Search

CN-121997180-A - Medical charging violation detection method based on knowledge graph and graph neural network

CN121997180ACN 121997180 ACN121997180 ACN 121997180ACN-121997180-A

Abstract

The invention belongs to the field of medical charge supervision, and in particular relates to a medical charge violation detection method based on a knowledge graph and a graph neural network, which comprises the steps of acquiring medical charge data, inputting the medical charge data into a trained graph neural network, and obtaining a violation detection result; the training process of the graphic neural network comprises the steps of obtaining a medical charging rule base and medical charging data, preprocessing the medical charging data according to the medical charging rule base, constructing a multi-mode heterogeneous knowledge graph according to the medical charging rule base and the preprocessed medical charging data, generating a graphic sample set according to the medical charging rule base, the preprocessed medical charging data and the multi-mode heterogeneous knowledge graph, training the graphic neural network by using the graphic set to obtain a trained graphic neural network, and simultaneously detecting violation types by constructing a relation among the multi-mode knowledge graph unified expression rule, the project, the price calculating unit and the unstructured text.

Inventors

  • SANG CHUNYAN
  • HU XIAOBAO
  • LIAO SHIGEN

Assignees

  • 重庆邮电大学

Dates

Publication Date
20260508
Application Date
20260204

Claims (10)

  1. 1. A medical charging violation detection method based on a knowledge graph and a graph neural network is characterized by comprising the steps of obtaining medical charging data, inputting the medical charging data into the trained graph neural network to obtain a violation detection result, wherein the training process of the graph neural network comprises the following steps: s1, acquiring a medical charging rule base and medical charging data, and preprocessing the medical charging data according to the medical charging rule base to obtain preprocessed medical charging data; s2, constructing a multi-mode heterogeneous knowledge graph according to the medical charging rule base and the preprocessed medical charging data; S3, generating a pattern book set containing violation labels according to the medical charging rule base, the preprocessed medical charging data and the multi-mode heterogeneous knowledge graph; and S4, training the graphic neural network by using the graphic book set containing the violation labels to obtain a trained graphic neural network.
  2. 2. The medical charging violation detection method based on the knowledge graph and the neural network of claim 1, wherein the medical charging data comprises real charging data and unstructured text data, the real charging data comprises charging records of a plurality of case numbers, the unstructured text data comprises unstructured texts of a plurality of case numbers, and the preprocessing of the medical charging data comprises: s11, cleaning the real charging data, and screening charging records constrained by the medical charging rule base from the cleaned real charging data to obtain preprocessed real charging data; s12, preprocessing the unstructured text of each case number in the unstructured text data to obtain the preprocessed unstructured text data; and S13, combining the preprocessed real charging data with unstructured text data to obtain preprocessed medical charging data.
  3. 3. The medical charging violation detection method based on the knowledge graph and the graph neural network of claim 2, wherein the medical charging rule base comprises a plurality of rules, each rule comprises an item and a charging unit thereof, each charging record comprises a case number, a date, an item and an item charging condition, and constructing the multi-mode heterogeneous knowledge graph comprises: respectively constructing a rule node and characteristics thereof for each rule in the medical charging rule base; respectively constructing a price unit node and characteristics of each price unit in the medical charging rule base; Respectively constructing a project node and characteristics thereof for each project in the preprocessed real charging data; respectively constructing a text semantic node and an embedded vector of the text semantic node for each case number unstructured text in the unstructured text data after preprocessing; Identifying entities in the pre-processed unstructured text data, and respectively constructing an entity node and an embedded vector thereof for each entity; Constructing cross-modal edges among the nodes, wherein the cross-modal edges comprise limited relation edges from item nodes to rule nodes, reverse relation edges from rule nodes to item nodes, possession unit relation edges from item nodes to price unit nodes, reverse relation edges from price unit nodes to item nodes, mutual exclusion relation edges among item nodes, semantic support relation edges between text semantic nodes and item nodes and semantic mapping relation edges between entity nodes and item nodes; The weights of the restricted relation side, the reverse relation side, the owned unit relation side and the exclusive relation side are 1, the weights of the semantic support relation side are the semantic similarity between the text semantic node and the item node, and the weights of the semantic mapping relation side are the semantic similarity between the entity node and the item node.
  4. 4. A method of medical charging violation detection based on knowledge-graph and graph neural networks according to claim 3, characterized in that generating a pattern set comprising violation tags comprises: s31, aggregating charging records of the same case number on the same date in the preprocessed medical charging data to obtain a case-level sample of each case number on each date; S32, carrying out violation detection on each case-level sample according to the medical charging rule base, and generating a violation label of each case-level sample; S33, extracting a plurality of non-violating case-level samples from the case-level samples, and disturbing each extracted case-level sample based on a medical charging rule base to generate a plurality of violating case-level samples; S34, for each case-level sample in the balanced case-level sample set Each item in (a) Generating dynamic characteristics and semantic embedding, wherein m is an index of a case-level sample in a balanced case-level sample set; s35, balancing each case-level sample in the case-level sample set Each item in (a) Respectively adding the dynamic characteristics and semantic embedding of the multiple-mode heterogeneous knowledge patterns to obtain each case-level sample in the balanced case-level sample set Pattern book of (2) ; S36, combining each case-level sample in the balance case-level sample set Pattern book of (2) And obtaining the pattern book set containing the violation labels.
