Medical insurance medical document auditing method and system based on knowledge graph

Abstract

The invention relates to the technical field of medical insurance intelligent auditing, and discloses a medical insurance medical document auditing method and system based on a knowledge graph. The method comprises the steps of obtaining evaluation index data and auditing results of an intelligent auditing system in a hospital in different time periods before and after being online, and determining knowledge graph nodes according to data changes; constructing a medical insurance auditing knowledge graph based on the data of each time point and the logic dependency relationship between the nodes; calculating a baseline level value and a change trend value of the evaluation index, and obtaining auditing influence coefficients of the node in different auditing links after the system is online in combination with graph node logic dependence; and screening nodes exceeding a preset threshold value according to the auditing influence coefficient as nodes to be intervened, and generating an optimized knowledge graph for document auditing. According to the method, the knowledge graph is dynamically constructed by comparing the data before and after the system is online, and the node contribution is quantified by using the influence coefficient to realize accurate optimization, so that the adaptive capacity and the intervention efficiency of the auditing system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical insurance verification technology, specifically to a method and system for verifying medical insurance documents based on knowledge graphs. Background Technology

[0002] Current intelligent medical insurance document review systems generally rely on pre-built, static rule knowledge bases. These rules are typically manually extracted and set by domain experts based on policy documents and historical experience, forming fixed logical judgment conditions. When reviewing documents, the system strictly matches these rules, lacking a quantitative evaluation mechanism for the effectiveness and impact of the rules themselves in a real, dynamic medical review environment. When the review process changes or a new system is launched, it is difficult to accurately identify the key rule nodes affecting the overall review performance.

[0003] Existing technical solutions for constructing knowledge graphs largely rely on static structuring of policy documents and medical knowledge for their nodes and relationships, aiming to enable associative queries and reasoning. This approach severs the dynamic connection between the knowledge graph and actual system performance data. Once established, the graph becomes relatively fixed, unable to be self-verified and optimized through changes in real business metrics generated during system operation. When improvements in review efficiency or accuracy are needed, maintenance personnel can only rely on subjective experience to adjust rules or the graph, lacking data-driven, targeted optimization criteria, resulting in long adjustment cycles and unpredictable effects. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph-based method and system for reviewing medical insurance documents, in order to solve the problems mentioned in the background.

[0005] To achieve the above objectives, this invention provides a knowledge graph-based method for verifying medical insurance documents, the method comprising: Obtain evaluation indicator data for different time periods before and after the launch of the hospital's intelligent review system, and determine the knowledge graph nodes in the review process based on the changes in the evaluation indicator data and the review results. Based on the evaluation indicator data corresponding to each time point before and after the system goes online, and the logical dependencies between the knowledge graph nodes, a medical insurance audit knowledge graph containing knowledge graph nodes and logical dependencies is constructed. Based on the medical insurance review knowledge graph, calculate the baseline level and trend value of the evaluation indicator data before and after the system goes online; Based on the baseline level and trend values ​​of the evaluation indicator data after the system goes online, and combined with the logical dependencies of the knowledge graph nodes in the medical insurance review knowledge graph, the review impact coefficients of the knowledge graph nodes in different review stages after the system goes online are obtained. Based on the audit impact coefficient, knowledge graph nodes with an audit impact coefficient greater than a preset impact threshold are selected as nodes to be intervened. The medical insurance audit knowledge graph containing the nodes to be intervened is used as an optimized audit knowledge graph, and the optimized audit knowledge graph is used to audit medical insurance documents.

[0006] Preferably, the step of acquiring evaluation indicator data for different time periods before and after the launch of the hospital's intelligent review system, and determining knowledge graph nodes in the review process based on the changes in the evaluation indicator data and the review results, includes: The evaluation indicator data consists of the monthly medical insurance refusal amount, refusal rate, and number of refusal details during different time periods before and after the launch of the hospital's intelligent review system. The evaluation indicator data corresponds to the medical insurance medical document review results. Based on the changes in evaluation indicator data such as the monthly medical insurance refusal amount, refusal rate, and number of refusal details during different time periods before and after the system went online, and combined with the medical insurance document review results corresponding to the evaluation indicator data, the knowledge graph nodes of the hospital's intelligent review system in different review stages of the review process are determined. The knowledge graph nodes include document type nodes, review rule nodes, and refusal category nodes. The method for determining the knowledge graph nodes at different stages of the review process in the in-hospital intelligent review system specifically includes: Before the hospital's intelligent review system went online, obtain the monthly medical insurance refusal amount sequence, the refusal rate sequence, and the refusal detail number sequence. The changes in the amount of refusal to pay in adjacent months, the changes in the refusal rate in adjacent months, and the changes in the number of refusal details in adjacent months are calculated before the system goes online. Based on these changes, the set of baseline knowledge graph nodes corresponding to the medical insurance document review results before the system goes online is determined. During the period after the hospital's intelligent review system was launched, the following data were collected: the monthly medical insurance refusal amount sequence after the system was launched; the refusal rate sequence after the system was launched; and the number of refusal details after the system was launched. Align the monthly data after the system goes live with the data before the system goes live. Based on the differences between the sequence of refusal amount after intervention, the sequence of refusal rate after intervention, and the sequence of number of refusal details after intervention and the sequence of refusal amount, refusal rate, and number of refusal details, select knowledge graph nodes from the baseline knowledge graph node set as knowledge graph nodes for different review stages in the hospital's intelligent review system.

