Construction and retrieval enhancement method and device based on medical insurance knowledge graph

Abstract

The invention provides a construction and retrieval enhancement method and device based on a medical insurance knowledge graph, and relates to the field of retrieval. Taking a target relationship among the target drug chemical component pharmacological attribute, the target medical insurance reimbursement rule and the target fixed-point medical institution distribution related to the anchor node as an edge correspondingly connected with the anchor node, and setting a domain label corresponding to the target relationship for the edge, according to the pharmacological attributes of the target drug chemical components and the target medical insurance reimbursement rule, adding an aging window on the attributes of the edges, and obtaining a medical insurance knowledge graph constructed based on the field sub-graph; analyzing the query request by using the classification model, identifying multiple user intentions of the query request, and matching the user intentions with the domain tags to obtain multiple target domain tags matched with the user intentions; and according to the target domain label, disassembling the query request into atomized query data corresponding to the target domain sub-graphs for different domains.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of retrieval, in particular to a construction and retrieval enhancement method and device based on a medical insurance knowledge graph. BACKGROUND

[0002] At present, the construction of a vertical field knowledge graph usually adopts a general modeling and storage method centered on entities, that is, all types of relationships (such as pharmacological properties of chemical components of a drug, medical insurance reimbursement rules, and distribution of designated medical institutions) are connected to the same entity node as equal edges, forming a high-density monomer network. When the existing RAG system performs a knowledge graph retrieval task, it usually locates a central entity based on semantic similarity or keyword matching, and then uniformly expands and recalls all associated information around the entity through a "k-hop" neighbor query. The retrieval result data accuracy of this knowledge graph retrieval method is low. SUMMARY

[0003] The purpose of the present application is to provide a construction and retrieval enhancement method and device based on a medical insurance knowledge graph to solve the technical problem of low retrieval result data accuracy of the current medical insurance retrieval method.

[0004] In a first aspect, the present application provides a construction and retrieval enhancement method based on a medical insurance knowledge graph, comprising: Obtaining medical insurance data containing pharmacological properties of chemical components of a drug, medical insurance reimbursement rule data, and distribution of designated medical institutions, establishing anchor nodes across fields in a graph database, taking target relationships between target pharmacological properties of chemical components of a drug, target medical insurance reimbursement rules, and target distribution of designated medical institutions related to the anchor nodes as edges connected by the anchor nodes, setting field labels corresponding to the target relationships for the edges, adding a time window for the target pharmacological properties of chemical components of a drug and the target medical insurance reimbursement rules on the attributes of the edges, and obtaining a medical insurance knowledge graph constructed based on a field subgraph; the anchor nodes include at least one of a drug, a person, and a medical institution; the field labels include at least one of a policy field, a resource field, and a medical field; In response to a user's query request, a classification model is used to analyze the query request and identify multiple user intents of the query request, the user intents are matched with the field labels, and multiple target field labels matching the user intents are obtained; The query request is decomposed into atomic query data for target domain subgraphs corresponding to different domains based on the target domain tags. Time parameters are injected into the atomic query data based on the current time and the time data in the query request, resulting in multiple parallel sub-retrieval tasks. Shortcut edge retrieval calculations and time window compliance filtering are performed in the medical insurance knowledge graph based on the sub-retrieval tasks, the target domain subgraphs, and the time parameters. Pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results. Based on the multiple pre-calculated shortcut edge results, they are fused in memory through logical intersection operations to generate the final query result.

[0005] In one possible implementation, the time window is an effective interval attribute introduced on the relation edges of the graph to construct a slowly changing dimension SCD model in the medical insurance knowledge graph; the multiple target domain labels are labels corresponding to multiple target domains identified by analyzing the user's natural language corresponding to the query request using an intent router before RAG retrieval; The process involves injecting time parameters into the atomic query data based on the current time and the time data in the query request, resulting in multiple parallel sub-retrieval tasks. Based on these sub-retrieval tasks, the target domain subgraph, and the time parameters, shortcut edge retrieval calculations and timeliness window compliance filtering are performed in the medical insurance knowledge graph, including: When generating a database query statement, the current timestamp and / or the user-specified timestamp in the query request are automatically inserted as variables into the WHERE clause. The current timestamp and / or the user-specified timestamp are injected as filter parameters into the Prompt or code generated by the retrieval statement, resulting in multiple sub-retrieval tasks that are executed in parallel. Based on the sub-retrieval task, the target domain subgraph, and the time parameter, the target domain subgraph is dynamically activated in the medical insurance knowledge graph for retrieval. The retrieval calculation of shortcut edges and the compliance filtering of the timeliness window are performed by calling the query interface corresponding to the target domain subgraph.

[0006] In one possible implementation, the results of multiple pre-computed shortcut edges are fused in memory through a logical intersection operation to generate the final query result, including: The first pre-computed shortcut edge result is detected by the fusion engine. If the corresponding drug in the first pre-computed shortcut edge result is reimbursable, the hospital data corresponding to the first pre-computed shortcut edge result is combined with the policy data corresponding to the second pre-computed shortcut edge result by the fusion engine to construct a structured context. Based on the structured context, natural language responses are generated through LLM. According to the high-frequency inference path that appears more frequently than a specified frequency in the asynchronous analysis query log of the fusion engine, a new shortcut edge is established between the two nodes corresponding to the high-frequency inference path. Based on the pharmacological properties of the target drug's chemical components and the target medical insurance reimbursement rules in the high-frequency inference path, a timeliness window is added to the attributes of the new shortcut edge to generate the fixed shortcut edge in the medical insurance knowledge graph.

[0007] In one possible implementation, the step of performing shortcut edge retrieval calculations and time-sensitive window compliance filtering in the medical insurance knowledge graph based on the sub-retrieval task, the target domain subgraph, and the time parameter, and determining the pre-calculated shortcut edge results based on the frequency of occurrence of the retrieval calculation results, includes: Based on the sub-retrieval task, the target domain subgraph, and the time parameter, the shortcut edge retrieval calculation and the timeliness window compliance filtering are performed in the medical insurance knowledge graph using the following formula, and the pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results:

[0008] in, This indicates the co-occurrence frequency of retrieval node A and retrieval node B in the historical query data within the sub-retrieval task; This represents the shortest path length from retrieval node A to retrieval node B in the target domain subgraph. This represents the time decay coefficient, used to control the rate at which the correlation strength decays when the time deviates from the center of the aging window; This represents the time parameter in the query request; This refers to the center time of the time window set on the shortcut edge between search node A and search node B; This represents the pre-calculated association strength value of the shortcut edge.

