Intelligent medical insurance compliance examination method and system based on graph retrieval enhanced generation

Abstract

The invention provides an intelligent medical insurance compliance examination method and system based on graph retrieval enhanced generation, relates to the field of medical informatization and artificial intelligence, and solves the technical problem that recessive medical insurance violation behaviors with context dependence and structure combination cannot be effectively identified in the prior art. The method comprises the following steps: constructing a compliance knowledge graph; extracting clinical entities from the diagnosis and treatment records, and mapping the clinical entities to corresponding nodes in the compliance knowledge graph; based on the mapped clinical entity as a query starting point, executing multi-hop graph retrieval in the compliance knowledge graph to obtain a policy rule sub-graph related to the current diagnosis and treatment behavior, and generating a compliance review result through structured reasoning; constructing a violation risk propagation model based on the historical auditing data, and performing violation risk assessment on the diagnosis and treatment behaviors to obtain a risk early warning level; and generating an audit priority queue for the plurality of to-be-audited diagnosis and treatment records based on the violation confidence and the risk early warning level. The method and the device are used in a medical insurance compliance examination process.

Description

Technical Field

[0001] This application relates to the fields of medical informatics and artificial intelligence, and in particular to an intelligent medical insurance compliance review method and system based on graph retrieval-enhanced generation. Background Technology

[0002] With the continuous improvement of the medical security system and the sustained expansion of the medical insurance fund, the security supervision of the medical insurance fund faces increasingly severe challenges. In recent years, medical insurance audits have used big data analysis or machine learning models to identify abnormal behavior.

[0003] However, existing technologies have a significant drawback: they cannot effectively identify implicit violations with context-dependent and structurally combinable characteristics. Specifically, many medical insurance violations do not stem from obvious errors in a single item, but rather from multiple seemingly compliant items violating restrictive or prohibitive logical constraints in policies within specific treatment scenarios. Because existing methods lack structured modeling of the logical relationships between policy rules, and generative AI models are not deeply coupled with authoritative policy knowledge during reasoning, they either miss complex violations due to rule fragmentation or generate unfounded review conclusions due to "illusions," failing to meet the stringent requirements of accuracy, interpretability, and compliance in medical insurance audits.

[0004] This deficiency has become a core technical bottleneck restricting the improvement of the efficiency of intelligent medical insurance supervision, and a new paradigm that can deeply integrate structured policy knowledge and generative reasoning ability is urgently needed to solve it. Therefore, this application provides an intelligent medical insurance compliance review method and system based on graph retrieval and enhanced generation. Summary of the Invention

[0005] This application provides an intelligent medical insurance compliance review method and system based on graph retrieval enhancement, which solves the technical problem that existing technologies cannot effectively identify hidden medical insurance violations with context dependence and structural composition.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for intelligent medical insurance compliance review based on graph retrieval enhancement is provided, including: Construct a compliance knowledge graph that includes medical insurance policies, rules, and their logical constraints; Clinical entities are extracted from medical records and mapped to corresponding nodes in the compliance knowledge graph, and marked as target mapping nodes; the clinical entities include names of chargeable items, names of surgical procedures, or names of medical consumables; Based on the mapped clinical entities as the query starting point, a multi-hop graph retrieval is performed in the compliance knowledge graph to obtain a policy rule subgraph related to the current diagnosis and treatment behavior; The medical records and the policy rule subgraph are input into the generative reasoning model, and the compliance review results are generated through structured reasoning. The compliance review results include violation determination, policy basis, rectification suggestions and violation confidence level. A violation risk propagation model is constructed based on historical audit data, and the compliance knowledge graph is used to assess the violation risk of the diagnosis and treatment behavior to obtain a risk warning level. Based on the aforementioned confidence level of violation and the aforementioned risk warning level, an audit priority queue is generated for multiple pending medical records.

[0007] Based on the above technical solutions, this application provides an intelligent medical insurance compliance review method based on graph retrieval enhancement. It constructs a logically consistent medical insurance compliance knowledge graph, transforming policy provisions into structured rule nodes and constraint edges with validity levels. By accurately mapping clinical entities in medical records and using them as starting points for multi-hop, intent-aware graph retrieval, it dynamically acquires policy subgraphs related to the current behavior. These subgraphs are embedded as rule anchors in generative model prompts, driving the model to perform structured reasoning according to a five-stage chain template, forcibly referencing specific rule IDs to ensure the review results are traceable and auditable. Simultaneously, it combines a graph neural network risk propagation model trained on historical audit data to assess potential combined risks and output risk warning levels. It integrates violation confidence and warning levels to generate an audit priority queue, optimizing the allocation of manual review resources. This solution deeply integrates knowledge graphs and generative AI, significantly improving the ability to identify concealed and structured insurance fraud, achieving an efficient closed loop from intelligent judgment to business implementation, and effectively overcoming the shortcomings of existing audit technologies such as isolated rules, lack of contextual understanding, difficulty in identifying combined violations, and unexplainable AI output.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for constructing the compliance knowledge graph includes: The medical insurance policy document is segmented at multiple granularities to generate hierarchical text units covering general instructions, chapters, clauses, and sub-items; Based on the hierarchical text units, corresponding violation scenario nodes are constructed, and the logical constraint relationships between each node are identified through the medical insurance semantic pattern library. The logical constraint relationships include one or more of the following: project connotation inclusion relationship, exclusion content exclusion relationship, and mutual exclusion relationship of not being able to charge at the same time. The medical insurance semantic pattern library is constructed in the following way: collecting the violation judgment logic in historical medical insurance audit cases, summarizing it into a structured semantic template, and binding it with policy keywords to guide the automatic annotation of logical constraint edges; Based on the violation scenario nodes and their logical constraints, an initial compliance knowledge graph is constructed.

[0009] Among them, the logic consistency verification is performed on the initial compliance knowledge graph. If contradictory constraints are detected derived from different policy clauses, the conflict is resolved according to the preset medical insurance rule priority strategy, and the conflict status and retention basis are marked in the graph.

[0010] In conjunction with the first aspect above, in one possible implementation, the extraction of clinical entities and mapping them to corresponding nodes in the compliance knowledge graph includes: Based on a pre-built multi-source coding mapping index library, multiple candidate knowledge graph nodes are generated for the clinical entity. The multi-source coding mapping index library integrates medical insurance catalog codes, ICD surgical codes, UDI consumable codes, and local alias dictionaries. Obtain the contextual information of the clinical entity in the medical record, including the patient diagnosis, the department that performed the procedure, the associated billing items, and the operation time; A semantic compatibility assessment is performed based on the context information and the policy applicability conditions associated with each candidate knowledge graph node to obtain a compatibility score for each candidate node. The policy applicability conditions include one or more of the following: medical institution level restrictions, mandatory accompanying diagnostics, mutually exclusive billing items, or limited use scenarios. Based on the compatibility score, a target mapping node is selected, and structured mapping information containing the mapping result and confidence level is output.

[0011] In conjunction with the first aspect above, in one possible implementation, the method for analyzing the compatibility scores of each candidate node includes: The context information is encoded into a context feature vector; The policy applicability conditions associated with each candidate knowledge graph node are parsed into a set of structured constraints; Pre-defined weights are assigned to different types of policy application conditions; among them, prohibitive or restrictive conditions have a higher weight than advisory conditions. Based on the context feature vector and the structured constraint set, a compatibility score is calculated using a weighted semantic matching function. The weighted semantic matching function imposes a significant negative penalty when any hard constraint is not satisfied, and adds a positive contribution to the satisfied constraints according to their weights.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the policy rule subgraph related to the current medical practice includes: Using the mapped clinical entities as initial query nodes, a query intent representation is generated by combining the current diagnosis and treatment context. In each hop graph traversal, the path expansion priority is dynamically determined based on the semantic relevance of neighboring nodes to the query intent and the effectiveness level of the medical insurance rules corresponding to the connecting edges; wherein, the effectiveness level of the medical insurance rules includes prohibitive rules, restrictive rules and descriptive rules, and the edges corresponding to prohibitive rules enjoy the highest priority in path expansion; Multi-hop retrieval is performed based on the path expansion priority to obtain an initial policy rule subgraph. The initial policy rule subgraph is then semantically redundantly pruned, and node branches with a similarity to the query intent representation below a preset threshold are removed to obtain the policy rule subgraph.

[0013] In conjunction with the first aspect above, in one possible implementation, the generation of the compliance review result includes: Construct a chain-thinking reasoning template that includes five stages: fact reconstruction, rule matching, logic verification, violation determination, and rectification suggestions; The policy rule subgraph is converted into a set of structured rule statements with unique identifiers and embedded as rule anchors into the context of cue words in the generative reasoning model; The generative reasoning model fills in the content of each stage in sequence according to the chain-like reasoning template, and each reasoning step must reference the unique identifier of at least one of the rule anchor points and output the compliance review results.

[0014] In conjunction with the first aspect above, in one possible implementation, the construction of the violation risk propagation model includes: Obtain historical medical insurance audit data, which includes historical audit tags for each charge item, and the historical audit tags indicate the violation status of the item; The knowledge graph nodes mapped to the fee items with the corresponding historical audit labels as violations are taken as positive samples, and the nodes with the corresponding historical audit labels as non-violations are taken as negative samples. A graph neural network is constructed on the compliance knowledge graph. The graph neural network is configured to output a violation risk score for each node and dynamically allocate propagation weights according to the medical insurance rule type of the connecting edges during message transmission. Among them, the edges corresponding to prohibitive rules are given higher propagation weights than the edges of restrictive rules and explanatory rules. The graph neural network is trained under supervision based on the positive and negative samples, so that the violation risk score is consistent with the historical audit label, and a violation risk propagation model is obtained.