[0007] Preferably, the step of selecting knowledge graph nodes from the baseline knowledge graph node set based on the differences between the aligned post-intervention refusal amount sequence, post-intervention refusal rate sequence, and post-intervention refusal detail number sequence and the refusal amount sequence, refusal rate sequence, and refusal detail number sequence specifically includes: By comparing the data of the post-intervention chargeback amount sequence for each corresponding month after the system went online with the data of the chargeback amount sequence before the system went online, the monthly chargeback amount difference sequence is obtained. The data of the post-intervention rejection rate sequence for each corresponding month after the system went online were compared with the data of the rejection rate sequence before the system went online to obtain the monthly rejection rate difference sequence. The data of the number of refusal details after intervention in each corresponding month after the system went online was compared with the data of the number of refusal details before the system went online to obtain the monthly refusal detail difference sequence. Normalize each difference data in the monthly chargeback amount difference sequence, monthly chargeback rate difference sequence, and monthly chargeback detail number difference sequence to obtain a normalized difference value. Select knowledge graph nodes with normalized difference values ​​greater than a preset difference threshold from the baseline knowledge graph node set, and use them as knowledge graph nodes in different review stages of the hospital's intelligent review system.

[0008] Preferably, the step of calculating the baseline level and trend value of the evaluation indicator data before and after the system goes online based on the medical insurance review knowledge graph specifically includes: Based on the logical dependencies between the knowledge graph nodes in the medical insurance review knowledge graph, the correspondence between the knowledge graph nodes before and after the launch of the hospital's intelligent review system is determined. Based on the evaluation index data of the monthly medical insurance refusal amount, refusal rate and number of refusal details before the hospital's intelligent review system went online, and combined with the correspondence between knowledge graph nodes, the change trend value of the evaluation index data corresponding to each knowledge graph node before the system went online was calculated, and the mean of the change trend value was used as the baseline level value. Based on the evaluation index data of the monthly medical insurance refusal amount, refusal rate, and number of refusal details after the launch of the hospital's intelligent review system, and combined with the correspondence between knowledge graph nodes, the changing trend value of the evaluation index data corresponding to each knowledge graph node after the system went online was calculated.

[0009] Preferably, the method for obtaining the audit impact coefficient specifically includes: The audit impact coefficients include the audit impact coefficients for document types, audit impact coefficients for audit rules, and audit impact coefficients for chargeback categories; The trend value of the evaluation index data corresponding to each knowledge graph node after the system goes online is compared with the baseline level value to obtain the level change value of each knowledge graph node. Based on the level change value of each knowledge graph node after the system goes online, and combined with the logical dependency relationship of the knowledge graph nodes in the medical insurance review knowledge graph, the contribution of the level change of the knowledge graph node in different review stages is calculated. The normalized value of the product of the horizontal change value of the knowledge graph node and the contribution of the horizontal change after the system goes online is used as the review influence coefficient of the knowledge graph node in different review stages. The calculation yields the contribution of knowledge graph nodes to changes in their level at different review stages, specifically including: In the medical insurance review knowledge graph, the number of out-degree nodes and the number of in-degree nodes of each knowledge graph node are extracted from the logical dependencies of the knowledge graph nodes. Based on the number of out-degree nodes and in-degree nodes of each knowledge graph node, the node influence of each knowledge graph node is calculated; the product of the node influence of the knowledge graph node and the weight of the knowledge graph node in different review stages is used as the initial contribution of the knowledge graph node. In the medical insurance review knowledge graph, the logical dependencies of knowledge graph nodes are used to form logical dependency paths. The path length of each logical dependency path is obtained, and the product of the reciprocal of the path length and the initial contribution of the knowledge graph node is used as the path contribution of the knowledge graph node. The average path contribution of all knowledge graph nodes corresponding to logically dependent paths is used as the contribution of knowledge graph nodes to the level changes in different review stages.

[0010] Preferably, the step of selecting knowledge graph nodes with an audit impact coefficient greater than a preset impact threshold as nodes to be intervened based on the audit impact coefficient specifically includes: Based on the different categories of audit impact coefficients for document type, audit impact coefficients for audit rules, and audit impact coefficients for refusal categories, the magnitudes of the audit impact coefficients and the preset classification impact thresholds are compared respectively. In the document type's audit impact coefficient, the knowledge graph nodes corresponding to the document type's audit impact coefficient that are greater than the corresponding classification impact threshold are selected as the document type's nodes to be intervened. In the review impact coefficient of the review rules, the knowledge graph nodes corresponding to the review impact coefficients of the review rules that are greater than the classification impact threshold of the review rules are selected as the nodes to be intervened in the review rules. In the review impact coefficient of the chargeback category, the knowledge graph nodes corresponding to the review impact coefficients of the chargeback categories that are greater than the classification impact threshold of the chargeback category are selected as nodes to be intervened in the chargeback category. The nodes to be intervened include nodes to be intervened for document types, nodes to be intervened for audit rules, and nodes to be intervened for refusal categories.

[0011] Preferably, the step of using the medical insurance review knowledge graph containing the nodes to be intervened as the optimized review knowledge graph specifically includes: Obtain the medical insurance review knowledge graph containing all nodes to be intervened, extract the logical dependencies between all nodes to be intervened in the medical insurance review knowledge graph, and construct a sub-knowledge graph of the nodes to be intervened. In the medical insurance review knowledge graph, non-intervention nodes that have a logical dependency relationship with the node to be intervened are regarded as associated nodes; The sub-knowledge graph is then integrated with all associated nodes according to the logical dependencies of the original medical insurance review knowledge graph to form an optimized review knowledge graph that includes nodes to be intervened and associated nodes.