[0009] In one possible implementation, after the step of obtaining the healthcare knowledge graph constructed based on the domain subgraph, the following is also included: In response to a medical insurance guidance request for a target user, the user's pathological data is determined based on the target user's historical medical insurance data through the domain subgraph of the medical field in the medical insurance knowledge graph. Based on the user's pathological data and the specified seasonal alternation pattern, the medication rules for the target user are generated. Based on the medication rules, the medical prescription data corresponding to the target user, and the physical examination results of the target user, the domain subgraph of the policy domain in the medical insurance knowledge graph is used to predict the individual medication to be taken and the medication dosage adjustment method for the target user. Based on the target user and the method of drug dosage adjustment, the AI ​​medical system generates medical insurance guidance data for the target user using the domain subgraph of the resource domain in the medical insurance knowledge graph.

[0010] In one possible implementation, after predicting the individual medication recipients and dosage adjustment methods for the target user based on the medication rules and medical prescription data through the policy domain subgraph in the health insurance knowledge graph, the method further includes: Based on the target users corresponding to the drugs to be taken and the drug dosage adjustment methods, the AI ​​medical system is used to predict the drug demand of various target users through the domain subgraphs of the resource domain and policy domain in the medical insurance knowledge graph, and the drug demand prediction results are obtained. In response to a drug guidance request for a target medical institution in the distribution of designated medical institutions, drug quantity adjustment guidance data is generated for the target medical institution based on the target drug demand forecast results corresponding to the target medical institution and the historical drug output data corresponding to the current season in the specified seasonal alternation pattern, so as to carry out advance stockpiling of drugs.

[0011] In one possible implementation, after predicting the individual medication recipients and dosage adjustment methods for the target user based on the medication rules, the target user's corresponding medical prescription data, and the target user's physical examination results through the policy domain subgraph in the medical insurance knowledge graph, the following steps are also included: Based on the medication rules, the target patient, the medication dosage adjustment method, and the physical examination results corresponding to the target user, the AI ​​medical system uses the domain subgraph of the medical field in the medical insurance knowledge graph to predict the target user's chronic disease medication use, and obtains the target user's chronic disease medication prediction result. Based on the ratio data between the predicted chronic disease medication use results and the current chronic disease medication dosage for multiple target users, the domain subgraph of the resource domain in the medical insurance knowledge graph is updated to obtain the updated medical insurance knowledge graph for the newly added chronic disease situation.

[0012] In one possible implementation, the step of generating medical insurance guidance data for the target user based on the subject of the medication to be taken and the medication dosage adjustment method, using a domain subgraph of the resource domain in the medical insurance knowledge graph and an AI medical system, includes: Based on the target user's historical medication data, an AI medical system generates the target user's personal medication habit data; the personal medication habit data includes at least one of the following: drug price, drug manufacturer, drug model / version, and drug source; Based on the personal medication habit data, the individuals to be taken the medication, and the medication dosage adjustment methods, the AI ​​medical system generates personalized medical insurance guidance data for the target users through the domain subgraph of the resource domain in the medical insurance knowledge graph.

[0013] Secondly, this application provides a device for constructing and enhancing retrieval based on a medical insurance knowledge graph, including: A construction module is used to acquire medical insurance data including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and the distribution of designated medical institutions. Cross-domain anchor nodes are established in a graph database. The target relationships between the target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target designated medical institution distributions related to the anchor nodes are used as edges connecting the anchor nodes. Domain labels corresponding to the target relationships are set on the edges. Time windows are added to the attributes of the edges according to the target drug chemical composition and pharmacological attributes and the target medical insurance reimbursement rules, resulting in a medical insurance knowledge graph constructed based on domain subgraphs. The anchor nodes include at least one of drugs, personnel, and medical institutions; the domain labels include at least one of policy domains, resource domains, and medical domains. The identification module is used to respond to a user's query request, analyze the query request using a classification model and identify multiple user intentions of the query request, match the user intentions with the domain tags, and obtain multiple target domain tags that match the user intentions; The retrieval module is used to decompose the query request into atomic query data for target domain subgraphs corresponding to different domains based on the target domain tags. It injects time parameters into the atomic query data based on the current time and time data in the query request, resulting in multiple parallel sub-retrieval tasks. Based on the sub-retrieval tasks, the target domain subgraphs, and the time parameters, it performs shortcut edge retrieval calculations and timeliness window compliance filtering in the medical insurance knowledge graph. It determines pre-calculated shortcut edge results based on the frequency of the retrieval calculation results, and fuses these pre-calculated shortcut edge results in memory through logical intersection operations to generate the final query result.

[0014] Thirdly, this application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method described in the first aspect above.

[0015] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in the first aspect above.

[0016] This application brings the following beneficial effects: This application provides a method and apparatus for constructing and enhancing retrieval based on a medical insurance knowledge graph. It can acquire medical insurance data including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and the distribution of designated medical institutions. Cross-domain anchor nodes are established in a graph database. The target relationships between the target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target designated medical institution distributions related to the anchor nodes are used as edges connecting the anchor nodes. Domain labels corresponding to the target relationships are set on the edges. Time windows are added to the attributes of the edges based on the target drug chemical composition and pharmacological attributes and the target medical insurance reimbursement rules, resulting in a domain-based sub-database. A medical insurance knowledge graph is constructed using graphs; the anchor nodes include at least one of drugs, personnel, and medical institutions; the domain labels include at least one of policy domains, resource domains, and medical domains; in response to user query requests, a classification model is used to analyze the query request and identify multiple user intentions of the query request, and the user intentions are matched with the domain labels to obtain multiple target domain labels that match the user intentions; based on the target domain labels, the query request is decomposed into atomic query data for target domain subgraphs corresponding to different domains, and time is injected into the atomic query data based on the current time and the time data in the query request. The parameters are used to obtain multiple parallel sub-retrieval tasks. Based on these sub-retrieval tasks, the target domain subgraph, and the time parameter, shortcut edge retrieval calculations and time-sensitive window compliance filtering are performed on the medical insurance knowledge graph. Pre-calculated shortcut edge results are determined based on the frequency of the retrieval calculation results. These pre-calculated shortcut edge results are then fused in memory through logical intersection operations to generate the final query result. In this scheme, by logically dividing the graph into independent subgraphs for policy, resources, medicine, and other fields, and combining intent recognition for routing retrieval, it ensures that when a user queries a specific domain, the system only recalls the relevant subgraphs for that domain, completely blocking... This eliminates noise interference from other domains, thereby significantly improving the accuracy and relevance of RAG-generated answers and achieving a significant improvement in the retrieval signal-to-noise ratio. Moreover, by introducing query decomposition and hotspot shortcut edge mechanisms, for complex cross-domain problems, the system decomposes requests into parallel small queries and merges them in memory. For frequently occurring cross-domain query patterns, the system automatically builds shortcut edges or cache indexes. This strategy of combining computing power for flexibility and storage for speed effectively solves the performance bottleneck caused by super node traversal, significantly reduces the latency of complex queries, and improves throughput. Furthermore, a time-interval-based version control mechanism (Slowly Changing Dimensions) is introduced into the graph architecture.By adding an effective time window to the edge attributes and injecting the current time filter condition by default into the retrieval layer, the system can accurately distinguish between currently valid and historically invalid information, ensuring that the medical insurance information obtained by users is accurate and reliable in the time dimension, thus solving the technical problem of low accuracy of retrieval results in current medical insurance retrieval methods.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the construction and retrieval enhancement method based on medical insurance knowledge graph provided in this application embodiment; Figure 2 Another flowchart illustrating the construction and retrieval enhancement method based on medical insurance knowledge graph provided in this application embodiment; Figure 3 An example of a knowledge graph data model in the construction and retrieval enhancement method based on medical insurance knowledge graph provided in the embodiments of this application; Figure 4 An example of a system architecture diagram in the construction and retrieval enhancement method based on medical insurance knowledge graph provided in the embodiments of this application; Figure 5 A schematic diagram of the structure of a medical insurance knowledge graph-based construction and retrieval enhancement device provided in this application embodiment; Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this application, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0022] Currently, in a graph containing the entity "pharmaceuticals," this entity may simultaneously connect to medical relationships describing chemical components, relationships describing reimbursement ratios, and resource relationships describing the location of medical facilities. When existing RAG systems perform retrieval tasks, they typically locate the central entity (anchor point) based on semantic similarity or keyword matching, and then expand and recall all related information around this entity indiscriminately through "k-hop" neighbor queries. Subsequently, this information, which mixes multiple domain attributes, is uniformly input as context into the LLM, allowing the model itself to filter and generate answers. This "unified" graph construction and retrieval model has a certain degree of universality in handling simple question-and-answer tasks and is currently the mainstream implementation solution in the industry.