[0015] In conjunction with the first aspect above, in one possible implementation, the assessment of the violation risk of the medical treatment includes: The mapped clinical entities are used as initial activation nodes and input into the trained violation risk propagation model. Multi-hop risk signal diffusion is performed on the compliance knowledge graph to obtain the violation risk score of each relevant node. Based on the violation risk score, a risk activation subgraph is constructed, and connected component clustering is performed on the risk activation subgraph to identify high-risk project combinations; Calculate the combined risk intensity of each high-risk project combination, which is obtained by weighted fusion of the average violation risk score of the nodes within the combination and the subgraph structure density; The combined risk intensity is mapped to a preset threshold range to determine the risk warning level.

[0016] In conjunction with the first aspect above, in one possible implementation, generating an audit priority queue for multiple pending medical records includes: Obtain the violation confidence level and risk warning level for each of the multiple pending medical records; The risk warning level is mapped to a preset value, and the violation confidence level and the mapped risk warning level are weighted and fused based on dynamic weights to obtain the audit priority score of each medical record. The multiple pending medical records are sorted according to the audit priority score, and an audit priority queue is generated in combination with the preset audit resource capacity. The audit priority queue is pushed to the medical insurance audit management system to guide manual review.

[0017] Secondly, this application provides an intelligent medical insurance compliance review system based on graph retrieval enhancement, comprising: a compliance knowledge graph construction module, an entity mapping module, a policy rule subgraph retrieval module, a compliance review module, a risk assessment module, and a priority scheduling module; wherein, the compliance knowledge graph construction module is used to construct a compliance knowledge graph containing medical insurance policy rules and their logical constraints; the entity mapping module is used to extract clinical entities from medical records and map them to corresponding nodes in the compliance knowledge graph, marking them as target mapping nodes; the policy rule subgraph retrieval module is used to retrieve the mapped clinical entities as... The query starts by performing a multi-hop graph retrieval within the compliance knowledge graph to obtain a policy and rule subgraph related to the current medical practice. The compliance review module inputs the medical records and the policy and rule subgraph into a generative reasoning model, generating compliance review results through structured reasoning. The risk assessment module constructs a violation risk propagation model based on historical audit data and uses the compliance knowledge graph to assess the violation risk of the medical practice, obtaining a risk warning level. The priority scheduling module generates an audit priority queue for multiple pending medical records based on the violation confidence level and the risk warning level.

[0018] This application provides an intelligent medical insurance compliance review method and system based on graph retrieval-enhanced generation, which can effectively solve the core problems in existing technologies such as rule fragmentation, lack of context, difficulty in identifying combined violations, and uninterpretable AI review results. The method constructs a structured and logically consistent medical insurance compliance knowledge graph, transforming scattered policy provisions into nodes and constraint edges with validity levels. During the review process, clinical entities in the treatment records are first accurately mapped to graph nodes, and multi-hop, intent-guided graph retrieval is performed from this starting point to dynamically obtain policy rule subgraphs highly relevant to the current treatment behavior. This subgraph is then embedded as a rule anchor point into the context of the generative reasoning model, forcing the model to perform structured reasoning according to a five-stage chain template of "fact restoration—rule matching—logic verification—violation judgment—rectification suggestion," ensuring that each conclusion cites specific policy basis, significantly improving interpretability and compliance credibility. Simultaneously, a violation risk propagation model trained based on historical audit data is used to assess potential combined risks of treatment behavior and output risk warning levels. The violation confidence level and warning level are integrated to generate an audit priority queue, guiding manual review resources towards high-risk cases. This application achieves a deep integration of knowledge-driven and data-driven approaches, which not only significantly improves the ability to detect hidden and coordinated fraudulent activities, but also constructs a complete closed loop from intelligent discovery to business implementation, providing efficient, accurate, and auditable intelligent regulatory support for the security of medical insurance funds.

[0019] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0020] Figure 1 A system architecture diagram of an intelligent medical insurance compliance review system based on graph retrieval enhancement provided in this application embodiment; Figure 2 A flowchart illustrating an intelligent medical insurance compliance review method based on graph retrieval enhancement provided in this application embodiment; Figure 3This is a flowchart illustrating a method for generating medical insurance compliance review results, provided as an embodiment of this application. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The intelligent medical insurance compliance review method based on graph retrieval enhancement provided in this application embodiment can be applied to, for example... Figure 1 The intelligent medical insurance compliance review system based on graph retrieval enhancement is shown. The system includes: a compliance knowledge graph construction module, an entity mapping module, a policy rule subgraph retrieval module, a compliance review module, a risk assessment module, and a priority scheduling module. The compliance knowledge graph construction module is used to build a compliance knowledge graph that includes medical insurance policies, rules, and their logical constraints. The entity mapping module is used to extract clinical entities from medical records and map them to the corresponding nodes in the compliance knowledge graph, marking them as target mapping nodes; The policy rule subgraph retrieval module is used to perform multi-hop graph retrieval in the compliance knowledge graph based on the mapped clinical entities as the query starting point, and obtain the policy rule subgraph related to the current diagnosis and treatment behavior. The compliance review module is used to input medical records and policy rule subgraphs into a generative reasoning model, and generate compliance review results through structured reasoning. The risk assessment module is used to build a violation risk propagation model based on historical audit data and to use a compliance knowledge graph to assess the violation risk of medical treatment behavior and obtain a risk warning level. The priority scheduling module is used to generate an audit priority queue for multiple pending medical records based on the violation confidence level and risk warning level.

[0023] To address the technical problem that existing technologies cannot effectively identify hidden medical insurance violations with contextual dependencies and structural composition, embodiments of this application provide an intelligent medical insurance compliance review method based on graph retrieval-enhanced generation, such as... Figure 2 As shown, it includes: S201. Construct a compliance knowledge graph that includes medical insurance policies, rules, and their logical constraints.

[0024] Among them, medical insurance policies and rules are the charging standards, applicable conditions, limited payment scope and exceptions in the medical service item catalog, drug catalog and medical consumable catalog issued by various medical security departments. Natural language processing technology is used to extract and structure them into nodes in the knowledge graph. It should be noted that logical constraint relationships refer to the business semantic associations between different policies and rules, including mutual exclusion relationships (such as "Project A shall not be charged at the same time as Project B"), inclusion relationships (such as "Project C already includes consumable D"), dependency relationships (such as "Using drug E requires accompanying diagnosis F"), and hierarchical restriction relationships (such as "only applicable to tertiary hospitals"). These relationships are modeled as directed edges with type labels and together with nodes, they form a compliance knowledge graph that supports multi-hop reasoning.

[0025] S202. Extract clinical entities from medical records and map them to the corresponding nodes in the compliance knowledge graph.

[0026] Clinical entity extraction is based on a multi-granularity medical text parsing model, which identifies key elements with medical insurance audit significance from electronic medical records, expense lists or medical order records, including the name of the chargeable item, surgical operation code (such as ICD-9-CM3), generic name of drug, specifications and model of medical consumables and diagnosis conclusion (such as ICD-10 code). It should be noted that the mapping process is not a simple string matching, but a multi-dimensional matching mechanism that integrates semantic similarity, encoding system alignment rules and context compatibility verification. This mechanism accurately links the extracted clinical entities to policy nodes in the compliance knowledge graph that are semantically equivalent or business-equivalent. The compatibility verification dynamically calculates the mapping confidence based on the policy application conditions (such as the level of medical institutions, patient age, and companion diagnostics) to ensure that subsequent inference is based on the correct rule context.

[0027] S203. Based on the mapped clinical entities as the query starting point, perform multi-hop graph retrieval in the compliance knowledge graph to obtain the policy rule subgraph related to the current diagnosis and treatment behavior.

[0028] Among them, multi-hop graph retrieval uses the mapped clinical entity nodes as seeds and performs depth-first or breadth-first expansion traversal along the logical constraint edges (such as mutual exclusion, inclusion, limitation, exclusion, etc.) in the compliant knowledge graph. During each hop expansion, low-relevance paths are dynamically pruned in combination with the current diagnosis and treatment context, and the maximum number of hops is limited to control the size of the subgraph. It should be noted that the policy rule subgraph is not a simple neighborhood aggregation, but rather a dual-channel guidance mechanism that integrates the effectiveness level and semantic relevance of medical insurance rules. It prioritizes the retention of rule paths with high regulatory intensity (such as prohibited and restricted categories) and high relevance to medical treatment behavior. The resulting subgraph covers both explicit conflicting rules and implicit logical chains on which potential combined violations depend, providing a precise and interpretable structured context for subsequent generative reasoning.

[0029] S204. Input medical records and policy rule subgraphs into the generative reasoning model, and generate compliance review results through structured reasoning; Among them, the medical records are in the form of a structured sequence of clinical events (including billing items, diagnostic codes, surgical procedures and consumable usage), which are context-aligned with the semantic representation of the policy rule subgraph after graph embedding encoding, and together construct the prompt for the generative reasoning model to ensure that the model performs factual reasoning within the policy boundaries; It should be noted that structured reasoning is not free text generation, but rather a forced model to output in five stages through a predefined chain-like thinking template: "fact restoration → rule matching → logic verification → violation determination → rectification suggestions". Each step of the reasoning requires explicit reference to nodes or edges in the policy rule subgraph as a basis. The generated compliance review result is machine-readable structured data containing a violation determination Boolean value, policy basis identifier, confidence score, and executable rectification suggestions.

[0030] S205. Construct a violation risk propagation model based on historical audit data, and use a compliance knowledge graph to assess the violation risk of medical treatment behavior, outputting risk warning information including risk warning level.

[0031] Among them, the violation risk propagation model is built on a graph neural network architecture. It uses the violation and non-violation samples marked in historical audit data as supervision signals. It learns the propagation rules of risk signals along different types of rule edges (such as prohibited, restricted, and explanatory) on the compliance knowledge graph, and assigns learnable propagation weights to each type of edge, so that rules with high regulatory intensity play a leading role in risk diffusion. It should be noted that the risk assessment of violations does not rely solely on single-point risk scores. Instead, it uses the mapped clinical entities as initial activation nodes, runs a trained risk propagation model on the compliance knowledge graph to perform multi-hop diffusion, and then identifies highly cohesive risk subgraphs composed of multiple low-risk items. Based on the average node risk and structural density of the subgraph, the combined risk intensity is calculated, and this intensity is mapped to a preset threshold range to determine the risk warning level, thereby achieving a proactive warning of hidden and combined violations.