[0012] Preferably, the step of using the optimized audit knowledge graph to audit medical insurance documents specifically includes: Obtain information on the document type, review rules, and historical refusal categories of medical insurance documents pending review; The document type information, audit rule information, and historical chargeback category information are matched with the document type, audit rule, and chargeback category nodes to be intervened in the optimized audit knowledge graph. If the document type information matches the document type of the node to be intervened, then the medical insurance document to be reviewed will be marked as a high-concern document. If the audit rule information matches the node to be intervened in the audit rule, the medical insurance document to be audited will be marked as a high-concern document; If the historical refusal category information matches the intervention node of the refusal category, the medical insurance document to be reviewed will be marked as a high-concern document; For medical insurance documents marked as high-concern documents, perform in-depth rule verification according to the logical dependencies of the optimized audit knowledge graph.

[0013] Preferably, the step of constructing a medical insurance review knowledge graph containing knowledge graph nodes and logical dependencies based on the evaluation indicator data corresponding to each time point before and after the system goes live, and the logical dependencies between the knowledge graph nodes, specifically includes: Starting from the time point before the system went online and ending from the time point after the system went online, we extracted the evaluation indicator data for the monthly medical insurance refusal amount, refusal rate, and number of refusal details corresponding to each time point. The evaluation index data corresponding to each time point is associated with the knowledge graph node to construct an initial knowledge graph with time point as the dimension and knowledge graph node as the entity. In the initial knowledge graph, logical dependencies between knowledge graph nodes are defined based on their sequential relationship in the review process. These logical dependencies include inclusion, dependency, and conflict relationships. The initial knowledge graph containing knowledge graph nodes and logical dependencies is then used as the medical insurance review knowledge graph.

[0014] Preferably, the present invention also includes a medical insurance document review system based on a knowledge graph. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the medical insurance document review method based on the knowledge graph described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Using system launch as a dividing point, the system continuously collects data on changes in key business indicators before and after launch, and correlates this data with specific review conclusions. By comparing and analyzing this time-series data, the system can automatically identify the entities and rules that play a role in the actual changes in review effectiveness and dynamically establish them as core nodes of the knowledge graph. This shifts the source of knowledge graph construction from static clauses to dynamic business evidence, ensuring the causal relationship between graph nodes and fluctuations in real-world review efficiency. The graph itself becomes a mapping model of the system's operational effectiveness.

[0016] Static knowledge graphs lack quantitative metrics for node contribution, leading to vague optimization goals. This solution defines a quantitative indicator called the "review impact coefficient." This indicator integrates the baseline state of a node's relevant business metrics, its trend over time, and the node's position and dependencies within the knowledge graph's logical network. Through calculation, the degree of influence of each node on the final result at different review stages is transformed into a specific numerical value. Maintenance personnel can accurately locate the few key nodes with the greatest impact on current review performance—those worthy of priority attention and adjustment—from the complex knowledge network without relying on vague experience, based on clearly defined coefficient thresholds. This transforms the system's iterative optimization from extensive trial and error to targeted, precise intervention, improving operational efficiency and the effectiveness of optimization measures. Attached Figure Description

[0017] Figure 1This is a schematic diagram illustrating the working principle of the knowledge graph-based medical insurance document verification method described in this invention. Figure 2 A flowchart for filtering knowledge graph nodes; Figure 3 A flowchart for calculating baseline level values ​​and trend values; Figure 4 An analysis diagram of the impact and contribution of knowledge graph nodes for medical insurance review; Figure 5 A graph for screening the impact coefficients of knowledge graph nodes in medical insurance review. Detailed Implementation

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

[0019] Please see Figure 1 This invention provides a knowledge graph-based method for reviewing medical insurance documents. The method includes: acquiring evaluation indicator data for different time periods before and after the launch of an in-hospital intelligent review system; determining knowledge graph nodes in the review process based on changes in the evaluation indicator data and review results; constructing a medical insurance review knowledge graph containing knowledge graph nodes and logical dependencies based on the evaluation indicator data at each time point before and after system launch, and the logical dependencies between knowledge graph nodes; calculating the baseline level and trend values ​​of the evaluation indicator data before and after system launch based on the medical insurance review knowledge graph; obtaining the review impact coefficient of knowledge graph nodes in different review stages after system launch based on the baseline level and trend values ​​of the evaluation indicator data after system launch, combined with the logical dependencies of knowledge graph nodes in the medical insurance review knowledge graph; and selecting knowledge graph nodes with review impact coefficients greater than a preset impact threshold as nodes to be intervened in based on the review impact coefficients. The medical insurance review knowledge graph containing the nodes to be intervened in is then used as an optimized review knowledge graph, and the optimized review knowledge graph is used to review medical insurance documents.

[0020] In one embodiment of the present invention, see [reference] Figure 2In practice, evaluation indicator data is obtained, comprising monthly medical insurance refusal amounts, refusal rates, and the number of refusal details, across different time periods before and after the launch of the hospital's intelligent review system. These evaluation indicator data directly correspond to the results of medical insurance document review. Based on the changes in these evaluation indicator data across different time periods before and after the system's launch, and combined with the corresponding medical insurance document review results, knowledge graph nodes for different review stages in the hospital's intelligent review system are determined. These knowledge graph nodes include document type nodes, review rule nodes, and refusal category nodes. In some embodiments, the method for determining the knowledge graph nodes of different review stages in the review process of the in-hospital intelligent review system specifically includes the following operations: During the period before the in-hospital intelligent review system goes online, obtain the monthly medical insurance refusal amount sequence, the monthly refusal rate sequence, and the monthly refusal detail number sequence; calculate the change values ​​of adjacent months in the refusal amount sequence, the change values ​​of adjacent months in the refusal rate sequence, and the change values ​​of adjacent months in the refusal detail number sequence before the system goes online, and determine the baseline knowledge graph node set corresponding to the medical insurance medical document review results before the system goes online based on these change values. During the period following the launch of the hospital's intelligent review system, the following data were collected: monthly medical insurance refusal amount sequence, monthly refusal rate sequence, and monthly refusal detail number sequence. The monthly data before and after the system launch were aligned. Based on the differences between the aligned post-intervention refusal amount, post-intervention refusal rate, and post-intervention refusal detail number sequences and the refusal amount, refusal rate, and refusal detail number sequences, knowledge graph nodes were selected from the baseline knowledge graph node set to serve as knowledge graph nodes for different review stages in the hospital's intelligent review system.