[0023] While existing technical solutions can achieve basic knowledge retrieval in general scenarios, their design philosophy of "single-entity graph, full-scale recall" has fundamental shortcomings when facing vertical fields with complex logic and huge differences in data dimensions, such as medical insurance. First, existing technologies suffer from severe semantic noise and low signal-to-noise ratio issues in the retrieval context. Because current graphs do not logically segment relationships along domain dimensions, when a user only queries "the reimbursement rate for a certain drug under medical insurance" (policy domain intent), the system often recalls a large amount of irrelevant information such as the drug's "chemical structure," "side effects," and "list of all hospitals selling it." This redundant data not only wastes the valuable context window length of the LLM but also introduces a significant amount of interference noise, easily leading the LLM to retrieve incorrect information, creating "illusions," or providing answers that deviate from the user's intent.

[0024] Secondly, existing technologies suffer from low retrieval efficiency and a lack of flexibility when handling complex cross-domain queries and high-frequency hotspots. In existing technologies, when faced with combined queries that cross resource and policy domains, such as "whether a certain top-tier hospital offers a reimbursable drug," the system typically needs to perform a full, multi-hop traversal of the vast graph. Since popular entities (such as common drugs) often connect to tens of thousands of institutional nodes, this traversal leads to the "supernode" problem, causing significant computational overhead and query latency. Furthermore, existing technologies lack a "shortcut" mechanism for high-frequency cross-domain query paths, resulting in the repeated calculation of the same complex reasoning process, making it difficult for the system's response speed to meet real-time interactive requirements.

[0025] Finally, existing technologies struggle to effectively address the issues of "timeliness" and "version control" of knowledge. Medical insurance policies have strict effective cycles (e.g., annual updates), while medical knowledge is relatively static. Existing technologies typically store these two types of data together, lacking fine-grained time-series management. When policies are updated, simple add / delete operations can easily lead to the loss of historical data, or the inability to distinguish between "currently effective" and "archived" data during retrieval, causing the system to incorrectly reference outdated policy data when users inquire about current reimbursement policies.

[0026] Based on this, this application provides a method and apparatus for constructing and enhancing retrieval based on a medical insurance knowledge graph. This method can solve the technical problem of low accuracy of retrieval results in current medical insurance retrieval methods.

[0027] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating a method for constructing and enhancing retrieval based on a medical insurance knowledge graph, provided as an embodiment of this application. Figure 1 As shown, the method includes: Step S110: Obtain medical insurance data including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and distribution of designated medical institutions. Establish cross-domain anchor nodes in the graph database. Use the target relationships between the target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target distribution of designated medical institutions related to the anchor nodes as the edges connected to the anchor nodes. Set domain labels corresponding to the target relationships on the edges. Add time windows to the attributes of the edges according to the target drug chemical composition and pharmacological attributes and target medical insurance reimbursement rules to obtain a medical insurance knowledge graph constructed based on the domain subgraph.

[0029] Anchor nodes include at least one of pharmaceuticals, personnel, and medical institutions; domain labels include at least one of policy domains, resource domains, and medical domains.

[0030] Background of this application's embodiments: Suppose we need to build a "Smart Medical Insurance Question-Answer System" for a certain city. The knowledge graph contains entities such as "Amoxicillin," "Peking Union Medical College Hospital," and related medical insurance policy documents. A user raises a typical cross-domain complex question: "Which tertiary hospital in Chaoyang District can reimburse amoxicillin?" The system will process this request according to the steps provided in this application's embodiments.

[0031] For the construction of the aforementioned medical insurance knowledge graph, an example is, as follows: Figure 3 As shown, the system constructs a multi-dimensional, segmented time-series knowledge graph (data layer construction): This step aims to establish an underlying data structure that supports logical isolation and timeliness management, which is the foundation for subsequent accurate retrieval. This system constructs the graph in the graph database using a method combining anchor entities, domain labels, and timeliness attributes. Define Base Entities: Establish unique anchor nodes across domains. For example, Drug (ID: Drug_001, Name: "Amoxicillin") and Institution (ID: Hosp_101, Name: "Peking Union Medical College Hospital"). Construct Domain Subgraphs: Instead of directly mixing all edges, domain labels are added to the edges: Policy Domain: Stores [:BELONGS_TO_POLICY] relationships. Resource Domain: Stores [:IN_STOCK] and [:LOCATED_IN] relationships. Inject Temporal Attributes: For policy-related relationships, lifecycle attributes must be included. For example: (Amoxicillin) - [:BELONGS_TO {valid_from:'2024-01-01', valid_to:'2099-12-31', status:'ACTIVE'}] -> (2024 National Reimbursement Drug List). For expired historical policies, their valid_to is truncated to the expiration date, and the status is marked as 'EXPIRED', achieving soft deletion and historical archiving. A domain subgraph refers to a subset of entities and relationships logically divided from the complete knowledge graph based on business attributes (such as policy, resources, and medicine). A base entity refers to a common anchor node (such as a "drug ID") spanning different subgraphs, used to connect information from different domains, and is a key hub for achieving cross-domain integration.