[0032] S206. Based on the confidence level of violation and the risk warning level, generate an audit priority queue for multiple pending medical records.

[0033] The generation of the audit priority queue is based on a two-dimensional dynamic weighted fusion mechanism, which maps the violation confidence (reflecting the credibility of certain violations) and risk warning level (reflecting the severity of potential combined violations) of each pending medical record into numerical indicators, and adaptively adjusts the weights of the two according to the current backlog of audit tasks to calculate a unified audit priority score. It should be noted that the queue generation is not simply sorted in descending order of scores, but rather a resource-aware scheduling strategy is introduced: while ensuring that high-confidence cases are processed first, a certain proportion of high-risk, low-confidence records are forcibly retained to support the discovery of new violation patterns. The queue is truncated and labeled in combination with the preset daily processing capacity limit. The output audit priority queue includes priority score, case type identifier and estimated review time, which can be directly connected to the task assignment module of the medical insurance audit management system.

[0034] Based on the above technical solutions, this application provides an intelligent medical insurance compliance review method based on graph retrieval enhancement. This method constructs a logically consistent compliance knowledge graph, transforming fragmented policies into a structured reasoning network. Through precise mapping of clinical entities and multi-hop subgraph retrieval, it ensures that the review focuses on the real policy context. It introduces chain-like reasoning with mandatory reference rule anchors, making the generated results auditable. Simultaneously, it integrates generative confidence and graph neural network-driven risk warning levels, achieving a leap from "whether there is a violation" to "how many urgent cases need to be investigated." This not only significantly improves the ability to identify hidden and structured fraud, but also directs limited manpower to high-risk cases through an audit priority queue, truly achieving a "precise, efficient, explainable, and implementable" intelligent medical insurance supervision closed loop. This effectively overcomes the bottlenecks commonly found in existing medical insurance audit systems, such as "fragmented rules, lack of context, difficulty in identifying combined violations, unexplainable AI output, and low efficiency of manual review."

[0035] In one possible implementation of this application embodiment, the above-mentioned S201 can be specifically implemented by the following S301, S302 and S303, which are described in detail below: S301. Perform multi-granular segmentation on medical insurance policy documents to generate hierarchical text units covering general instructions, chapters, clauses, and sub-items.

[0036] Among them, multi-granularity segmentation refers to parsing the original PDF or Word format policy document into a sequence of text units with semantic hierarchy based on the inherent structure of the medical insurance policy document (such as table of contents, chapter titles, clause numbers, project codes, etc.) by combining a rule engine with a deep learning model. This includes general instructions, chapters, sections, articles, clauses, items and sub-items.

[0037] In some implementations, regular expressions are used to identify the standard encoding format in the medical insurance catalog (such as "CL001" and "HC-2024-XXX"). A BERT-based paragraph classifier is then used to determine the semantic role of each paragraph (such as "scope of application", "exclusions", and "connotation"). The output is a structured JSON object with hierarchical tags. Each object contains the original text, location path, encoding ID, and semantic type.

[0038] It should be noted that this segmentation process not only preserves the original text order, but also explicitly marks the parent-child / sibling relationships between each unit, providing a topological foundation for subsequent knowledge graph construction and avoiding common problems such as "incorrect attribution of excluded content" caused by flattening.

[0039] For example, after segmenting "Item CL001: Percutaneous Coronary Intervention" in the "Medical Service Price Item Specification of xxxx", we get: Parent node: "CL001 Project Name and Code"; Child node 1: "[Content] Includes guidewire, balloon dilation..."; Child node 2: "[Excluded Content] Drug-eluting stents require separate charging"; Child node 3: "[Restricted Payment] Only applicable to cardiology departments of tertiary hospitals".

[0040] S302. Based on hierarchical text units, construct corresponding violation scenario nodes, and identify the logical constraint relationships between nodes through the medical insurance semantic pattern library; Among them, logical constraints include one or more of the following: project connotation inclusion relationship, exclusion relationship of excluded content, and mutual exclusion relationship of not being able to charge at the same time; The medical insurance semantic pattern library is constructed by collecting violation judgment logic from historical medical insurance audit cases, summarizing it into structured semantic templates, and binding them with policy keywords to guide the automatic annotation of logical constraint edges.

[0041] Among them, violation scenario nodes refer to the smallest business units with independent audit significance extracted from each hierarchical text unit, such as a charging item, an excluded consumable, or a usage restriction condition. These nodes serve as the basic entities of the compliance knowledge graph.

[0042] In some implementations, the system calls a pre-built medical insurance semantic pattern library (containing 200+ regulatory keyword templates, such as "must not...at the same time", "only...for use", "already includes...no additional charge") to perform syntactic dependency analysis and pattern matching on each text unit, automatically identifying and labeling the logical constraint relationships between nodes, including mutual exclusion, inclusion, limitation and exclusion relationships.

[0043] It should be noted that logical constraint relationships are not simple co-occurrence associations, but rather directed edges with directionality and business semantics. For example, "Project A → Mutually Exclusive → Project B" means that A and B cannot be charged at the same time, while "Project C → Includes → Consumable D" means that D cannot be charged separately. These relationships will be directly used for subsequent graph retrieval and risk propagation. For example, from the sub-node "[Excluded Content] Drug-eluting stents require separate charging" of "CL001", the system identifies the following violations: Violation scenario node 1: "CL001 (PCI surgery)"; Violation scenario node 2: "Drug-eluting stent (code STENT-DES)"; Logical constraint relationship: "CL001→Excluded→STENT-DES", that is, stents can be charged separately, which does not violate the "inclusion" rule.

[0044] S303. Based on the violation scenario nodes and their logical constraint relationships, an initial compliance knowledge graph is constructed.

[0045] The initial compliance knowledge graph uses violation scenario nodes (such as chargeable items, consumables, and diagnostic codes) extracted from medical insurance policies as entity nodes in the graph, and establishes directed edges through logical constraints between them (such as mutual exclusion, inclusion, exclusion, and limitation), thereby forming a structured and semantic knowledge network to support subsequent graph retrieval and intelligent reasoning.

[0046] In some implementations, the system uses graph databases (such as Neo4j) or graph neural network frameworks (such as DGL) to construct the graph. Each node contains a unique identifier (such as a medical insurance project code), a text description, the source policy clause, and a semantic type. Each edge is labeled with a relation type (such as EXCLUDES, INCLUDES, MUTUAL_EXCLUSION) and a weight. The weight is determined by the rule's effectiveness level, ensuring that rules with high regulatory intensity take precedence in inference.

[0047] It should be noted that the knowledge base is not statically stored, but dynamically expandable. When new policy clauses are added, the system automatically identifies the new nodes and infers their relationship with existing nodes, while triggering a consistency verification process to avoid logical breaks in the knowledge base due to information updates, thus ensuring the continuous effectiveness of the knowledge base.

[0048] For example, for the violation scenario node "CL001: Percutaneous coronary intervention", the system establishes an "EXCLUDES" edge with "Drug-eluting stent (STENT-DES)" to indicate that the stent can be charged separately; at the same time, it establishes an "INCLUDES" edge with "Balloon dilation" to indicate that it is already included in the project content; if it is subsequently found that "STENT-DES cannot be charged together with bioresorbable stents", then a "MUTUAL_EXCLUSION" edge is added to form a complete combined violation detection path.

[0049] It should be noted that when performing logical consistency verification on the initial compliance knowledge graph, if contradictory constraints derived from different policy provisions are detected, the conflict will be resolved according to the preset medical insurance rule priority strategy, and the conflict status and retention basis will be marked in the graph.

[0050] Among them, logical consistency verification refers to the system automatically detecting whether there are contradictory logical constraints derived from different policy provisions after the initial compliance knowledge graph is constructed. For example, there may be conflicting edges between the same pair of nodes, such as "mutual exclusion" and "allowing simultaneous acceptance", thereby ensuring semantic consistency and reliable reasoning within the graph.

[0051] In some implementations, the verification process uses graph traversal algorithms (such as DFS or BFS) to perform path analysis on each pair of related nodes in the graph, and calculates whether the semantic relationships derived from all possible paths conflict. If a conflict is found, the preset medical insurance rule priority strategy engine is called to arbitrate according to the rule effectiveness level (such as prohibited class > restricted class > explanatory class) or the order of publication time, retaining the edges corresponding to the high-priority rules and marking the low-priority edges as "covered".

[0052] It should be noted that this verification mechanism does not simply delete conflict edges, but achieves traceability through conflict annotation with metadata—that is, it retains all original rule paths in the graph and records the conflict type, arbitration basis and source policy provisions, supporting subsequent manual review and auditing, and avoiding information loss due to rule rulings.

[0053] For example, if an old policy in a certain region stipulates that "CT scans (CT001) and MRI (MR001) cannot be charged on the same day" (mutually exclusive), while a new policy adds that "tumor patients can be charged on the same day" (a restricted exception), after the system detects that the two conflict, it retains the new rule and marks the old rule as "overridden" according to the priority strategy of "newly issued policies take precedence over old policies". At the same time, it marks "Applicable conditions: diagnostic code C50.9" in the atlas to achieve refined compliance expression.

