weighted graph traversal and pruning along the predicted relationship path to generate a knowledge subgraph highly relevant to the query requirements. Specifically, starting from the core entity, execute a biased graph traversal algorithm, such as Best-First Search. When selecting the next node to expand, prioritize nodes located on the predicted relationship path connecting the core entity from the previous step. Dynamically assign a comprehensive weight to each node during the traversal. Stop when the traversal reaches a preset depth (e.g., 3 hops) or the number of candidate nodes reaches a threshold. Sort all candidate nodes according to the comprehensive weight, retaining only the top few nodes (e.g., the top 20 nodes) and the connecting edges between nodes, forming a knowledge subgraph highly relevant to the query requirements.

[0030] The overall weight , To determine the semantic similarity between the text description of node n and the user's query requirement Q, models such as Sentence-BERT can be used to calculate the cosine similarity of the text embeddings. Let n be the center score (e.g., PageRank) of node n in the entire high-risk domain medical decision-making knowledge graph, representing its importance within the domain. The path matching score is assigned to node n. A higher score is given if node n is on the predicted relational path connecting the core entities. , , This is a hyperparameter and can be set based on experience.

[0031] Constraint-based generation and sourcing: The knowledge subgraph is linearized into structured text to facilitate processing by the large language model. Input to the large language model is constructed using a pre-defined prompt template, driving the model to reason based solely on the knowledge subgraph. The prompt template includes the linearized structured text of the knowledge subgraph and the user's query requirements. Under the strong constraints of the prompt template, the large language model generates a logically clear natural language answer entirely based on the facts provided by the knowledge subgraph. Simultaneously, regular expressions are used to extract all referenced knowledge IDs from the natural language answer. These knowledge IDs are matched with knowledge fragments recorded during the linearization of the knowledge subgraph to accurately identify each knowledge ID in the natural language answer. This determines which nodes and edges in the knowledge subgraph correspond to, ensuring that the final natural language answer is not only logically clear and entirely based on the facts provided by the knowledge subgraph, but also includes machine-readable sourcing identifiers.

[0032] Alongside displaying the final natural language answer to the user, a graph data JSON object with highlighted information is also presented to parse the source identifiers in the answer. A graph visualization library (such as D3.js or Cytoscape.js) is used to render the knowledge subgraph, and an event listener (such as onHover) is added to each traceable knowledge ID in the natural language answer. When the user's mouse hovers over a knowledge ID, the corresponding node and edge in the graph data are highlighted, thus realizing an interactive chain of evidence that links text and graph, such as... Figure 3 As shown.

Claims

1. A knowledge graph enhanced reasoning system for high-risk field medical decision-making, comprising a user interaction interface, a core engine and a result output interface. The user interaction interface is used for receiving natural language query instructions of the user and displaying query results in the form of text and images; The result output interface is used for integrating and rendering the query results processed by the core engine; and the core engine comprises a knowledge graph storage module, a mixed query driven subgraph retrieval module and a constraint generation and tracing module; The knowledge graph storage module is used for storing a high-risk field medical knowledge graph; The mixed query driven subgraph retrieval module is used for cutting out a knowledge subgraph from the high-risk field medical knowledge graph according to the natural language query requirement input by the user; The constraint generation and tracing module translates the knowledge subgraph into a format that can be understood by a large language model, and constrains the large language model to generate a query result that is faithful and traceable.

2. The knowledge graph enhanced reasoning system for high-risk field medical decision-making according to claim 1, wherein: The system can be deployed on a single server, a distributed server cluster or a cloud platform. 3.The knowledge graph enhanced reasoning system for high-risk field medical decision-making according to claim 1, characterized in that: The knowledge graph storage module adopts a graph database or a relational database supporting graph calculation.

4. A knowledge graph enhanced reasoning method for medical decision in high-risk fields, characterized in that, The system comprises the following steps: Building a knowledge graph; for a specific high-risk field, integrating multi-source heterogeneous data to build a high-risk field medical decision knowledge graph containing field entities, relationships and attributes; Mixed query driven subgraph retrieval; Receiving a natural language query requirement input by the user; Analyzing the query requirement; identifying core entities and relationship intentions; Based on the identification result, predicting a relationship path connecting the core entities in the high-risk field medical decision knowledge graph, and performing weighted graph traversal and pruning along the predicted path; Generating a knowledge subgraph highly related to the query requirement; Constraint generation and tracing; linearizing the knowledge subgraph into structured text, combining a preset prompt template to build an input for the large language model, driving the large language model to perform reasoning based on the knowledge subgraph, generating a natural language answer containing machine-readable traceable identifiers; mapping the traceable identifiers in the answer back to the nodes and edges in the knowledge subgraph to form an interactive visual evidence chain.

5. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 4, characterized in that: The analysis of the query requirement specifically comprises identifying core entities in the natural language query instructions of the user based on a named entity recognition component specific to the high-risk field; using an entity linking component to finally screen out the unique entity ID with the highest confidence as the linking target from the high-risk field medical decision knowledge graph; using dependency syntax analysis and keyword technology to analyze the syntax structure of the user query input text and identify the core relationship intention of the user.

6. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 4, characterized in that: The prediction of the relationship path connecting the core entities in the high-risk field medical decision knowledge graph based on the identification result specifically comprises outputting one or more relationship paths that are most likely to connect the core entities in logic based on a rule template or a learning model according to the identified core entities and relationship intentions.

7. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 6, characterized in that: The rule template is a series of high-frequency, medically logical reasoning path templates predefined by experts in the high-risk field; The learning model is a lightweight classification or sorting model that has existed a large number of real paths and can sort all possible paths according to the input entity type.

8. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 4, characterized in that: The weighted graph traversal and pruning are performed along the predicted path; and generating a knowledge subgraph highly relevant to the query requirement specifically includes: performing a biased graph traversal algorithm, and preferentially selecting a node located on the predicted path when selecting a next node to be expanded; dynamically assigning a comprehensive weight to each node in the traversal process; and sorting all candidate nodes according to the comprehensive weights, and only retaining a plurality of candidate nodes with the highest comprehensive weights to form the knowledge subgraph.

9. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 8, characterized in that: The comprehensive weight , is a semantic similarity score of the text description of the node n and the user query requirement Q; is a centrality score of the node n in the entire high-risk field medical decision knowledge graph; is a path matching score of the node n; , , is a hyperparameter, which can be set according to experience.

10. The knowledge graph enhanced reasoning method for high-risk field medical decision-making according to claim 4, characterized in that: The preset prompt template includes a structure text linearized from the knowledge subgraph and the query requirement of the user.

Images