[0032] Step S120: In response to the user's query request, the query request is analyzed using a classification model and various user intents of the query request are identified. The user intents are matched with domain labels to obtain various target domain labels that match the user intents.

[0033] For the aforementioned user intent identification, an example is intent identification and hotspot shortcut routing (L1 retrieval layer): This step aims to initially triage user queries, prioritizing high-frequency hotspots to achieve extremely fast response. When a user enters a question, such as... Figure 2 and Figure 4 As shown, the Intent Router first intervenes: Hotspot matching: The system queries the Redis cache or graph database for shortcut edges. The system checks if there are any high-frequency pre-calculated results for "Amoxicillin + Chaoyang District + Reimbursement + Top-tier Hospital". Judgment logic: If a pre-calculated edge exists in the system: If (Drug_001)-[:FAST_RECOMMEND{criteria:'Chaoyang+Sanjia+Reimbursement'}]->(Hosp_List), and the TTL (Time to Live) of this edge has not expired, then return the result directly and skip all subsequent steps.

[0034] The above user intent classification: If the hotspot is not matched, the router uses a lightweight classification model to analyze the intent. In this embodiment, the user intent is identified as including "find location (resource domain)" and "confirm reimbursement (policy domain)," which is a cross-domain combined query. Therefore, the request is forwarded to the "query decomposer."

[0035] The aforementioned Intent Routing refers to the process of analyzing the user's natural language before Retrieval Enhancement Generation (RAG) retrieval to determine which domains are involved, thereby deciding which subgraph query interfaces to invoke. The aforementioned Shortcut Edge refers to a pre-computed edge that directly connects two non-directly adjacent nodes, its purpose being to skip complex intermediate reasoning paths and directly obtain the results of high-frequency queries. Retrieval Enhancement Generation (RAG) technology effectively solves the problems of lagging model knowledge updates and lack of domain-specific knowledge by injecting external knowledge base retrieval results into the context window of a Large Language Model (LLM). In professional vertical fields such as healthcare, insurance, and finance, Knowledge Graphs (KG), due to their structured semantic representation capabilities, are often used as the core data foundation of RAG systems.

[0036] Step S130: Based on the target domain label, the query request is decomposed into atomic query data for the target domain subgraph corresponding to different domains. Time parameters are injected into the atomic query data according to the current time and the time data in the query request to obtain multiple parallel sub-retrieval tasks. Based on the sub-retrieval tasks, target domain subgraphs and time parameters, shortcut edge retrieval calculation and time window compliance filtering are performed in the medical insurance knowledge graph. The pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results. Based on the multiple pre-calculated shortcut edge results, they are fused in memory through logical intersection operation to generate the final query result.

[0037] In one possible implementation, the aforementioned time window is an effective interval attribute introduced on the relation edges of the graph to construct a slowly changing dimension SCD model in the medical insurance knowledge graph; the multiple target domain labels are labels corresponding to multiple target domains identified by analyzing the user's natural language corresponding to the query request before RAG retrieval using an intent router; Based on the current time and the time data in the query request, a time parameter is injected into the atomic query data to obtain multiple parallel sub-retrieval tasks. Then, based on the sub-retrieval tasks, the target domain subgraph, and the time parameter, shortcut edge retrieval calculations and time-sensitivity filtering are performed in the medical insurance knowledge graph. Specifically, this may include the following steps: When generating database query statements, the current timestamp and / or the user-specified timestamp in the query request are automatically inserted as variables into the WHERE clause. The current timestamp and / or the user-specified timestamp are injected as filter parameters into the Prompt or code generated by the retrieval statement, resulting in multiple sub-retrieval tasks that are executed in parallel. Based on the sub-retrieval task, the target domain subgraph, and the time parameter, the target domain subgraph is dynamically determined and activated in the medical insurance knowledge graph for retrieval. The retrieval of shortcut edges and the compliance filtering of the time window are performed by calling the query interface corresponding to the target domain subgraph.

[0038] For the above-mentioned decomposition and retrieval process, for example, such as Figure 2 and Figure 4As shown, the system employs time-sensitive query decomposition and parallel retrieval (L2 retrieval layer): It decomposes complex natural language problems into atomic queries targeting different subgraphs and injects time parameters. The implementation scheme uses a query decomposer to break down the problem into two parallel subtasks: Subtask A (resource domain retrieval): "Which tertiary hospitals in Chaoyang District offer amoxicillin?"; The generated Cypher query is: Cypher; MATCH (d:Drug {name:'Amoxicillin'})-[:IN_STOCK]->(h:Hospital); WHERE h.district = 'Chaoyang District' AND h.level = 'tertiary'; RETURN h.id; Subtask B (Policy Domain Retrieval): "What is the current reimbursement policy for amoxicillin?", time-sensitive filtering: The system automatically obtains the current system time (e.g., 2025-11-25) and injects it into the query to filter out expired policies. The generated Cypher query: Cypher; WITH date('2025-11-25') AS today; MATCH (d:Drug {name:'Amoxicillin'})-[r:BELONGS_TO_POLICY]->(p:Policy); WHERE r.valid_from<= today AND r.valid_to>=today / / Strict time window filtering; RETURN p.ratio, p.type; The system executes the above two queries in parallel. Assume subtask A returns [Central Hospital, Chaoyang Hospital], and subtask B returns [Category A, 80% reimbursement].

[0039] In one optional implementation, the above-mentioned retrieval calculation of shortcut edges in the medical insurance knowledge graph based on sub-retrieval tasks, target domain subgraphs, and time parameters, as well as the conformity filtering of the time window, and the determination of pre-calculated shortcut edge results based on the frequency of occurrence of the retrieval calculation results, may specifically include the following steps: Based on the sub-retrieval task, target domain subgraph, and time parameter, the following formula is used to perform shortcut edge retrieval calculation and timely window compliance filtering in the medical insurance knowledge graph, and the pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results:

[0040] in, This indicates the co-occurrence frequency of retrieval node A and retrieval node B in the historical query data within the sub-retrieval task; This represents the shortest path length from node A to node B in the target domain subgraph. This represents the time decay coefficient, used to control the rate at which the correlation strength decays when the time deviates from the center of the aging window; This represents the time parameter in the query request; This indicates the center time of the time window set on the shortcut edge between search node A and search node B; This represents the pre-calculated association strength value of the shortcut edge.

[0041] In this embodiment of the application, the calculation process of the above calculation formula enables more accurate data calculation and retrieval in the medical insurance knowledge graph based on the sub-retrieval task, the target domain subgraph, and the time parameter, making the calculation data of the pre-calculated shortcut edge results more accurate.