[0054] Based on the above technical solutions, in the scenario of intelligent medical insurance supervision, policy rules come from diverse sources, are frequently updated, and have complex semantics. If the original text is directly used for compliance review, it is easy to cause reasoning errors due to rule conflicts (such as contradictions between old and new clauses, inconsistencies in local policies) or logical deficiencies (such as combinations of behaviors that are not explicitly prohibited but are actually violated), resulting in "missed reviews" or "misjudgments," which seriously affect the security of the fund and the trust of medical institutions. Therefore, it is urgent to build a compliance knowledge foundation that is logically consistent, structurally clear, and can support deep reasoning. This application systematically solves the above problems through a three-step process of "multi-granular segmentation → construction of scenario nodes and logical relationships → logical consistency verification." Multi-granular segmentation preserves the hierarchical semantics of the original policy text; based on the medical insurance semantic pattern library, it accurately extracts business constraint relationships such as mutual exclusion, inclusion, and limitation to form a machine-readable graph structure; most importantly, it introduces a conflict arbitration mechanism based on the rule validity level to automatically resolve contradictions and ensure the internal logical consistency of the graph. This solution not only avoids the high cost and subjective bias of manual mapping, but also fundamentally ensures the reliability of subsequent graph retrieval, risk propagation and generative reasoning, providing a solid foundation for achieving intelligent medical insurance auditing with high accuracy, interpretability and auditability.

[0055] In one possible implementation of this application embodiment, the above-mentioned S202 can be specifically implemented by the following S401, S402, S403 and S404, which are described in detail below: S401. Based on a pre-built multi-source coding mapping index library, multiple candidate knowledge graph nodes are generated for clinical entities.

[0056] Among them, the multi-source coding mapping index library is a unified mapping system that integrates medical insurance catalog codes, ICD surgical codes, UDI consumable codes and local alias dictionaries. It is used to solve the problem of inconsistent naming of clinical entities under different standard systems, thereby achieving cross-system semantic alignment and accurate matching.

[0057] In some implementations, the system constructs a global coding mapping table to associate the identifiers of the same clinical entity under different coding systems. For example, it establishes a bidirectional mapping relationship between the medical insurance project code "CL001" and the ICD-9-CM3 surgical code "37.21", the UDI consumable code "0123456789", and the hospital's internal name "PCI surgery", supporting reverse lookup of all possible corresponding nodes from any coding form.

[0058] It should be noted that this mapping process is not a simple string matching, but a multi-dimensional verification that combines semantic similarity, encoding hierarchy structure and context compatibility to ensure that the generated candidate nodes not only have consistent encoding, but also match the business meaning, thus avoiding incorrect mapping due to encoding ambiguity.

[0059] For example, when "coronary angiography + stent placement" appears in the medical record, the medical insurance code "CL001" corresponding to "coronary angiography" is identified, and its ICD code "37.21" and UDI consumable association "STENT-DES" are found through the mapping index database. At the same time, it is found that local hospitals often refer to it as "coronary intervention". Therefore, three candidate nodes are generated: "CL001", "37.21" and "STENT-DES" for subsequent map retrieval.

[0060] S402. Obtain the contextual information of the clinical entity in the medical record. The contextual information includes the patient's diagnosis, the department that performed the procedure, the associated billing items, and the operation time.

[0061] Contextual information refers to business environment data surrounding a clinical entity in actual diagnosis and treatment scenarios, used to determine whether the entity meets the applicable conditions stipulated by medical insurance policies, such as "only applicable to tertiary hospitals" or "requires a specific diagnosis", thereby improving the accuracy of mapping and reasoning.

[0062] In some implementations, the system automatically extracts key fields by parsing electronic medical records, expense lists, and medical orders: the patient's ICD-10 diagnostic code (e.g., C50.9), the department where the surgery was performed (e.g., cardiology), the associated chargeable items (e.g., CL001, STENT-DES), and the operation time. These are then structured into a contextual metadata vector for subsequent rule matching and graph retrieval.

[0063] It should be noted that this contextual information is not only a static attribute, but also a dynamic basis for verification. For example, if a project stipulates that it is "only for use by tertiary hospitals", but the actual implementing department is a secondary hospital, then even if the code matches, it should be marked as high risk to avoid misjudging compliance.

[0064] For example, when the system identifies the clinical entity "drug-eluting stent (STENT-DES)," it simultaneously obtains its contextual information: Patient diagnosis: "C50.9 adenocarcinoma"; Department: "Tertiary hospital cardiology department"; Related project: "CL001 percutaneous coronary intervention"; Operation time: "xxxx-xx-xx". Based on this, the system determines that the stent use meets the exception clause of "cancer patients can be charged separately" and is performed in a compliant institution, thus confirming it as a legitimate charging behavior.

[0065] S403. Based on the context information and the policy applicability conditions associated with each candidate knowledge graph node, perform a semantic compatibility assessment to obtain the compatibility score of each candidate node. The policy's applicable conditions include one or more of the following: restrictions on the level of medical institutions, mandatory accompanying diagnostics, mutually exclusive billing items, or limited usage scenarios. Specifically, the context information is encoded into a context feature vector; The policy applicability conditions associated with each candidate knowledge graph node are parsed into a set of structured constraints; Pre-defined weights are assigned to different types of policy application conditions; among them, prohibitive or restrictive conditions have a higher weight than advisory conditions. Based on the context feature vector and the set of structured constraints, a compatibility score is calculated using a weighted semantic matching function. The weighted semantic matching function imposes a significant negative penalty when any hard constraint is not satisfied, and adds a positive contribution to the satisfied constraints according to their weights.

[0066] Semantic compatibility assessment refers to matching and analyzing the contextual information of clinical entities in medical records (such as patient diagnosis, performing department, and operation time) with the medical insurance policy application conditions associated with each candidate node in the knowledge graph (such as "limited to tertiary hospitals" or "requires accompanying tumor diagnosis") to determine whether the entity meets the policy-stipulated usage prerequisites, thereby providing a quantitative basis for subsequent accurate mapping.

[0067] In some implementations, contextual information is encoded into structured feature vectors (e.g., {diagnosis:"C50.9", department:"Third-level Hospital Cardiology", time:"xxxx-xx-xx"}), and the policy applicability conditions of each candidate node are parsed into a set of structured constraints (e.g., {required_diagnosis:["C50.9"], allowed_departments:["Third-level Hospital"], forbidden_with:["CT001"]}). Different types of constraints are weighted based on preset weights, such as prohibitive or restrictive conditions (e.g., "cannot be admitted simultaneously" or "limited to tertiary level") having a weight of 0.8 or higher, and advisory conditions (e.g., "generally recommended") having a weight of 0.3 or lower. A compatibility score is calculated using a weighted semantic matching function, and a significant negative penalty is imposed when any hard constraint is not met, while the satisfied constraints are added positively according to their weights.

[0068] It should be noted that this evaluation mechanism is not a simple Boolean judgment, but introduces a soft scoring mechanism to support fuzzy matching and multi-condition comprehensive decision-making, avoiding the direct exclusion of reasonable cases due to the failure of a single rule, thereby improving the robustness of the system and the rationality of medical treatment.

[0069] For example, when the system identifies “drug-eluting stent (STENT-DES)” and generates a candidate node, its associated policy conditions are: “for use only in tertiary hospitals” (restricted category, weight 0.8), “must be accompanied by tumor diagnosis” (necessary condition, weight 0.9), and “must not be received together with ordinary stents” (prohibited category, weight 0.95).

[0070] Current context information is: Diagnosis: C50.9 (tumor); Department: Cardiology Department of a secondary hospital (violating the "tertiary hospital" requirement); Related items: No conventional stent; The system calculates the score as follows: meeting the "tumor diagnosis" requirement → +0.9; not meeting the "tertiary hospital" requirement → -1.0 (hard penalty); meeting the "cannot admit patients simultaneously" requirement → +0.95; compatibility score = 0.9 + 0.95 - 1.0 = 0.85 → although high, it is marked as "high risk" and prompts for manual review.

[0071] S404. Select the target mapping node based on the compatibility score and output structured mapping information containing the mapping results and confidence levels.

[0072] Among them, the target mapping node refers to the optimal matching node among multiple candidate knowledge graph nodes, which has the highest semantic compatibility score with the clinical entity context information and meets the core policy constraints, and is used for subsequent compliance review and risk reasoning.

[0073] In some implementations, the system selects the mapping target from the candidate node list by setting a threshold strategy or sorting mechanism. For example, if the compatibility score is higher than 0.8, the highest-scoring node is directly selected as the target; if the score is lower than 0.5, it is marked as "undetermined" and a manual review process is triggered. At the same time, the mapping result and the calculated compatibility score are encapsulated together as structured data, and the output format includes: {original_entity:"drug-eluting stent",mapped_node:"STENT-DES",compatibility_score:0.85,status:"high_risk"}.

[0074] It should be noted that the output not only includes the mapping result, but also includes the confidence level (i.e., compatibility score) and status label (such as normal, high risk, pending confirmation), which enables downstream systems to decide whether to process automatically or intervene manually based on the confidence level, avoiding erroneous audit judgments due to low-quality mapping.

[0075] For example, when the system maps "stents used in coronary intervention", it generates two candidate nodes: "STENT-DES" (compliance score 0.85) and "STENT-STD" (ordinary stent, score 0.32). Since "STENT-DES" meets key conditions such as tumor diagnosis and tertiary hospital, and its score is significantly higher than other options, the system selects it as the target mapping node and outputs structured information.

[0076] Based on the above technical solutions, in intelligent medical insurance compliance review, clinical entities (such as surgeries and consumables) often suffer from mismatches or omissions when directly matching knowledge graph nodes due to inconsistent coding systems, diverse aliases, and strong contextual dependencies. For example, "drug-eluting stents" may appear in hospital systems as trade names, medical insurance codes, or ICD terms. If only string matching is relied upon, "STENT-DES" is easily mismapped as a regular stent, leading to incorrect compliance judgments. Therefore, a high-precision, context-aware intelligent mapping mechanism is urgently needed. This application constructs a multi-source coding mapping index library to unify heterogeneous naming; extracts the diagnosis and treatment context (such as diagnosis, department, and time); and performs semantic compatibility scoring based on policy applicability conditions, selecting high-confidence target nodes and outputting structured results. This method not only overcomes the problem of coding fragmentation but also integrates the conditional constraints of medical insurance rules into the mapping process, making the mapping results both "correctly coded" and "semantically compliant." Compared with traditional static matching, this solution significantly improves mapping accuracy and audit interpretability, laying a reliable foundation for subsequent graph retrieval and risk reasoning.