[0021] In specific implementation, the data of the post-intervention refusal amount sequence for each corresponding month after the system goes live is compared with the data of the pre-system launch refusal amount sequence to obtain a monthly refusal amount difference sequence. The data of the post-intervention refusal rate sequence for each corresponding month after the system goes live is compared with the data of the pre-system launch refusal rate sequence to obtain a monthly refusal rate difference sequence. The data of the post-intervention refusal detail quantity sequence for each corresponding month after the system goes live is compared with the data of the pre-system launch refusal detail quantity sequence to obtain a monthly refusal detail quantity difference sequence. It can be understood that each difference data in the monthly refusal amount difference sequence, monthly refusal rate difference sequence, and monthly refusal detail quantity difference sequence is normalized to obtain a normalized difference value. Knowledge graph nodes with normalized difference values ​​greater than a preset difference threshold are selected from the baseline knowledge graph node set and used as knowledge graph nodes for different review stages in the hospital's intelligent review system. Optionally, the formula used for normalization is:

[0022] in: Indicates the first Monthly Normalized difference values ​​of each indicator series Indicates the first Monthly The original difference values ​​of each indicator series, Indicates the first The minimum value among all monthly raw difference values ​​of the indicator series. Indicates the first The maximum value among all monthly raw differences in the indicator sequence.

[0023] In one embodiment of the present invention, see [reference] Figure 3In specific implementation, based on the logical dependencies between knowledge graph nodes in the medical insurance review knowledge graph, the correspondence between knowledge graph nodes before and after the launch of the hospital's intelligent review system is determined. Based on the evaluation indicator data of monthly medical insurance refusal amounts, refusal rates, and the number of refusal details before the launch of the hospital's intelligent review system, and combined with the correspondence between knowledge graph nodes, the trend value of the evaluation indicator data corresponding to each knowledge graph node before the system launch is calculated, and the mean of the trend value is used as the baseline level value. Based on the evaluation indicator data of monthly medical insurance refusal amounts, refusal rates, and the number of refusal details after the launch of the hospital's intelligent review system, and combined with the correspondence between knowledge graph nodes, the trend value of the evaluation indicator data corresponding to each knowledge graph node after the system launch is calculated. In some embodiments, the trend value can be calculated based on the monthly evaluation indicator data sequence associated with each knowledge graph node, using linear fitting or moving average methods to obtain the slope of the sequence over time or the average change within a specific period.

[0024] In specific implementation, the methods for obtaining the audit impact coefficient include: audit impact coefficients for document types, audit impact coefficients for audit rules, and audit impact coefficients for refusal categories. The trend value of the evaluation indicator data corresponding to each knowledge graph node after the system goes live is compared with the baseline level value to obtain the level change value of each knowledge graph node. Based on the level change value of each knowledge graph node after the system goes live, combined with the logical dependencies of knowledge graph nodes in the medical insurance audit knowledge graph, the contribution of the level change of the knowledge graph node in different audit stages is calculated. The normalized value of the product of the level change value of the knowledge graph node after the system goes live and the contribution of the level change is used as the audit impact coefficient of the knowledge graph node in different audit stages. It can be understood that the level change value of a knowledge graph node directly reflects the magnitude of change of the evaluation indicators associated with that node after the system goes live.

[0025] In practical implementation, the contribution of knowledge graph nodes to the level changes in different review stages is calculated, specifically including the following operations: In the medical insurance review knowledge graph, the number of out-degree nodes and in-degree nodes of each knowledge graph node are extracted from the logical dependencies of the knowledge graph nodes. Based on the number of out-degree nodes and in-degree nodes of each knowledge graph node, the node influence of each knowledge graph node is calculated. The product of the node influence of the knowledge graph node and the weight of the knowledge graph node in different review stages is used as the initial contribution of the knowledge graph node. Optionally, the weight of the knowledge graph node in different review stages is pre-set according to the importance of the review stage in the overall process. In the medical insurance review knowledge graph, the logical dependencies of knowledge graph nodes are used to form logical dependency paths. The path length of each logical dependency path is obtained, and the product of the reciprocal of the path length and the initial contribution of the knowledge graph node is used as the path contribution of the knowledge graph node. The average of the path contributions of all knowledge graph nodes corresponding to logical dependency paths is used as the contribution of the knowledge graph node to the level changes in different review stages.

[0026] In some embodiments, the formula used to calculate the node influence of a knowledge graph node is:

[0027] in: Indicates the first The influence of a knowledge graph node. Indicates the first The number of out-degree nodes of a knowledge graph node. Indicates the first The number of in-degree nodes for each knowledge graph node.

[0028] In one embodiment of the present invention, the review impact coefficients are compared with preset classification impact thresholds according to different classifications: the review impact coefficient of document type, the review impact coefficient of review rule, and the review impact coefficient of refusal category. Among the review impact coefficients of document types, knowledge graph nodes corresponding to the review impact coefficients of document types that are greater than the classification impact threshold are selected as nodes to be intervened in for that document type. Similarly, among the review impact coefficients of review rules, knowledge graph nodes corresponding to the review impact coefficients of review rules that are greater than the classification impact threshold are selected as nodes to be intervened in for that review rule. Likewise, among the review impact coefficients of refusal categories, knowledge graph nodes corresponding to the review impact coefficients of refusal categories that are greater than the classification impact threshold are selected as nodes to be intervened in for that refusal category. The nodes to be intervened in include those for document types, those for review rules, and those for refusal categories.