[0042] As an optional implementation, the above-mentioned results of multiple pre-computed shortcut edges are fused in memory through logical intersection operations to generate the final query result, which may specifically include the following steps: The fusion engine detects the results of the first pre-computation shortcut edge. If the corresponding drug in the first pre-computation shortcut edge result is reimbursable, the fusion engine combines the hospital data in the first pre-computation shortcut edge result with the policy data in the second pre-computation shortcut edge result to construct a structured context. Based on the structured context, natural language responses are generated through LLM. According to the high-frequency inference paths that appear more frequently than a specified frequency in the asynchronous analysis query log of the fusion engine, new shortcut edges are established between the two nodes corresponding to the high-frequency inference paths. Based on the chemical composition, pharmacological properties, and target medical insurance reimbursement rules of the target drug in the high-frequency inference paths, a timeliness window is added to the attributes of the new shortcut edges to generate solidified shortcut edges in the medical insurance knowledge graph.

[0043] For the fusion process of the aforementioned fusion engine, an example is as follows: Figure 2 and Figure 4 As shown, the memory fusion and answer generation (L3 generation layer) step performs a logical intersection operation on the scattered search results in memory and generates the final response. Implementation scheme: Logical judgment and fusion: The fusion engine first checks the results of subtask B. If the medicine is "not reimbursable," the process is terminated directly and the user is notified. In this example, the medicine is reimbursable. The engine combines the hospital list from subtask A with the policy information from subtask B to construct a structured context.

[0044] Context assembly: "Drug: Amoxicillin. Policy status: Currently in effect (2024 version), Category A, 80% reimbursement. Eligible resources: China-Japan Friendship Hospital (Grade A Tertiary Hospital / Chaoyang Hospital), Chaoyang Hospital (Grade A Tertiary Hospital / Chaoyang Hospital)." LLM Generation: The large language model generates a natural language response based on the clean context described above: "According to the currently effective 2024 version of the medical insurance policy, amoxicillin is a Class A drug with a reimbursement rate of approximately 80%. In Chaoyang District, you can not only get reimbursed, but the following top-tier hospitals currently have it in stock: China-Japan Friendship Hospital and Chaoyang Hospital." Asynchronous Construction of Cross-Domain Shortcuts (Closed-Loop Optimization): This step uses log analysis to solidify high-frequency inference paths into shortcuts in the graph, achieving "faster with use." Implementation: Log Mining: The system's backend asynchronously analyzes and queries logs, discovering that the pattern "drug + region + reimbursement" appeared more than 100 times in the past 24 hours. Shortcut Solidification (Materialization): The system triggers a batch processing task to establish a new shortcut edge [:RECOMMENDED_LOCATION] between the "Amoxicillin" node and a matching "Hospital" node. Setting Validity Period: The edge is set to expire_time = NOW() + 4 hours to ensure that if hospital inventory changes, the shortcut expires promptly, forcing the system to re-run the real-time query in step three.

[0045] By logically segmenting the graph into independent subgraphs for fields such as policy, resources, and medicine, and combining intent recognition for routing retrieval, the system ensures that when a user queries a specific field, it only recalls the relevant subgraph for that field, completely shielding noise interference from other fields. This significantly improves the accuracy and relevance of RAG-generated answers, achieving a significant improvement in the retrieval signal-to-noise ratio. Furthermore, by introducing query decomposition and hotspot shortcuts mechanisms, the system decomposes complex cross-domain problems into parallel small queries and merges them in memory. For frequently occurring cross-domain query patterns, the system automatically builds shortcut edges or cache indexes. This strategy, which combines computing power for flexibility and storage for speed, effectively solves the performance bottleneck caused by supernode traversal, significantly reduces the latency of complex queries, and improves throughput. Moreover, a time-interval-based version control mechanism (SlowlyChanging Dimensions) is introduced into the graph architecture. By adding an effective time window to the edge attributes and injecting the current time filter condition by default into the retrieval layer, the system can accurately distinguish between "currently valid" and "historically invalid" policy information, ensuring that the medical insurance information obtained by users is accurate and reliable in the time dimension, and guaranteeing the timeliness and accuracy of knowledge in a certain field.

[0046] This solution reconstructs the traditional flat knowledge graph into logically decoupled domain subgraphs and transforms passive full-scale retrieval into intent-based proactive routing retrieval and hotspot shortcut acceleration. Through a knowledge graph retrieval enhancement generation (RAG) method and system based on multi-dimensional subgraph segmentation and dynamic routing, the traditional monolithic, flat vertical domain knowledge graph is decoupled into logically independent "domain subgraphs" (such as policy domains and resource domains). Combined with query decomposition and hotspot shortcut mechanisms, it addresses noise interference and performance bottlenecks in complex cross-domain queries. Furthermore, it introduces time-series version control to ensure the timeliness of policy knowledge.

[0047] In this application embodiment: 1. Significantly improves the signal-to-noise ratio and accuracy of retrieval: Existing technologies typically perform full-graph retrieval, which easily introduces irrelevant noise (such as recalling pharmacological side effects when asking about reimbursement). This invention ensures that the retrieval process is strictly limited to a specific "domain subgraph" that matches the user's intent by "constructing a multi-dimensional segmented knowledge graph" and "intent recognition and dynamic routing". For example, when asking about policies, only the policy subgraph is searched, physically shielding the interference of inventory and medical knowledge, thereby eliminating the "illusion" caused by the cluttered context of LLM from the root. 2. Solves the performance bottleneck of complex cross-domain queries. Existing technologies often face high latency caused by "super node" traversal when dealing with cross-domain multi-hop queries that "want both A and B". This invention decomposes complex O(N*M) level association queries into two simple O(1) level search tasks and merges them in memory by "timeliness-based query decomposition", which greatly reduces the computational complexity. Simultaneously, by combining "asynchronous construction of cross-domain shortcuts," high-frequency inference results are solidified into physical edges, achieving a qualitative leap from "online computation" to "direct reading," significantly improving system throughput. 3. It perfectly solves the challenges of timeliness and version control of policy knowledge. Existing technologies often result in outdated responses due to the mixing of old and new data. This invention introduces a lifecycle-based temporal attribute design in the "data layer construction" and forcibly injects "current system time" as a filtering condition in the "retrieval layer." This mechanism ensures that the system can accurately distinguish between "currently effective" and "historically archived" policies, like a "time machine," guaranteeing both the accuracy of immediate question answering and the ability to trace history.

[0048] This solution breaks away from the inefficient "single graph, full retrieval" model of traditional vertical domain RAG systems, proposing a multi-dimensional subgraph architecture based on logical partitioning and a dynamic routing retrieval mechanism. This mechanism decouples complex cross-domain queries into parallel subtasks and introduces time-series version control and hotspot shortcuts, solving the three major challenges of "low signal-to-noise ratio," "poor timeliness," and "high latency in cross-domain retrieval" in complex domains such as medical insurance.