[0077] In one possible implementation of this application embodiment, the above-mentioned S203 can be specifically implemented by the following S501, S502 and S503, which are described in detail below: S501. Using the mapped clinical entity as the initial query node, generate a query intent representation in combination with the current diagnosis and treatment context.

[0078] Among them, query intent representation refers to the fusion encoding of the mapped clinical entity (such as the medical insurance project node) with its actual diagnosis and treatment scenario (such as patient diagnosis, performing department, and related operations) to form a semantic vector that can reflect whether the current behavior conforms to the policy boundary, which serves as a guiding signal for subsequent multi-hop graph retrieval.

[0079] In some implementations, the mapping result (such as node ID "CL001") is concatenated with contextual information (such as diagnosis code "C50.9", department "cardiology", and time "2024-03-15") to form a structured prompt, which is then input into a pre-trained medical semantic encoder (such as BioBERT or LLM fine-tuned in the medical insurance domain). The output is a dense vector with a fixed dimension, which contains both entity semantics and embedded contextual constraints of policy applicability conditions.

[0080] It should be noted that this query intent is not only used to locate graph nodes, but also serves as the initial state of a graph neural network or multi-hop retrieval system, dynamically guiding risk signals to propagate along highly relevant paths. For example, when the context includes "tumor diagnosis", the system will prioritize activating subgraph branches related to "exceptions" rather than the default prohibited path.

[0081] For example, when the mapping result is "STENT-DES" and the context shows that the patient's diagnosis is "C50.9" and the implementing agency is "tertiary hospital", the query intent representation generated by the system will strengthen the semantic direction of "can be charged separately"; while in another case, if the context is "secondary hospital", the intent representation will be biased towards the "violation of limiting conditions" path, thus prioritizing the triggering of the "institution level mismatch" related rule subgraph in subsequent graph retrieval.

[0082] S502. In each hop graph traversal, the path expansion priority is dynamically determined based on the semantic relevance of neighboring nodes to the query intent and the effectiveness level of the medical insurance rules corresponding to the connecting edges.

[0083] Among them, the effectiveness levels of medical insurance rules include prohibitive rules, restrictive rules, and descriptive rules, with the edges corresponding to prohibitive rules enjoying the highest priority in path expansion.

[0084] Among them, the path expansion priority is used to determine which neighboring nodes should be explored first in multi-hop graph retrieval. It takes into account both semantic relevance (whether the neighboring node matches the current diagnosis and treatment intention) and rule effectiveness level (the strength of the policy constraints represented by the edge).

[0085] In some implementations, the system checks each candidate neighbor node. Calculate its representation with the query intent. Cosine similarity as a semantic relevance score ;in, For nodes The embedding vector; simultaneously, assigning rule effectiveness weights based on the type of connecting edges. ,like The priority score is the weighted sum of the two.

[0086] It should be noted that this mechanism ensures that rules with high regulatory intensity (such as "cannot be accepted at the same time") can obtain a high expansion priority even if their semantic relevance is slightly low, thereby avoiding the omission of key violation paths.

[0087] For example, when traversing starting from "CL001", the neighbor node "STENT-DES" passes through the edge "EXCLUDES" (restricted class, The connection is made through the "INCLUDES" edge (description class). (Connection). If the current context is a tumor patient, then the semantic relevance of "STENT-DES" is... The combined effect of these two factors results in a significantly higher priority for expansion than "balloon dilation," leading the system to prioritize in-depth retrieval along this path.

[0088] S503. Perform multi-hop retrieval based on path expansion priority to obtain the initial policy rule subgraph, and perform semantic redundancy pruning on the initial policy rule subgraph, removing node branches with similarity to the query intent representation below a preset threshold to obtain the policy rule subgraph.

[0089] The multi-hop retrieval starts with the mapped clinical entity and dynamically selects high-value neighbor nodes for hop-by-hop traversal based on the expansion priority of each path to generate an initial policy rule subgraph covering potential violation logic. The subsequent semantic redundancy pruning calculates the similarity between each node in the subgraph and the query intent representation, and removes irrelevant or weakly relevant branches to ensure that the subgraph focuses on the true compliance boundary of the current diagnosis and treatment behavior.

[0090] In some implementations, a priority-driven traversal with at most L hops (e.g., L=2) is performed, forming an initial subgraph from all visited nodes and edges. ;right Each node Calculate its embedding vector Representation of query intent cosine similarity ,like (like If ), then remove from the subgraph The entire subtree rooted at the root is used to compress noisy paths and improve inference efficiency.

[0091] It should be noted that this pruning mechanism is not a simple truncation based on distance, but an intelligent filtering based on semantic relevance. Even if a node is reachable within 2 hops, it will still be removed if its semantics are irrelevant to the current context (such as tumor diagnosis, tertiary hospital) (e.g., "children's vaccination restrictions"), thus avoiding irrelevant rules from interfering with generative reasoning.

[0092] For example, when searching for the combination "CL001+STENT-DES", the initial subgraph may contain: relevant branches: CL001→EXCLUDES→STENT-DES (similarity 0.82); irrelevant branches: CL001→INCLUDES→pediatric sedatives (because CL001 is also defined in the pediatric catalog, but the current patient is an adult tumor, the similarity is only 0.31); after pruning, the system removes "pediatric sedatives" and its downstream nodes, and the final subgraph only retains rule paths that are highly related to "interventional treatment", "stent charges", "institution level" and "tumor exception", which significantly improves the accuracy and interpretability of subsequent compliance review.

[0093] Based on the above technical solutions, in intelligent medical insurance auditing, if compliance review of medical treatment behaviors relies solely on static rule matching or full graph traversal, the large scale of the policy graph and the abundance of semantic noise can easily lead to redundant search results or deviations from the actual scenario. For example, a "coronary intervention" procedure may be associated with hundreds of rules, but only a few exceptions or restrictions are truly relevant to the current patient (e.g., tumor diagnosis, tertiary hospital). Therefore, a context-aware, rule-sensitive, and highly focused graph retrieval mechanism is urgently needed. This application generates a query intent representation by integrating clinical entities and the treatment context, giving the retrieval a business context; in each hop traversal, it dynamically calculates path priority by combining semantic relevance and the effectiveness level of medical insurance rules, ensuring that high-regulatory prohibitive rules (such as "cannot be accepted simultaneously") are explored first; and it performs semantic redundancy pruning on the initial subgraph, eliminating branches irrelevant to the current intent. This method avoids the computational waste of blind traversal and prevents the omission of key violation paths. Compared to traditional k-hop neighbor retrieval, this solution significantly improves the relevance, inference efficiency, and audit accuracy of subgraphs, providing accurate and interpretable knowledge context for generative compliance review.

[0094] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S204 can be specifically implemented through the following S601, S602, S603 and S604, which are explained in detail below: S601. Construct a chain-thinking reasoning template that includes five stages: fact reconstruction, rule matching, logic verification, violation determination, and rectification suggestions.

[0095] Among them, the chain-based reasoning template is a structured prompt framework that forces the generative reasoning model to output intermediate reasoning steps in sequence according to five predefined logical stages, ensuring that the compliance review process is traceable, explainable, and business compliant, and avoiding logical jumps or illusions caused by the free generation of large models.

[0096] In some implementations, the template is embedded in the model input as a natural language instruction, for example: Please deduce the following steps: 1. [Fact Reconstruction] Based on the medical records, list the actual procedures performed; 2. [Rule Matching] Identify relevant clauses from the policy rule subgraph; 3. [Logical Verification] Determine if there are any conflicts or exceptional conditions; 4. [Violation Determination] A comprehensive assessment is made to determine whether a violation has occurred and the confidence level. 5. [Suggestions for Rectification] If a violation is found, provide a feasible corrective plan. The model must be filled in segment by segment, and each segment must reference a specific node or edge in the subgraph as a basis.

[0097] It should be noted that this template not only constrains the output format, but also transforms black-box generation into white-box decision-making through an explicit inference chain—the output of each stage can be verified by auditors, and the system can automatically extract structured fields (such as boolean "whether it is a violation" and numerical "confidence level") for subsequent priority queue generation.

[0098] For example, in the case of "CL001+STENT-DES being implemented in a secondary hospital", the model is output according to the template: 1. [Fact Reconstruction] The patient underwent percutaneous coronary intervention (CL001) and used a drug-eluting stent (STENT-DES). 2. [Rule Matching] The policy stipulates that "STENT-DES is an exception to CL001 and can be charged separately," but "it is only for use by tertiary hospitals." 3. [Logical Verification] The current executing agency is a secondary hospital, which does not meet the limiting conditions; 4. [Violation Judgment] A violation has been committed, with a confidence level of 0.92; 5. [Suggestions for Rectification] It is recommended to refund the STENT-DES fee or provide a referral certificate from a tertiary hospital.

[0099] S602. Convert the policy rule subgraph into a set of structured rule statements with unique identifiers, and embed them as rule anchors into the context of prompt words in the generative reasoning model.

[0100] The structured rule statement set refers to transforming each node-edge-node triple (such as "CL001—EXCLUDES→STENT-DES") in the final policy rule subgraph into a standardized statement that is readable in natural language and parsable by machines, and assigning it a globally unique identifier (such as rule_id:EXCL_2024_CL001_STENT) to serve as an immutable "rule anchor" in the generative reasoning process, ensuring that the model output is strictly based on real policy basis; In some implementations, the system traverses each directed edge in the subgraph and automatically generates rule statements based on predefined templates. For example: for "mutual exclusion" relationships: [rule_id:MUT_001] Project CL001 and Project CT001 cannot be charged on the same day; for "exclusion" relationships: [rule_id:EXCL_002] Drug-eluting stents (STENT-DES) are excluded from CL001 and can be charged separately; for "restriction" relationships: [rule_id:LIM_003] CL001 is only permitted to be performed in the cardiology department of tertiary hospitals. These statements are prioritized and then embedded as context prefixes with prompts, placed before the chain-like thinking template.