[0029] In practice, the impact coefficients of document types, audit rules, and refusal categories are compared with preset category impact thresholds. For document types, knowledge graph nodes corresponding to their impact coefficients exceeding the corresponding category impact threshold are selected as nodes requiring intervention. Similarly, for audit rules, knowledge graph nodes corresponding to their impact coefficients exceeding the corresponding category impact threshold are selected as nodes requiring intervention. Likewise, for refusal categories, knowledge graph nodes corresponding to their impact coefficients exceeding the corresponding category impact threshold are selected as nodes requiring intervention. In essence, these operations result in nodes requiring intervention for document types, audit rules, and refusal categories.

[0030] In some embodiments, the classification impact threshold corresponding to the document type, the classification impact threshold corresponding to the audit rule, and the classification impact threshold corresponding to the refusal category are three independently set values. In specific implementation, the comparison operation is performed independently for each type of audit impact coefficient. For example, for all calculated document type audit impact coefficients, they are compared one by one with the preset document type classification impact threshold. Document type nodes whose audit impact coefficient is greater than the document type classification impact threshold are marked as nodes requiring intervention. The same comparison and filtering logic is performed for the audit impact coefficients of audit rules and refusal categories, but the audit rule classification impact threshold and the refusal category classification impact threshold are used as the filtering criteria, respectively. Optionally, the classification impact threshold can be determined based on the statistical quantiles of the audit impact coefficients of various types of nodes in historical audit data analysis.

[0031] In some embodiments, the filtering process can be described by a conditional relation: ,in, Show the first The review impact coefficient of a knowledge graph node within its own category. Indicates the category to which the node belongs. The corresponding preset classification influence threshold. When the above relationship holds, the first... A knowledge graph node was selected as the node to be intervened under this category. It is understood that the three categories of nodes—document type, review rule, and refusal category—have different semantics and functions in the knowledge graph; therefore, setting classification impact thresholds for each allows for differentiated sensitivity control. Optionally, the node to be intervened for document type might correspond to specific types such as "outpatient Western medicine prescription" or "inpatient surgery fee details," the node to be intervened for review rule might correspond to "indication and treatment item matching rule" or "drug overdose rule," and the node to be intervened for refusal category might correspond to categories such as "restricted medication" or "non-medical insurance catalog." In this way, a specific set of nodes whose review impact coefficient significantly exceeds the general level of this type of node after the hospital's intelligent review system is implemented is accurately identified from all knowledge graph nodes.

[0032] In one embodiment of the present invention, a medical insurance review knowledge graph containing all nodes to be intervened is obtained. Logical dependencies between all nodes to be intervened in the medical insurance review knowledge graph are extracted, and a sub-knowledge graph of the nodes to be intervened is constructed. In the medical insurance review knowledge graph, non-nodes that have logical dependencies with the nodes to be intervened are designated as associated nodes. The sub-knowledge graph and all associated nodes are merged according to the logical dependencies of the original medical insurance review knowledge graph to form an optimized review knowledge graph containing the nodes to be intervened and associated nodes. Obtain the document type information, review rule information, and historical refusal category information of the medical insurance documents to be reviewed; match the document type information, review rule information, and historical refusal category information with the nodes to be intervened for document type, review rule, and refusal category in the optimized review knowledge graph; if the document type information matches the node to be intervened for document type, mark the medical insurance document to be reviewed as a high-concern document; if the review rule information matches the node to be intervened for review rule, mark the medical insurance document to be reviewed as a high-concern document; if the historical refusal category information matches the node to be intervened for refusal category, mark the medical insurance document to be reviewed as a high-concern document; for medical insurance documents marked as high-concern documents, perform in-depth rule verification according to the logical dependencies of the optimized review knowledge graph.

[0033] In specific implementation, a medical insurance review knowledge graph containing all nodes to be intervened is obtained. Logical dependencies between all nodes to be intervened in the medical insurance review knowledge graph are extracted, and a sub-knowledge graph of the nodes to be intervened is constructed. In the medical insurance review knowledge graph, non-intervention nodes that have logical dependencies with the nodes to be intervened are designated as associated nodes. The sub-knowledge graph and all associated nodes are merged according to the logical dependencies of the original medical insurance review knowledge graph to form an optimized review knowledge graph containing the nodes to be intervened and associated nodes. In some embodiments, the extraction of logical dependencies is achieved by traversing all edges in the medical insurance review knowledge graph that start or end with the node to be intervened. If both nodes connected by these edges are nodes to be intervened, they are included in the sub-knowledge graph; if one end is a node to be intervened and the other end is a non-intervention node, the non-intervention node is identified as an associated node.

[0034] In practice, the medical insurance documents are reviewed using an optimized review knowledge graph. This includes the following operations: obtaining document type information, review rule information, and historical refusal category information for the medical insurance documents to be reviewed; matching the document type information, review rule information, and historical refusal category information with the nodes to be intervened for document type, review rule, and refusal category in the optimized review knowledge graph; if the document type information matches a node to be intervened for a document type, the medical insurance document to be reviewed is marked as a high-concern document; if the review rule information matches a node to be intervened for a review rule, the medical insurance document to be reviewed is marked as a high-concern document; if the historical refusal category information matches a node to be intervened for a refusal category, the medical insurance document to be reviewed is marked as a high-concern document; and for medical insurance documents marked as high-concern documents, in-depth rule verification is performed according to the logical dependencies of the optimized review knowledge graph. As can be understood, the matching process involves a precise comparison of information with node attributes. For example, if the document type information "Inpatient Surgery Fee Details" matches the "Inpatient Surgery Fee Details" attribute stored in the node to be intervened for the same document type, then the match is considered successful. See Table 1.