[0049] 1. A Dynamic Routing Retrieval Method Based on Intent Recognition for "Multi-Dimensional Subgraphs": This is the most fundamental architectural innovation of this invention. This invention protects a method for logically dividing a knowledge graph into multiple independent domain subgraphs (such as policy domains and resource domains), and a corresponding Intent Router. Its key feature is that the system does not traverse the entire graph, but dynamically determines which subgraph(s) to activate for retrieval based on the user's natural language query intent classification results. The aim is to protect this retrieval architecture of "physical / logical isolation storage + on-demand routing activation," which fundamentally shields against noise interference from irrelevant domains.

[0050] 2. Policy Version Control Retrieval Technology Based on Temporal Injection: Addressing the pain point of frequent updates to medical insurance policies, this invention protects a special time-series processing mechanism. Temporal injection refers to a technique where, when generating database query statements, the program automatically inserts the current timestamp or a user-specified timestamp as a variable into the WHERE clause to filter out expired data.

[0051] (1) Storage layer: Introduce the "Valid Interval" attribute on the relation edges of the graph to construct the "Slowly Changing Dimension (SCD)" model.

[0052] (2) Retrieval layer: Innovatively, "current system time" is forcibly injected as a filtering parameter into the Prompt or code generated by the retrieval statement.

[0053] (3) To protect this technical solution of "the storage layer retains all historical versions + the retrieval layer dynamically injects time windows", it ensures that the RAG system can accurately distinguish between "currently effective" and "historical archive" knowledge, and prevent the large model from generating the illusion of spatiotemporal confusion.

[0054] 3. Asynchronous Construction Mechanism for "Cross-Domain Hotspot Shortcuts" Based on Log Mining: This invention protects a self-optimizing system that "gets faster with use." This system analyzes user query logs to identify high-frequency cross-domain query patterns (such as "medicine + location + reimbursement") and automatically generates "shortcut edges" in the graph or builds "aggregate indexes" in the cache layer. This invention aims to protect this closed-loop acceleration method that "uses historical query logs to drive graph structure updates, solidifying complex multi-hop reasoning into single-hop direct reading."

[0055] 4. A Complex Reasoning Framework of "Query Decomposition - Parallel Retrieval - Memory Fusion": For complex cross-domain problems (requiring queries for both A and B), this invention protects a decomposition and fusion processing flow. Specifically, the natural language problem is decomposed into atomic queries targeting different subgraphs, the retrieval is executed in parallel, and finally, a logical intersection operation is performed on the heterogeneous results in memory. This "computing power for flexibility" processing flow is also a key protection point of this invention.

[0056] In some embodiments, after obtaining the medical insurance knowledge graph constructed based on the domain subgraph as described above, the method may further include the following steps: In response to a request for medical insurance guidance for a target user, the system determines the user's pathological data based on the target user's historical medical insurance data using the medical domain subgraph in the medical insurance knowledge graph. Based on the user's pathological data and specified seasonal patterns, it generates medication rules for the target user. Then, based on the medication rules, the target user's corresponding medical prescription data, and the target user's physical examination results, it predicts the individual medication recipients and dosage adjustment methods for the target user using the policy domain subgraph in the medical insurance knowledge graph. Finally, based on the medication recipients and dosage adjustment methods, the system uses the AI ​​medical system to generate medical insurance guidance data for the target user using the resource domain subgraph in the medical insurance knowledge graph.

[0057] By comprehensively analyzing target users' historical medical insurance data, pathological data, seasonal patterns, medical prescription data, and physical examination results, and utilizing a medical insurance knowledge graph and AI medical system, personalized and precise medical insurance guidance and medication dosage adjustment suggestions are provided. This not only helps improve treatment outcomes but also enhances users' health management, reduces risks arising from improper medication use or inappropriate dosage, and optimizes the allocation and efficiency of medical insurance resources.

[0058] In some embodiments, after predicting the target medication and dosage adjustment method for an individual user based on medication rules and medical prescription data through a subgraph of the policy domain in a health insurance knowledge graph, the method may further include the following steps: Based on the target users and their corresponding medication needs and dosage adjustment methods, the AI ​​medical system is used to predict the medication needs of various target users by using the domain subgraphs of the resource and policy domains in the medical insurance knowledge graph, thus obtaining the medication demand prediction results. In response to drug guidance requests from target medical institutions distributed among designated medical institutions, drug quantity adjustment guidance data is generated for the target medical institutions based on the target drug demand forecast results for the target medical institutions and the historical drug output data corresponding to the current season in the specified seasonal alternation pattern, so as to carry out advance stockpiling of drugs.

[0059] By integrating and analyzing drug demand forecasts from multiple target users, resource and policy information from the health insurance knowledge graph, and considering the impact of seasonal patterns on drug use, precise guidance data for adjusting drug quantities is provided to target medical institutions. This helps medical institutions to stock up on drugs in advance, ensuring sufficient and reasonable drug supply in different seasons or specific time periods, thereby improving the efficiency and quality of medical services and reducing problems caused by drug shortages or surpluses.

[0060] In some embodiments, after predicting the individual medication recipients and dosage adjustment methods for the target user based on the medication rules, the target user's corresponding medical prescription data, and the target user's physical examination results through the policy domain subgraph in the medical insurance knowledge graph, the method may further include the following steps: Based on the target user's medication rules, the person to be taking the medication, the medication dosage adjustment method, and the physical examination results, the AI ​​medical system uses the domain subgraph of the medical field in the medical insurance knowledge graph to predict the target user's chronic disease medication use, and obtain the target user's chronic disease medication prediction results. Based on the ratio data between the predicted chronic disease medication use results and the current chronic disease medication amount for multiple target users, the domain subgraph of the resource domain in the medical insurance knowledge graph is updated to obtain the updated medical insurance knowledge graph for the newly added chronic disease situation.

[0061] By comprehensively analyzing target users' medication habits, medication recipients, dosage adjustment methods, and physical examination results, and utilizing medical domain information from the AI ​​healthcare system and the medical insurance knowledge graph, the system accurately predicts the target users' chronic disease medication usage. Based on these predictions and the ratio of current chronic disease medication usage, the subgraph of the resource domain in the medical insurance knowledge graph is further updated to reflect new chronic disease cases. This process enables the medical insurance knowledge graph to more accurately reflect real-world chronic disease trends and changes in medication demand, thereby supporting more scientific and rational allocation and management of medical resources.