[0101] It should be noted that this mechanism effectively solves the problem of "illusory referencing" in large models—the model must explicitly refer to anchors such as "according to rule_id:LIM_003" during inference, rather than fabricating policy provisions; at the same time, the unique identifier supports downstream systems to automatically associate with the original policy source, achieving three-level traceability of "conclusion-basis-original text".

[0102] S603. Generative reasoning models fill in the content of each stage in sequence according to the chain-like reasoning template, and each reasoning step must reference the unique identifier of at least one rule anchor point.

[0103] This mechanism uses a forced generative reasoning model to output structured reasoning content in five stages: fact reconstruction, rule matching, logic verification, violation determination, and rectification suggestions. It requires that the conclusion of each stage must display a unique identifier of at least one rule anchor point associated with the policy rule subgraph (such as rule_id:LIM_2024_002). This constrains the free generation of the large model within the boundaries of real and verifiable medical insurance policies, ensuring that the review process has a legal basis and audit traceability.

[0104] In some implementations, the system embeds instructions in the prompt: "At each inference stage, it must begin with 'according to [rule_id:XXX]' or explicitly cite the rule ID," and combines this with an output format validator to post-process the model response—if a stage does not contain a valid rule reference, it triggers regeneration or marks it as a low-confidence result; at the same time, the unique identifier of the rule anchor is bound to metadata such as the original policy text, effectiveness level, and applicable conditions, allowing auditors to jump to the original policy text with one click.

[0105] It is worth noting that this design fundamentally addresses the core pain points of generative AI in regulatory scenarios: illusion and lack of interpretability. Through a "reference-based verification" mechanism, it not only eliminates the risk of models fabricating policy clauses but also ensures that every conclusion can be traced back to specific nodes and edges in the knowledge graph. This achieves a leap from "black-box judgment" to "white-box reasoning," significantly improving the compliance, credibility, and feasibility of the medical insurance intelligent review system.

[0106] S604. Generative inference models output compliance review results in a predefined JSON format.

[0107] Based on the aforementioned technical solutions, directly using large language models for compliance judgment in medical insurance intelligent auditing can easily lead to "illusions"—that is, fabricating policy basis, logical leaps, or outputting unverifiable conclusions, resulting in a lack of authority and auditability in the review results, making them difficult for regulatory agencies to adopt. Therefore, there is an urgent need for a technical solution that can both leverage the expressive capabilities of generative AI and strictly constrain its reasoning process. This application systematically solves this problem through a three-pronged mechanism: constructing a five-stage chain-like thinking template (fact restoration → rule matching → logic verification → violation judgment → rectification suggestions), forcing the model to reason step by step according to regulatory logic; transforming policy rule subgraphs into structured rule statements with unique identifiers, embedding prompt words as tamper-proof "rule anchors"; and forcing each reasoning step to reference at least one rule ID, ensuring that every conclusion has a genuine policy basis. This design deeply integrates the accuracy of knowledge graphs with the expressive power of generative models, not only eliminating the risk of policy fabrication but also achieving three-level traceability of "conclusion—basis—original text". Compared to traditional end-to-end discrimination models, this solution significantly improves the interpretability, compliance, and business usability of review results, providing a reliable paradigm for the application of AI in highly regulated fields.

[0108] In one possible implementation of this application embodiment, the above-mentioned S205 can be specifically described as follows: The construction of a violation risk propagation model includes the following steps: S701. Obtain historical medical insurance audit data; Among them, the historical medical insurance audit data includes historical audit labels for each charge item, and the historical audit labels indicate the violations of the item; The knowledge graph nodes mapped to the fee items with the corresponding historical audit labels as illegal are used as positive samples, and the nodes with the corresponding historical audit labels as non-illegal are used as negative samples.

[0109] Among them, positive sample seed nodes refer to the charging items that have been confirmed to have violations by manual review from historical medical insurance audit cases (such as "charging CT enhancement scan fees without indication"), and their corresponding knowledge graph nodes are marked with high-risk labels; while negative samples are the charging items that have been confirmed to be compliant by audit in the same period (such as "routine examinations that conform to clinical pathways"), which are used to construct the contrast signal in supervised learning. Together, they constitute the basic dataset for training the violation risk propagation model. It should be noted that the sample construction strategy is not simply based on whether the project was rejected, but rather it is finely labeled by combining the type of violation, policy basis and context. For example, the same project "CT001" is compliant in cancer patients but non-compliant in patients with the common cold, so it is classified as a negative sample and a positive sample respectively, thereby ensuring that the model learns the violation logic based on rules and context, rather than the superficial statistical correlation.

[0110] S702. Construct a graph neural network on the compliance knowledge graph. The graph neural network is configured to output a violation risk score for each node and dynamically allocate propagation weights according to the medical insurance rule type of the connecting edge during message transmission.

[0111] In this system, the graph neural network performs multi-round message passing on the compliance knowledge graph. Each node aggregates the risk signals passed by its neighbors through the edges and outputs a violation risk score between 0 and 1. During the aggregation process, the system dynamically adjusts the propagation weight according to the type of medical insurance rule corresponding to the connecting edge. The edges corresponding to prohibitive rules (such as "cannot be charged at the same time") are given higher weights (e.g., 0.95), the edges corresponding to restrictive rules (such as "only tertiary hospitals are allowed") are given lower weights (e.g., 0.75), and the edges corresponding to explanatory rules (such as "the connotation is already included") are given the lowest weights (e.g., 0.40), thereby ensuring that rules with high regulatory intensity play a leading role in risk propagation. It should be noted that this weight allocation mechanism is not statically preset, but is combined with the learnable parameters of the graph neural network and fine-tuned based on historical audit data during the training process. This allows the model to inherit the prior effectiveness level of medical insurance policies and adaptively calibrate the real influence of different rules in actual violation scenarios, thereby improving the accuracy of risk propagation and its relevance to business.

[0112] S703. Supervised training of the graph neural network is performed based on positive and negative samples to ensure that the violation risk score is consistent with the historical audit label, thereby training a violation risk propagation model.

[0113] The supervised training uses a binary cross-entropy loss function, taking historical audit labels (positive sample labels are 1, negative sample labels are 0) as the true values ​​and the violation risk scores output by the model as the predicted values. The parameters of the graph neural network are optimized through backpropagation, so that the risk scores of high-risk nodes approach 1 and the risk scores of compliant nodes approach 0, thereby achieving end-to-end learning of medical insurance violation patterns. It should be noted that, in order to alleviate the problem of imbalance between positive and negative samples (usually, there are far fewer violations than compliance cases), a class weighting strategy or Focal Loss mechanism is introduced during training to enhance the model's sensitivity to the minority class (positive samples). At the same time, the loss calculation not only applies to the seed node, but also propagates to its multi-hop neighborhood through the graph structure, so that nodes that are not directly labeled but have close semantic relationships can also obtain reasonable risk estimates, thereby improving the model's generalization ability and subgraph-level risk identification effect.

[0114] Based on the aforementioned technical solutions, traditional rule engines struggle to identify hidden, combined violations (such as multiple compliant items pieced together to form a non-compliant package) in intelligent medical insurance supervision, while purely data-driven black-box models lack interpretability and policy basis. Therefore, a technical solution is urgently needed that can integrate knowledge from the medical insurance field and learn risk propagation patterns from historical audit experience. This application systematically addresses this problem by constructing a graph neural network based on a compliance knowledge graph: using historically audited non-compliant / non-compliant items as seed nodes for positive and negative samples, ensuring the training objective aligns with real-world regulatory scenarios; dynamically allocating propagation weights based on medical insurance rule types during message transmission; where prohibitive rules (such as "cannot be collected simultaneously") have higher weights, allowing high-intensity regulatory constraints to play a dominant role in risk diffusion; and through supervised training, aligning node risk scores with audit labels, achieving a leap from "isolated item judgment" to "graph-structured collaborative risk assessment." This method not only preserves the interpretability of policy logic but also discovers potential violation patterns that are not explicitly labeled but have similar structures. Compared to traditional methods, this solution combines domain knowledge guidance, risk signal generalization capabilities, and regulatory compliance, providing a reliable and auditable technical path for intelligent risk control of medical insurance funds.

[0115] Conducting a risk assessment of violations in medical practices includes the following steps: S801. Input the mapped clinical entity as the initial activation node into the trained violation risk propagation model, and perform multi-hop risk signal diffusion on the compliance knowledge graph to obtain the violation risk score of each relevant node.

[0116] In this process, the mapped clinical entities (such as the medical insurance project "CL001" or the consumable "STENT-DES") are used as the initial activation source for risk propagation. The corresponding knowledge graph nodes are assigned a high initial risk value (such as 1.0) at the start of inference. This risk signal spreads in multiple hops along the logical constraint edges (such as mutual exclusion, exclusion, limitation, etc.) in the graph. The neighbor information is aggregated layer by layer through the trained graph neural network to output a quantified violation risk score for all reachable related nodes.

[0117] In some implementations, the system employs a two-hop diffusion strategy: the first hop starts from the activation node and collects directly connected rule nodes (e.g., "CL001→EXCLUDES→STENT-DES"); the second hop continues to expand to the condition nodes associated with these rule nodes (e.g., "STENT-DES—REQUIRES→Tumor Diagnosis C50.9" or "CL001—LIMITED_TO→Tertiary Hospital"); the message transmission weight of each hop is dynamically adjusted according to the edge type, with prohibitive rule edges (e.g., weight 0.95) > restrictive rule edges (e.g., weight 0.75) > descriptive rule edges (e.g., weight 0.40), ensuring that high-regulatory-strength paths dominate risk transmission.