[0035] Table 1: Relationship between nodes to be intervened and associated nodes Nodes to be intervened (type) Node to be intervened (name) Associated Nodes (Type) Associated node (name) Logical Dependencies Review Rules Drug Overdose Rules Chargeback categories Overdose point to Document Type outpatient Western medicine prescriptions Review Rules Indication matching rules Include Chargeback categories Restricted medication Review Rules Medical insurance catalog restriction rules Dependence Optionally, deep rule verification refers to performing more complex and rigorous multi-rule reasoning and data validation than conventional verification, based on the specific logical dependency paths connecting the nodes to be intervened and related nodes in the optimized audit knowledge graph. In some embodiments, the fusion process of constructing the optimized audit knowledge graph can be formally represented as:

[0036] in: This represents the final optimized audit knowledge graph. This represents the sub-knowledge graph of the node to be intervened in. Represents the set of all associated nodes. This represents the set of logical dependencies (edges) between nodes to be intervened and associated nodes, as well as between associated nodes, in the original medical insurance review knowledge graph.

[0037] See Figure 4 This is an impact and contribution analysis chart of nodes in the medical insurance review knowledge graph. It is used to quantitatively display the review impact coefficient, node contribution, and overall impact of different nodes. The overall impact of "Drug Excess Rule" and "Excessive Drug Use" is significantly higher than that of other nodes, reflecting that these nodes are core influencing factors in the review process. The overall impact of "Indication Matching Rule" is the lowest, corresponding to its secondary position in the review logic. This type of chart is a core tool for optimizing the medical insurance review knowledge graph. It locates core nodes in the review process, assists in screening nodes to be intervened, quantifies the "direct impact-logical contribution" relationship of nodes, optimizes the node weight configuration of the knowledge graph, and provides a basis for node priority for in-depth verification of high-concern documents.

[0038] In one embodiment of the present invention, starting from the time point before the system goes live and ending at the time point after the system goes live, evaluation indicator data for the monthly medical insurance refusal amount, refusal rate, and number of refusal details corresponding to each time point are extracted. The evaluation indicator data corresponding to each time point is associated with knowledge graph nodes to construct an initial knowledge graph with time points as the dimension and knowledge graph nodes as entities. In the initial knowledge graph, logical dependencies between knowledge graph nodes are defined based on their sequential relationship in the review process. These logical dependencies include inclusion, dependency, and conflict relationships. The initial knowledge graph containing knowledge graph nodes and logical dependencies is used as the medical insurance review knowledge graph.

[0039] In the specific implementation, starting from the time point before the hospital's intelligent review system went live and ending at the time point after its launch, evaluation indicator data for each time point, including the monthly medical insurance refusal amount, refusal rate, and number of refusal details, were extracted. The evaluation indicator data for each time point was then associated with knowledge graph nodes to construct an initial knowledge graph with time points as the dimension and knowledge graph nodes as entities. Within the initial knowledge graph, logical dependencies between knowledge graph nodes were defined based on their sequential relationship in the review process. These logical dependencies included inclusion, dependency, and conflict relationships. The initial knowledge graph containing knowledge graph nodes and logical dependencies was then used as the medical insurance review knowledge graph.

[0040] In some embodiments, time points are extracted monthly, starting from the twelfth month before the launch of the hospital's intelligent review system and ending in the sixth month after its launch, thus forming a sequence of eighteen consecutive time points. In specific implementation, associating the evaluation indicator data corresponding to each time point with a knowledge graph node involves creating attribute fields for each knowledge graph node entity in the initial knowledge graph. These fields store the specific values ​​of the medical insurance reimbursement amount, reimbursement rate, and number of reimbursement details associated with that node at each time point. Optionally, the association operation can be implemented by establishing a triplet data table of "node-time-indicator value," where each triplet identifies the specific value of a specific evaluation indicator recorded at a specific time point and a specific knowledge graph node. It is understandable that defining logical dependencies is based on the inherent logic of the audit business process. For example, there is an "inclusion" relationship between the "outpatient Western medicine prescription" document type node and the "drug over-dosage rule" audit rule node, indicating that this rule must be applied to this document type; there is a "dependency" relationship between the "drug over-dosage rule" audit rule node and the "over-dosage" refusal category node, indicating that violating this rule may lead to such refusal; and there may be a "conflict" relationship between the "drugs within the medical insurance catalog" audit rule node and the "non-medical insurance catalog" audit rule node.

[0041] In some embodiments, the definition and assignment of logical dependencies can be described using a relation definition function:

[0042] in: and These represent any two knowledge graph nodes in the medical insurance review knowledge graph. It is a relational definition function, and its output is... It is a numerical value representing the type and strength of a relationship. A specific positive value indicates an "inclusion" or "dependency" relationship, a specific negative value indicates a "conflict" relationship, and a zero value indicates no direct logical dependency. It can be understood that through this functional definition, the complex relationships between knowledge graph nodes are transformed into a computable structure. In practical implementation, the initial knowledge graph containing knowledge graph nodes and logical dependencies is used as the medical insurance review knowledge graph. This means that the final medical insurance review knowledge graph is a dynamic, multi-dimensional graph structure, where nodes carry evaluation indicator data across time series, and edges represent logical dependencies between nodes with type and strength attributes.