[0062] In some embodiments, the above-mentioned generation of medical insurance guidance data for target users based on the target patient and the method of drug dosage adjustment through the domain subgraph of the resource domain in the medical insurance knowledge graph using an AI medical system may specifically include the following steps: Based on the target user's historical medication data, an AI-powered medical system generates the target user's personal medication habit data; the personal medication habit data includes at least one of the following: drug price, drug manufacturer, drug model / version, and drug source; Based on personal medication habit data, the individuals to be taken with medication, and medication dosage adjustment methods, the AI ​​healthcare system generates personalized medical insurance guidance data for target users by utilizing the domain subgraph of the resource domain in the medical insurance knowledge graph.

[0063] By integrating the target user's personal medication habits (such as preferences for drug price, manufacturer, model, or source) with their current treatment needs (the target of medication and dosage adjustment methods), and with the support of medical insurance knowledge graph resources, highly personalized medical insurance guidance data is generated, thereby improving the accuracy of medical insurance services, user compliance, and resource matching efficiency.

[0064] Figure 5 A schematic diagram of a structure for constructing and enhancing a retrieval device based on a medical insurance knowledge graph is provided. For example... Figure 5 As shown, the medical insurance knowledge graph-based construction and retrieval enhancement device 500 includes: Module 501 is used to acquire medical insurance data including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and the distribution of designated medical institutions. It establishes cross-domain anchor nodes in a graph database, using the target relationships between the target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target designated medical institution distributions related to the anchor nodes as edges connecting the anchor nodes. Domain labels corresponding to the target relationships are set on the edges. Time windows are added to the attributes of the edges based on the target drug chemical composition and pharmacological attributes and the target medical insurance reimbursement rules, resulting in a medical insurance knowledge graph constructed based on domain subgraphs. The anchor nodes include at least one of drugs, personnel, and medical institutions; the domain labels include at least one of policy domains, resource domains, and medical domains. The identification module 502 is used to respond to a user's query request, analyze the query request using a classification model and identify multiple user intentions of the query request, match the user intentions with the domain labels, and obtain multiple target domain labels that match the user intentions. The retrieval module 503 is used to decompose the query request into atomic query data for target domain subgraphs corresponding to different domains based on the target domain tags, inject time parameters into the atomic query data according to the current time and the time data in the query request, and obtain multiple parallel sub-retrieval tasks. Based on the sub-retrieval tasks, the target domain subgraphs and the time parameters, the module performs shortcut edge retrieval calculation and timeliness window compliance filtering in the medical insurance knowledge graph. Based on the frequency of occurrence of the retrieval calculation results, the module determines the pre-calculated shortcut edge results, and merges the multiple pre-calculated shortcut edge results in memory through logical intersection operation to generate the final query result.

[0065] The medical insurance knowledge graph-based construction and retrieval enhancement device provided in this application embodiment has the same technical features as the medical insurance knowledge graph-based construction and retrieval enhancement method provided in the above embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0066] An electronic device provided in this application embodiment, such as Figure 6 As shown, the electronic device 600 includes a processor 602 and a memory 601. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the method provided in the above embodiments.

[0067] See Figure 6 The electronic device also includes a bus 603 and a communication interface 604. The processor 602, the communication interface 604 and the memory 601 are connected through the bus 603. The processor 602 is used to execute executable modules, such as computer programs, stored in the memory 601.

[0068] The memory 601 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 604 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0069] Bus 603 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0070] The memory 601 is used to store programs. After receiving an execution instruction, the processor 602 executes the program. The method executed by the apparatus defined by the process disclosed in any of the preceding embodiments of this application can be applied to the processor 602 or implemented by the processor 602.

[0071] The processor 602 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 602 or by instructions in software form. The processor 602 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 601, and processor 602 reads the information from memory 601 and, in conjunction with its hardware, completes the steps of the above method.

[0072] Corresponding to the above-described method for constructing and enhancing retrieval based on medical insurance knowledge graph, this application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to perform the steps of the above-described method for constructing and enhancing retrieval based on medical insurance knowledge graph.

[0073] The medical insurance knowledge graph-based construction and retrieval enhancement device provided in this application embodiment can be specific hardware on a device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0074] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0075] For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0078] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the construction and retrieval enhancement method based on the medical insurance knowledge graph described in the various embodiments of this application. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0079] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0080] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for constructing and enhancing retrieval based on a medical insurance knowledge graph, characterized in that, include: This process involves acquiring medical insurance data, including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and the distribution of designated medical institutions. Cross-domain anchor nodes are established in a graph database. The target relationships between these relationships—target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target designated medical institution distributions—related to the anchor nodes are used as edges connecting the anchor nodes. Domain labels corresponding to these target relationships are set on the edges. Time windows are added to the attributes of the edges based on the target drug chemical composition and pharmacological attributes and the target medical insurance reimbursement rules, resulting in a medical insurance knowledge graph constructed based on domain subgraphs. The anchor nodes include at least one of drugs, personnel, and medical institutions; the domain labels include at least one of policy domains, resource domains, and medical domains. In response to a user's query request, the query request is analyzed using a classification model and multiple user intentions of the query request are identified. The user intentions are then matched with the domain tags to obtain multiple target domain tags that match the user intentions. The query request is decomposed into atomic query data for target domain subgraphs corresponding to different domains based on the target domain tags. Time parameters are injected into the atomic query data based on the current time and the time data in the query request, resulting in multiple parallel sub-retrieval tasks. Shortcut edge retrieval calculations and time window compliance filtering are performed in the medical insurance knowledge graph based on the sub-retrieval tasks, the target domain subgraphs, and the time parameters. Pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results. Based on the multiple pre-calculated shortcut edge results, they are fused in memory through logical intersection operations to generate the final query result.

2. The method according to claim 1, characterized in that, The time-effectiveness window is an effective interval attribute introduced on the relation edges of the graph to construct a slowly changing dimension SCD model in the medical insurance knowledge graph; the multiple target domain labels are labels corresponding to multiple target domains identified by analyzing the user's natural language corresponding to the query request before RAG retrieval using an intent router; The process involves injecting time parameters into the atomic query data based on the current time and the time data in the query request, resulting in multiple parallel sub-retrieval tasks. Based on these sub-retrieval tasks, the target domain subgraph, and the time parameters, shortcut edge retrieval calculations and timeliness window compliance filtering are performed in the medical insurance knowledge graph, including: When generating a database query statement, the current timestamp and / or the user-specified timestamp in the query request are automatically inserted as variables into the WHERE clause. The current timestamp and / or the user-specified timestamp are injected as filter parameters into the Prompt or code generated by the retrieval statement, resulting in multiple sub-retrieval tasks that are executed in parallel. Based on the sub-retrieval task, the target domain subgraph, and the time parameter, the target domain subgraph is dynamically activated in the medical insurance knowledge graph for retrieval. The retrieval calculation of shortcut edges and the compliance filtering of the timeliness window are performed by calling the query interface corresponding to the target domain subgraph.