[0118] It should be noted that the diffusion process is not a full graph computation, but a local subgraph reasoning centered on the activated node. This ensures computational efficiency while focusing on the policy context that is truly relevant to the current diagnosis and treatment behavior. More importantly, the risk score is a dynamic result of context awareness. The same node "STENT-DES" has a low risk in the cancer patient scenario but a high risk in the common cold patient scenario, which reflects the model's ability to understand the conditions for policy applicability.

[0119] For example, when a medical record contains "CL001" and "STENT-DES" and the executing institution is a "Level II Hospital", the system sets these two nodes as the initial activation source; the risk signal propagates along the path "CL001—LIMITED_TO→Level III Hospital", and a high-risk response is generated because the actual institution does not match; at the same time, although "STENT-DES" is excluded, its risk score is still raised to 0.87 because the institution level does not meet the precondition; the model outputs the risk distribution of the relevant nodes, providing a basis for subsequent high-risk combination identification.

[0120] S802. Construct a risk activation subgraph based on the violation risk score, and perform connected component clustering on the risk activation subgraph to identify high-risk project combinations.

[0121] Among them, the risk activation subgraph refers to the induced subgraph formed by selecting all nodes with violation risk scores higher than a preset threshold (such as 0.6) from the compliance knowledge graph and the connecting edges between them, which is used to focus on potential violation areas; connected component clustering is performed on the subgraph to group high-risk nodes that are directly or indirectly connected to each other through rule edges into the same group, thereby identifying a set of high-risk projects with collaborative violation characteristics.

[0122] In some implementations, the entire graph is traversed, and nodes with a risk score ≥ τ (τ = 0.6) are retained. All edges between these nodes in the original graph (including mutually exclusive, exclusion, and restricted types) are extracted to form a risk activation subgraph. A graph algorithm library (such as NetworkX or Neo4j) is called to calculate its weakly connected components, and each component corresponds to a high-risk project combination. The number of nodes in the combination, the average risk score, and the subgraph density will be used as inputs for subsequent calculations of the combination risk intensity.

[0123] It should be noted that this method can not only identify individual high-risk projects, but also capture hidden violation patterns formed by multiple medium- and high-risk projects through policy logic. For example, the risk scores of "CL001" and "STENT-DES" are both 0.7 when viewed individually, but because they have an "exclusion" relationship and appear in the secondary hospital scenario at the same time, they are identified as a high-risk combination after clustering, while traditional single-point scoring would miss such collaborative risks.

[0124] For example, as shown in the uploaded document example, in a certain medical record, the model outputs: "CT001 (CT examination)" has a risk score of 0.68, and "MR001 (MRI examination)" has a risk score of 0.72. There is a mutually exclusive edge between the two: "cannot be charged on the same day".

[0125] S803. Calculate the combined risk intensity of each high-risk project combination. The combined risk intensity is obtained by weighted fusion of the average violation risk score of the nodes in the combination and the subgraph structure density. Map the combined risk intensity to a preset threshold range to determine the risk warning level.

[0126] Among them, the portfolio risk intensity is used to quantify the overall severity of violations of a high-risk portfolio. Its calculation integrates two dimensions: first, the average violation risk score of all nodes in the portfolio, reflecting the individual risk level; second, the subgraph structure density, which is the ratio of the number of edges of the subgraph formed by the portfolio in the compliance knowledge graph to the theoretical maximum number of edges (or equivalent connectivity index), used to measure whether there are multiple mutually supporting policy logic connections between projects (such as mutual exclusion, exclusion, restriction, etc.), thereby determining whether systemic and collaborative violations are constituted. In some implementations, subgraph structure density Through formula Calculate; where N is the number of nodes within the combination, The number of logical constraint edges actually existing between these nodes; combined risk intensity. Then a weighted fusion method will be used: ; in, To calculate the average risk score for violations, Pre-defined weights (emphasizing individual risk) ensure that high-risk and highly interconnected combinations receive higher intensity scores; It should be noted that structural density is introduced to identify the "risk aggregation effect". If several medium-risk projects are closely linked through multiple prohibitive or restrictive rules (such as violating multiple conditions such as mutual exclusion, institutional restrictions, and accompanying diagnosis at the same time), the combined harm of their violations is far greater than that of isolated high-risk projects. Relying solely on average scores will underestimate such hidden but high-risk insurance fraud patterns, while the addition of structural density enables the model to perceive complex violation topologies.

[0127] For example, in a medical insurance audit, the system identified a high-risk item combination containing two nodes: "Percutaneous Coronary Intervention (CL001)" and "Drug-Eluting Stent (STENT-DES)". Their violation risk scores were 0.82 and 0.76, respectively, resulting in an average violation risk score of 0.79. Further analysis of the combination's connectivity in the compliance knowledge graph revealed two strong logical constraint edges: "STENT-DES is an exception to CL001" (restrictive rule) and "CL001 is limited to use in tertiary hospitals" (restrictive rule). The combination has 2 nodes, a theoretical maximum of 1 edge, and an actual number of edges (considering bidirectional or multi-element associations). The calculated subgraph structure density is 0.95. Using a weighted fusion formula The combined risk intensity is obtained; The preset risk warning level threshold range is: Corresponding to "high risk", Corresponding to "medium risk", [ Corresponding to "low risk"; Since 0.838 ≥ 0.8, the system judges the combination as high risk and outputs a risk warning message: "High-risk combination: CL001+STENT-DES, violates the institution level restriction (rule_id:LIM_2024_002), it is recommended to focus on reviewing the qualifications of the implementing hospital."

[0128] Based on the above technical solutions, traditional methods in intelligent medical insurance auditing often rely on single-item rule matching or isolated scoring, making it difficult to identify hidden violations composed of multiple seemingly compliant items (such as "reasonable items + reasonable consumables + unreasonable usage scenarios"), resulting in the failure to detect a large number of structural fraudulent activities. Therefore, there is an urgent need for a dynamic risk assessment mechanism that can start from the overall diagnosis and treatment behavior and explore the policy logic connections between items. This application uses mapped clinical entities as activation sources to perform multi-hop risk signal diffusion on the compliance knowledge graph, allowing risks to propagate along high-weight rules such as prohibition and restriction, achieving context-aware risk generalization; it aggregates semantically related high-risk nodes into combinations through connected component clustering, breaking through the limitations of single-point analysis; and it introduces a combination risk intensity index, integrating the average risk score and subgraph structure density to accurately quantify the severity of collaborative violations; and it generates graded early warning levels through threshold mapping to support differentiated audit handling. This method can not only discover novel combinations of violations that were not previously labeled, but also significantly improve the accuracy of review and the credibility of supervision by relying on logically consistent knowledge graphs and explainable propagation paths, thus providing intelligent and systematic risk control protection for the security of medical insurance funds.

[0129] In one possible implementation of this application embodiment, the above-mentioned S206 can be specifically implemented by the following S901 and S902, which are described in detail below: S901. Obtain the violation confidence level and risk warning level corresponding to multiple pending medical records; map the risk warning level to a preset value, and perform weighted fusion of the violation confidence level and the mapped risk warning level based on dynamic weights to obtain the audit priority score of each medical record.

[0130] Among them, the violation confidence level reflects the credibility of the generative reasoning model's judgment on whether a single medical record has a violation (range 0-1), while the risk warning level characterizes the overall severity of the combination of high-risk items identified in the record (e.g., high / medium / low). To achieve precise scheduling of audit resources, the system maps the risk warning level to preset values ​​(e.g., high=0.9, medium=0.6, low=0.3) and introduces a dynamic weighting mechanism to weight and fuse the violation confidence level with the mapped level value to generate a sortable audit priority score, which is used to guide the construction of the manual review queue.

[0131] In some implementations, dynamic weights are adjusted based on the current audit workload, historical false alarm rates, or policy priorities; for example, during a special campaign to combat "repeated inspections," the system automatically increases the weight of risk warning levels (e.g., =0.7), ensuring that high-risk portfolios receive high priority even with moderate confidence levels; the fusion formula is: ; in, For audit priority scores, This represents the mapped risk warning level value. For the degree of confidence in the violation, The dynamic weights are adaptively adjusted based on the current backlog of audit tasks, thereby achieving a balance between high-confidence violations and high-risk potential violations.

[0132] It should be noted that this fusion mechanism avoids the one-sidedness of a single indicator: relying solely on confidence level may overlook structurally high-risk behaviors (such as multiple medium-confidence items forming a strong correlation combination), while relying solely on warning level may amplify false alarms with low confidence level; through dynamic weighting, the system achieves an adaptive balance between accuracy and the breadth of risk coverage.

[0133] For example, on a certain day, the system processes three pending records: Record A: Violation confidence level 0.85, risk warning level "high" (mapping value 0.9). =0.7→ Record B: Confidence level 0.92, but no high-risk combination (grade "low", 0.3) → Record C: Confidence level 0.70, grade "Medium" (0.6) → The audit queue is sorted by P value as A>C>B to ensure that high-risk combinations are handled first and to optimize the allocation of limited manpower.

[0134] S902. Sort multiple pending medical records according to audit priority scores, generate an audit priority queue based on preset audit resource capacity, and push the audit priority queue to the medical insurance audit management system to guide manual review.

[0135] The audit priority score serves as a quantitative indicator, used to sort multiple pending medical records in descending order to form a risk sequence from high to low. Based on this, the system combines the preset audit resource capacity (such as the number of manual review cases that can be processed daily, the number of auditors, or the maximum working hours) to extract a number of records at the top of the ranking to generate an audit priority queue. This queue is then pushed to the medical insurance audit management system in the form of structured data (such as JSON or API messages) for auditors to conduct manual reviews according to priority, thus achieving precise allocation of limited regulatory resources to high-risk cases.