[0043] See Figure 5This is a screening chart of the impact coefficients of nodes in the medical insurance review knowledge graph. It is used to identify nodes whose impact coefficients exceed a threshold and require intervention. The impact coefficients of "Drug Over-dosage Rules," "Over-dosage," and "Out-of-scope Drug Use" are all ≥0.8, exceeding the preset threshold, and are therefore identified as nodes requiring intervention. The impact coefficients of the remaining nodes are below the threshold and are considered normal nodes. The impact coefficients of rule-based and refusal-based nodes are generally higher than those of document-based nodes, reflecting the characteristic that "review rules and refusal reasons have a more significant impact on the review results." This type of chart is a core screening tool for optimizing the medical insurance review knowledge graph. It quickly locates high-impact nodes, assists in constructing an optimized knowledge graph that includes nodes requiring intervention, verifies the rationality of the distribution of review impact coefficients, and optimizes threshold settings to improve screening accuracy.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-based method for verifying medical insurance documents, characterized in that, The method includes the following steps: Obtain evaluation indicator data for different time periods before and after the launch of the hospital's intelligent review system, and determine the knowledge graph nodes in the review process based on the changes in the evaluation indicator data and the review results. Based on the evaluation indicator data corresponding to each time point before and after the system goes online, and the logical dependencies between the knowledge graph nodes, a medical insurance audit knowledge graph containing knowledge graph nodes and logical dependencies is constructed. Based on the medical insurance review knowledge graph, calculate the baseline level and trend value of the evaluation indicator data before and after the system goes online; Based on the baseline level and trend values ​​of the evaluation indicator data after the system goes online, and combined with the logical dependencies of the knowledge graph nodes in the medical insurance review knowledge graph, the review impact coefficients of the knowledge graph nodes in different review stages after the system goes online are obtained. Based on the audit impact coefficient, knowledge graph nodes with an audit impact coefficient greater than a preset impact threshold are selected as nodes to be intervened. The medical insurance audit knowledge graph containing the nodes to be intervened is used as an optimized audit knowledge graph, and the optimized audit knowledge graph is used to audit medical insurance documents.

2. The knowledge graph-based medical insurance document verification method according to claim 1, characterized in that, The process involves acquiring evaluation indicator data from different time periods before and after the launch of the hospital's intelligent review system, and determining knowledge graph nodes in the review process based on the changes in the evaluation indicator data and the review results. The evaluation indicator data consists of the monthly medical insurance refusal amount, refusal rate, and number of refusal details during different time periods before and after the launch of the hospital's intelligent review system. The evaluation indicator data corresponds to the medical insurance medical document review results. Based on the changes in evaluation indicator data such as the monthly medical insurance refusal amount, refusal rate, and number of refusal details during different time periods before and after the system went online, and combined with the medical insurance document review results corresponding to the evaluation indicator data, the knowledge graph nodes of the hospital's intelligent review system in different review stages of the review process are determined. The knowledge graph nodes include document type nodes, review rule nodes, and refusal category nodes. The method for determining the knowledge graph nodes at different stages of the review process in the in-hospital intelligent review system specifically includes: Before the hospital's intelligent review system went online, obtain the monthly medical insurance refusal amount sequence, the refusal rate sequence, and the refusal detail number sequence. The changes in the amount of refusal to pay in adjacent months, the changes in the refusal rate in adjacent months, and the changes in the number of refusal details in adjacent months are calculated before the system goes online. Based on these changes, the set of baseline knowledge graph nodes corresponding to the medical insurance document review results before the system goes online is determined. During the period after the hospital's intelligent review system was launched, the following data were collected: the monthly medical insurance refusal amount sequence after the system was launched; the refusal rate sequence after the system was launched; and the number of refusal details after the system was launched. Align the monthly data after the system goes live with the data before the system goes live. Based on the differences between the sequence of refusal amount after intervention, the sequence of refusal rate after intervention, and the sequence of number of refusal details after intervention and the sequence of refusal amount, refusal rate, and number of refusal details, select knowledge graph nodes from the baseline knowledge graph node set as knowledge graph nodes for different review stages in the hospital's intelligent review system.

3. The knowledge graph-based medical insurance document verification method according to claim 2, characterized in that, The method involves selecting knowledge graph nodes from the baseline knowledge graph node set based on the differences between the aligned post-intervention refusal amount sequence, post-intervention refusal rate sequence, and post-intervention refusal detail number sequence and the refusal amount sequence, refusal rate sequence, and refusal detail number sequence. Specifically, this includes: By comparing the data of the post-intervention chargeback amount sequence for each corresponding month after the system went online with the data of the chargeback amount sequence before the system went online, the monthly chargeback amount difference sequence is obtained. The data of the post-intervention rejection rate sequence for each corresponding month after the system went online were compared with the data of the rejection rate sequence before the system went online to obtain the monthly rejection rate difference sequence. The data of the number of refusal details after intervention in each corresponding month after the system went online was compared with the data of the number of refusal details before the system went online to obtain the monthly refusal detail difference sequence. Normalize each difference data in the monthly chargeback amount difference sequence, monthly chargeback rate difference sequence, and monthly chargeback detail number difference sequence to obtain a normalized difference value. Select knowledge graph nodes with normalized difference values ​​greater than a preset difference threshold from the baseline knowledge graph node set, and use them as knowledge graph nodes in different review stages of the hospital's intelligent review system.

4. The knowledge graph-based medical insurance document verification method according to claim 1, characterized in that, The calculation of the baseline level and trend values ​​of the evaluation indicator data before and after the system's launch, based on the medical insurance review knowledge graph, specifically includes: Based on the logical dependencies between the knowledge graph nodes in the medical insurance review knowledge graph, the correspondence between the knowledge graph nodes before and after the launch of the hospital's intelligent review system is determined. Based on the evaluation index data of the monthly medical insurance refusal amount, refusal rate and number of refusal details before the hospital's intelligent review system went online, and combined with the correspondence between knowledge graph nodes, the change trend value of the evaluation index data corresponding to each knowledge graph node before the system went online was calculated, and the mean of the change trend value was used as the baseline level value. Based on the evaluation index data of the monthly medical insurance refusal amount, refusal rate, and number of refusal details after the launch of the hospital's intelligent review system, and combined with the correspondence between knowledge graph nodes, the changing trend value of the evaluation index data corresponding to each knowledge graph node after the system went online was calculated.