3. The method according to claim 1, characterized in that, The results of multiple pre-computed shortcut edges are fused in memory through logical intersection operations to generate the final query result, including: The first pre-computed shortcut edge result is detected by the fusion engine. If the corresponding drug in the first pre-computed shortcut edge result is reimbursable, the hospital data corresponding to the first pre-computed shortcut edge result is combined with the policy data corresponding to the second pre-computed shortcut edge result by the fusion engine to construct a structured context. Based on the structured context, natural language responses are generated through LLM. According to the high-frequency inference path that appears more frequently than a specified frequency in the asynchronous analysis query log of the fusion engine, a new shortcut edge is established between the two nodes corresponding to the high-frequency inference path. Based on the pharmacological properties of the target drug's chemical components and the target medical insurance reimbursement rules in the high-frequency inference path, a timeliness window is added to the attributes of the new shortcut edge to generate the fixed shortcut edge in the medical insurance knowledge graph.

4. The method according to claim 1, characterized in that, The process of performing shortcut edge retrieval calculations and time-sensitivity filtering in the medical insurance knowledge graph based on the sub-retrieval task, the target domain subgraph, and the time parameter, and determining the pre-calculated shortcut edge results based on the frequency of occurrence of the retrieval calculation results, includes: Based on the sub-retrieval task, the target domain subgraph, and the time parameter, the shortcut edge retrieval calculation and the timeliness window compliance filtering are performed in the medical insurance knowledge graph using the following formula, and the pre-calculated shortcut edge results are determined based on the frequency of occurrence of the retrieval calculation results: in, This indicates the co-occurrence frequency of retrieval node A and retrieval node B in the historical query data within the sub-retrieval task; This represents the shortest path length from retrieval node A to retrieval node B in the target domain subgraph. This represents the time decay coefficient, used to control the rate at which the correlation strength decays when the time deviates from the center of the aging window; This represents the time parameter in the query request; This refers to the center time of the time window set on the shortcut edge between search node A and search node B; This represents the pre-calculated association strength value of the shortcut edge.

5. The method according to claim 1, characterized in that, Following the step of obtaining the medical insurance knowledge graph constructed based on the domain subgraph, the method further includes: In response to a medical insurance guidance request for a target user, the user's pathological data is determined based on the target user's historical medical insurance data through the domain subgraph of the medical field in the medical insurance knowledge graph. Based on the user's pathological data and the specified seasonal alternation pattern, the medication rules for the target user are generated. Based on the medication rules, the medical prescription data corresponding to the target user, and the physical examination results of the target user, the domain subgraph of the policy domain in the medical insurance knowledge graph is used to predict the individual medication to be taken and the medication dosage adjustment method for the target user. Based on the target user and the method of drug dosage adjustment, the AI ​​medical system generates medical insurance guidance data for the target user using the domain subgraph of the resource domain in the medical insurance knowledge graph.

6. The method according to claim 5, characterized in that, After predicting the individual medication recipients and dosage adjustment methods for the target user based on the medication rules and medical prescription data through the policy domain subgraph in the medical insurance knowledge graph, the method further includes: Based on the target users corresponding to the drugs to be taken and the drug dosage adjustment methods, the AI ​​medical system is used to predict the drug demand of various target users through the domain subgraphs of the resource domain and policy domain in the medical insurance knowledge graph, and the drug demand prediction results are obtained. In response to a drug guidance request for a target medical institution in the distribution of designated medical institutions, drug quantity adjustment guidance data is generated for the target medical institution based on the target drug demand forecast results corresponding to the target medical institution and the historical drug output data corresponding to the current season in the specified seasonal alternation pattern, so as to carry out advance stockpiling of drugs.

7. The method according to claim 5, characterized in that, After predicting the individual medication recipients and dosage adjustment methods for the target user based on the medication administration rules, the target user's corresponding medical prescription data, and the target user's physical examination results through the policy domain subgraph in the medical insurance knowledge graph, the following steps are also included: Based on the medication rules, the target patient, the medication dosage adjustment method, and the physical examination results corresponding to the target user, the AI ​​medical system uses the domain subgraph of the medical field in the medical insurance knowledge graph to predict the target user's chronic disease medication use, and obtains the target user's chronic disease medication prediction result. Based on the ratio data between the predicted chronic disease medication use results and the current chronic disease medication dosage for multiple target users, the domain subgraph of the resource domain in the medical insurance knowledge graph is updated to obtain the updated medical insurance knowledge graph for the newly added chronic disease situation.

8. The method according to claim 5, characterized in that, The process of generating medical insurance guidance data for the target user based on the patient to be given the medication and the medication dosage adjustment method, using the domain subgraph of the resource domain in the medical insurance knowledge graph and an AI medical system, includes: Based on the target user's historical medication data, an AI medical system generates the target user's personal medication habit data; the personal medication habit data includes at least one of the following: drug price, drug manufacturer, drug model / version, and drug source; Based on the personal medication habit data, the individuals to be taken the medication, and the medication dosage adjustment methods, the AI ​​medical system generates personalized medical insurance guidance data for the target users through the domain subgraph of the resource domain in the medical insurance knowledge graph.

9. A device for constructing and enhancing retrieval based on a medical insurance knowledge graph, characterized in that, include: A construction module is used to acquire medical insurance data including drug chemical composition and pharmacological attributes, medical insurance reimbursement rules, and the distribution of designated medical institutions. Cross-domain anchor nodes are established in a graph database. The target relationships between the target drug chemical composition and pharmacological attributes, target medical insurance reimbursement rules, and target designated medical institution distributions related to the anchor nodes are used as edges connecting the anchor nodes. Domain labels corresponding to the target relationships are set on the edges. Time windows are added to the attributes of the edges according to the target drug chemical composition and pharmacological attributes and the target medical insurance reimbursement rules, resulting in a medical insurance knowledge graph constructed based on domain subgraphs. The anchor nodes include at least one of drugs, personnel, and medical institutions; the domain labels include at least one of policy domains, resource domains, and medical domains. The identification module is used to respond to a user's query request, analyze the query request using a classification model and identify multiple user intentions of the query request, match the user intentions with the domain tags, and obtain multiple target domain tags that match the user intentions; The retrieval module is used to decompose the query request into atomic query data for target domain subgraphs corresponding to different domains based on the target domain tags. It injects time parameters into the atomic query data based on the current time and time data in the query request, resulting in multiple parallel sub-retrieval tasks. Based on the sub-retrieval tasks, the target domain subgraphs, and the time parameters, it performs shortcut edge retrieval calculations and timeliness window compliance filtering in the medical insurance knowledge graph. It determines pre-calculated shortcut edge results based on the frequency of the retrieval calculation results, and fuses these pre-calculated shortcut edge results in memory through logical intersection operations to generate the final query result.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 8.

Images