[0136] In some implementations, audit resource capacity can be dynamically configured; for example, it can be set to 200 cases per day during the peak period of fund settlement at the end of the quarter and 100 cases per day during normal periods; the system automatically pulls the records to be audited at midnight every day, calculates the priority score and sorts them, and takes the top M records (M=capacity value) to generate the queue for that day; at the same time, each record in the queue is accompanied by a risk warning level, violation confidence level, high-risk project combination and rule anchor reference, which makes it easier for auditors to quickly locate the focus of the problem.

[0137] It should be noted that this mechanism not only improves audit efficiency but also enables risk-oriented intelligent task assignment: avoiding delays in high-risk cases under the traditional "first-come, first-served" or "random inspection" models, and ensuring that major violations are intercepted within the golden window period; in addition, the queue generation process can be linked with the audit feedback closed loop—if the false alarm rate of a certain type of high-priority case remains high after review, the system can retrospectively adjust the confidence level or weight parameters to achieve continuous collaborative optimization between the model and the business.

[0138] Based on the aforementioned technical solutions, in the practice of intelligent auditing of medical insurance, the number of medical records awaiting review is enormous, while manual review resources are limited. If a "first-come, first-served" or random sampling method is adopted, high-risk violations are easily delayed or even missed, resulting in fund losses. Furthermore, relying solely on a single indicator (such as model confidence level) for ranking may overlook structurally high-risk combinations (such as multiple medium-confidence items constituting strong policy conflicts), making accurate resource allocation difficult. Therefore, an intelligent priority scheduling mechanism that integrates multi-dimensional risk signals and adapts to business constraints is urgently needed. This application integrates violation confidence level and risk warning level; the former reflects the reliability of model judgment, while the latter reflects the severity of combined violations. Warning levels are mapped to numerical values, and dynamic weighted fusion is introduced, enabling the system to adaptively adjust its ranking strategy under different regulatory focuses (such as emphasizing structural risks during special campaigns). A priority queue is generated based on preset audit resource capacity, ensuring that limited manpower focuses on the cases most in need of intervention. This method not only significantly improves the timeliness of detecting high-risk cases, but also achieves efficient collaboration between AI and manual auditing through structured queue push, truly realizing the intelligent supervision goal of "letting data speak and using resources where they are most needed".

[0139] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A method for intelligent medical insurance compliance review based on graph retrieval enhancement, characterized in that, include: Construct a compliance knowledge graph that includes medical insurance policies, rules, and their logical constraints; Clinical entities are extracted from medical records and mapped to corresponding nodes in the compliance knowledge graph, and marked as target mapping nodes; the clinical entities include names of chargeable items, names of surgical procedures, or names of medical consumables; Based on the mapped clinical entities as the query starting point, a multi-hop graph retrieval is performed in the compliance knowledge graph to obtain a policy rule subgraph related to the current diagnosis and treatment behavior; The medical records and the policy rule subgraph are input into the generative reasoning model, and the compliance review results are generated through structured reasoning. The compliance review results include violation determination, policy basis, rectification suggestions and violation confidence level. A violation risk propagation model is constructed based on historical audit data, and the compliance knowledge graph is used to assess the violation risk of the diagnosis and treatment behavior to obtain a risk warning level. Based on the aforementioned confidence level of violation and the aforementioned risk warning level, an audit priority queue is generated for multiple pending medical records.

2. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The method for constructing the compliance knowledge graph includes: The medical insurance policy document is segmented at multiple granularities to generate hierarchical text units covering general instructions, chapters, clauses, and sub-items; Based on the hierarchical text units, corresponding violation scenario nodes are constructed, and the logical constraint relationships between each node are identified through the medical insurance semantic pattern library. The logical constraint relationships include one or more of the following: project connotation inclusion relationship, exclusion content exclusion relationship, and mutual exclusion relationship of not being able to charge at the same time. Based on the violation scenario nodes and their logical constraints, an initial compliance knowledge graph is constructed.

3. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The method for determining the target mapping node for extracting clinical entities includes: Based on a pre-built multi-source coding mapping index library, multiple candidate knowledge graph nodes are generated for the clinical entity. The multi-source coding mapping index library integrates medical insurance catalog codes, ICD surgical codes, UDI consumable codes, and local alias dictionaries. Obtain the contextual information of the clinical entity in the medical record, including the patient diagnosis, the department that performed the procedure, the associated billing items, and the operation time; A semantic compatibility assessment is performed based on the context information and the policy applicability conditions associated with each candidate knowledge graph node to obtain a compatibility score for each candidate node. The policy applicability conditions include one or more of the following: medical institution level restrictions, mandatory accompanying diagnostics, mutually exclusive billing items, or limited use scenarios. Based on the compatibility score, a target mapping node is selected, and structured mapping information containing the mapping result and confidence level is output.

4. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 3, characterized in that, The method for analyzing the compatibility scores of each candidate node includes: The context information is encoded into a context feature vector; The policy applicability conditions associated with each candidate knowledge graph node are parsed into a set of structured constraints; Pre-defined weights are assigned to different types of policy application conditions; among them, prohibitive or restrictive conditions have a higher weight than advisory conditions. Based on the context feature vector and the structured constraint set, a compatibility score is calculated using a weighted semantic matching function. The weighted semantic matching function imposes a significant negative penalty when any hard constraint is not satisfied, and adds a positive contribution to the satisfied constraints according to their weights.

5. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The method for obtaining the policy rule subgraph related to the current medical practice includes: Using the mapped clinical entities as initial query nodes, a query intent representation is generated by combining the current diagnosis and treatment context. In each hop graph traversal, the path expansion priority is dynamically determined based on the semantic relevance of neighboring nodes to the query intent and the effectiveness level of the medical insurance rules corresponding to the connecting edges; wherein, the effectiveness level of the medical insurance rules includes prohibitive rules, restrictive rules and descriptive rules, and the edges corresponding to prohibitive rules enjoy the highest priority in path expansion; Multi-hop retrieval is performed based on the path expansion priority to obtain an initial policy rule subgraph. The initial policy rule subgraph is then semantically redundantly pruned, and node branches with a similarity to the query intent representation below a preset threshold are removed to obtain the policy rule subgraph.

6. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The generation of the compliance review results includes: Construct a chain-thinking reasoning template that includes five stages: fact reconstruction, rule matching, logic verification, violation determination, and rectification suggestions; The policy rule subgraph is converted into a set of structured rule statements with unique identifiers and embedded as rule anchors into the context of cue words in the generative reasoning model; The generative reasoning model fills in the content of each stage in sequence according to the chain-like reasoning template, and each reasoning step must reference the unique identifier of at least one of the rule anchor points and output the compliance review results.

7. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The construction of the aforementioned violation risk propagation model includes: Obtain historical medical insurance audit data, which includes historical audit tags for each charge item, and the historical audit tags indicate the violation status of the item; The knowledge graph nodes mapped to the fee items with the corresponding historical audit labels as violations are taken as positive samples, and the nodes with the corresponding historical audit labels as non-violations are taken as negative samples. A graph neural network is constructed on the compliance knowledge graph. The graph neural network is configured to output a violation risk score for each node and dynamically allocate propagation weights according to the medical insurance rule type of the connecting edges during message transmission. Among them, the edges corresponding to prohibitive rules are given higher propagation weights than the edges of restrictive rules and explanatory rules. The graph neural network is trained under supervision based on the positive and negative samples, so that the violation risk score is consistent with the historical audit label, and a violation risk propagation model is obtained.

8. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The assessment of the violation risk of the aforementioned medical treatment includes: The mapped clinical entities are input as initial activation nodes into the trained violation risk propagation model, and multi-hop risk signal diffusion is performed on the compliance knowledge graph to obtain the violation risk score of each relevant node. Based on the violation risk score, a risk activation subgraph is constructed, and connected component clustering is performed on the risk activation subgraph to identify high-risk project combinations; Calculate the combined risk intensity of each high-risk project combination, which is obtained by weighted fusion of the average violation risk score of the nodes within the combination and the subgraph structure density; The combined risk intensity is mapped to a preset threshold range to determine the risk warning level.

9. The intelligent medical insurance compliance review method based on graph retrieval enhancement generation according to claim 1, characterized in that, The process of generating an audit priority queue for multiple pending medical records includes: Obtain the violation confidence level and risk warning level for each of the multiple pending medical records; The risk warning level is mapped to a preset value, and the violation confidence level and the mapped risk warning level are weighted and fused based on dynamic weights to obtain the audit priority score of each medical record; The multiple pending medical records are sorted according to the audit priority score, and an audit priority queue is generated in combination with the preset audit resource capacity. The audit priority queue is pushed to the medical insurance audit management system to guide manual review.

10. An intelligent medical insurance compliance review system based on graph retrieval enhancement generation, operating based on the intelligent medical insurance compliance review method based on graph retrieval enhancement generation as described in any one of claims 1-9, characterized in that, It includes a compliance knowledge graph construction module, an entity mapping module, a policy and rule subgraph retrieval module, a compliance review module, a risk assessment module, and a priority scheduling module; The compliance knowledge graph construction module is used to build a compliance knowledge graph that includes medical insurance policies, rules, and their logical constraints. The entity mapping module is used to extract clinical entities from medical records and map them to the corresponding nodes in the compliance knowledge graph, marking them as target mapping nodes; The policy rule subgraph retrieval module is used to perform multi-hop graph retrieval in the compliance knowledge graph based on the mapped clinical entity as the query starting point, and obtain the policy rule subgraph related to the current diagnosis and treatment behavior. The compliance review module is used to input the medical records and the policy rule subgraph into the generative reasoning model, and generate compliance review results through structured reasoning; The risk assessment module is used to build a violation risk propagation model based on historical audit data, and to use the compliance knowledge graph to assess the violation risk of the diagnosis and treatment behavior to obtain a risk warning level; The priority scheduling module is used to generate an audit priority queue for multiple pending medical records based on the violation confidence level and the risk warning level.

Images