5. The knowledge graph-based medical insurance document verification method according to claim 4, characterized in that, The method for obtaining the audit impact coefficient specifically includes: The audit impact coefficients include the audit impact coefficients for document types, audit impact coefficients for audit rules, and audit impact coefficients for chargeback categories; The trend value of the evaluation index data corresponding to each knowledge graph node after the system goes online is compared with the baseline level value to obtain the level change value of each knowledge graph node. Based on the level change value of each knowledge graph node after the system goes online, and combined with the logical dependency relationship of the knowledge graph nodes in the medical insurance review knowledge graph, the contribution of the level change of the knowledge graph node in different review stages is calculated. The normalized value of the product of the horizontal change value of the knowledge graph node and the contribution of the horizontal change after the system goes online is used as the review influence coefficient of the knowledge graph node in different review stages. The calculation yields the contribution of knowledge graph nodes to changes in their level at different review stages, specifically including: In the medical insurance review knowledge graph, the number of out-degree nodes and the number of in-degree nodes of each knowledge graph node are extracted from the logical dependencies of the knowledge graph nodes. Based on the number of out-degree nodes and in-degree nodes of each knowledge graph node, the node influence of each knowledge graph node is calculated; the product of the node influence of the knowledge graph node and the weight of the knowledge graph node in different review stages is used as the initial contribution of the knowledge graph node. In the medical insurance review knowledge graph, the logical dependencies of knowledge graph nodes are used to form logical dependency paths. The path length of each logical dependency path is obtained, and the product of the reciprocal of the path length and the initial contribution of the knowledge graph node is used as the path contribution of the knowledge graph node. The average path contribution of all knowledge graph nodes corresponding to logically dependent paths is used as the contribution of knowledge graph nodes to the level changes in different review stages.

6. The knowledge graph-based medical insurance document verification method according to claim 1, characterized in that, The step of selecting knowledge graph nodes with an impact coefficient greater than a preset impact threshold as nodes to be intervened based on the review impact coefficient specifically includes: Based on the different categories of audit impact coefficients for document type, audit impact coefficients for audit rules, and audit impact coefficients for refusal categories, the magnitudes of the audit impact coefficients and the preset classification impact thresholds are compared respectively. In the document type's audit impact coefficient, the knowledge graph nodes corresponding to the document type's audit impact coefficient that are greater than the corresponding classification impact threshold are selected as the document type's nodes to be intervened. In the review impact coefficient of the review rules, the knowledge graph nodes corresponding to the review impact coefficients of the review rules that are greater than the classification impact threshold of the review rules are selected as the nodes to be intervened in the review rules. In the review impact coefficient of the chargeback category, the knowledge graph nodes corresponding to the review impact coefficients of the chargeback categories that are greater than the classification impact threshold of the chargeback category are selected as nodes to be intervened in the chargeback category. The nodes to be intervened include nodes to be intervened for document types, nodes to be intervened for audit rules, and nodes to be intervened for refusal categories.

7. The knowledge graph-based medical insurance document verification method according to claim 6, characterized in that, The step of using the medical insurance review knowledge graph containing the nodes to be intervened as the optimized review knowledge graph specifically includes: Obtain the medical insurance review knowledge graph containing all nodes to be intervened, extract the logical dependencies between all nodes to be intervened in the medical insurance review knowledge graph, and construct a sub-knowledge graph of the nodes to be intervened. In the medical insurance review knowledge graph, non-intervention nodes that have a logical dependency relationship with the node to be intervened are regarded as associated nodes; The sub-knowledge graph is then integrated with all associated nodes according to the logical dependencies of the original medical insurance review knowledge graph to form an optimized review knowledge graph that includes nodes to be intervened and associated nodes.

8. The knowledge graph-based medical insurance document verification method according to claim 7, characterized in that, The process of using the optimized audit knowledge graph to audit medical insurance documents specifically includes: Obtain information on the document type, review rules, and historical refusal categories of medical insurance documents pending review; The document type information, audit rule information, and historical chargeback category information are matched with the document type, audit rule, and chargeback category nodes to be intervened in the optimized audit knowledge graph. If the document type information matches the document type of the node to be intervened, then the medical insurance document to be reviewed will be marked as a high-concern document. If the audit rule information matches the node to be intervened in the audit rule, the medical insurance document to be audited will be marked as a high-concern document; If the historical refusal category information matches the intervention node of the refusal category, the medical insurance document to be reviewed will be marked as a high-concern document; For medical insurance documents marked as high-concern documents, perform in-depth rule verification according to the logical dependencies of the optimized audit knowledge graph.

9. The knowledge graph-based medical insurance document verification method according to claim 1, characterized in that, The step of constructing a medical insurance review knowledge graph, which includes knowledge graph nodes and logical dependencies, based on the evaluation indicator data corresponding to each time point before and after the system goes live, and the logical dependencies between the knowledge graph nodes, specifically includes: Starting from the time point before the system went online and ending from the time point after the system went online, we extracted the evaluation indicator data for the monthly medical insurance refusal amount, refusal rate, and number of refusal details corresponding to each time point. The evaluation index data corresponding to each time point is associated with the knowledge graph node to construct an initial knowledge graph with time point as the dimension and knowledge graph node as the entity. In the initial knowledge graph, logical dependencies between knowledge graph nodes are defined based on their sequential relationship in the review process. These logical dependencies include inclusion, dependency, and conflict relationships. The initial knowledge graph containing knowledge graph nodes and logical dependencies is then used as the medical insurance review knowledge graph.

10. A knowledge graph-based medical insurance document verification system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the knowledge graph-based medical insurance document verification method according to any one of claims 1 to 9.

Images