  5. 5. The method for detecting medical charging violations based on knowledge-graph and neural network of claim 4, wherein the method is a case-level sample Items in (a) Generating dynamic features and semantic embedding includes: at the case level sample In acquiring information about an item For charging records concerning items Feature encoding all charging records to obtain items Dynamic characteristics of (2); Extracting case-level samples from multi-modal heterogeneous knowledge-graph Text semantic nodes corresponding to the included case numbers according to the items Aggregating the embedded vectors of the extracted text semantic nodes with the edge weights of the extracted text semantic nodes to obtain items Semantic embedding of (c).
  6. 6. The medical toll violation detection method based on the knowledge graph and the graph neural network of claim 1, wherein the graph neural network comprises an embedding layer, a multi-head relationship attention convolution module, a hierarchical aggregation module and a depth classification module, and training the graph neural network by using a pattern set containing violation labels comprises: S41, pattern book Features of each node in the pattern book are respectively input into the embedded layer to obtain the pattern book Is embedded into the vector of each node of (a); s42, pattern book The embedded vector of the node is input into a multi-head relation attention convolution module to obtain a pattern book Enhancement features of each node in (a); S43, pattern book The enhanced features of all nodes in the network are input into a hierarchical aggregation module to obtain a pattern book Is a multi-scale feature representation of (2); S44, pattern book Is input into a depth classification module to obtain a pattern book Is a rule violation detection result and probability distribution thereof; s45, according to the pattern book Enhanced features and pattern book for each node in (a) And (3) calculating a loss function value according to the probability distribution of the violation detection result, updating parameters of the graph neural network according to the loss function value, and obtaining the trained graph neural network when the loss function value is minimum.
  7. 7. The medical toll violation detection method based on the knowledge graph and the graphic neural network, according to claim 6, is characterized in that the multi-head relationship attention convolution module comprises a multi-layer multi-head relationship attention convolution layer and a multi-layer residual error gating and layer normalization module, wherein the multi-head relationship attention convolution module is used for pattern book The processing of the embedded vector of the node comprises the following steps: S421, pattern book The embedded vector of the node is input into a first multi-head relation attention convolution layer, and the output of the first multi-head relation attention convolution layer is input into a first residual error gating and layer normalizing module; S422, inputting the output of the first residual error gating and layer normalizing module into a second multi-head relationship attention convolution layer, and inputting the output of the second multi-head relationship attention convolution layer into the second residual error gating and layer normalizing module; s423, inputting the output of the current residual error gating and layer normalizing module into a next multi-head relation attention convolution layer, and inputting the output of the next multi-head relation attention convolution layer into the next residual error gating and layer normalizing module; S424, repeating the step S423 until the output of the last layer residual error gating and layer normalizing module is obtained, namely the pattern book Enhanced features of each node in (a).
  8. 8. The medical charging violation detection method based on a knowledge graph and graph neural network according to claim 7, wherein the processing of the output of the current residual gating and layer normalizing module by the next multi-head relationship attention convolution layer comprises: ; ; ; Wherein, the For the dimension of the kth attention head, 、 、 Respectively a query, a key, a value matrix of the edge type r in the kth attention header, The activation function is represented as a function of the activation, For the offset of edge type r in the kth attention header, Representation of a graphic sample Is a node of (2) In the first place Layer multi-headed relationship the eigenvectors of the attention convolution layer, 、 Respectively represent the picture samples Is a node of (2) 、 In the first place The feature vectors of the layer residual gating and layer normalization module, Representing a set of edge types, The number of attention points is indicated, Is a pattern book Is a node of (2) In-edge type A set of the next neighbor nodes, Is a pattern book Is a node of (2) 、 Edge type of (2) In the first place Semantic gating of the individual attention heads, Is a pattern book Is a node of (2) 、 Edge type of (2) In the first place The weight of the individual attention header(s), Representing edge type In the first place A linear transformation matrix of the individual attention heads, Is of edge type In the first place The learnable semantic gating parameters of the individual attention heads, Is a node 、 Edge type of (2) Is a weight of (2).
  9. 9. The medical toll violation detection method based on the knowledge graph and the graphic neural network of claim 7, wherein the hierarchical aggregation module comprises a hierarchical aggregator and a Set Transformer module, wherein the Set Transformer is a Set Transformer, and the hierarchical aggregator pairs the graphic script The processing of the enhanced features of all nodes in the network comprises: Pattern book The enhanced features of all nodes in the graph are input into a hierarchical aggregator to obtain the graph book Is characterized by the aggregation of (3) ; For polymerized features Respectively carrying out mean value pooling and maximum pooling to obtain mean value pooling characteristics Maximum pooling feature ; Pattern book The enhanced features of all the project nodes in the Set are combined to obtain a project enhanced feature Set, and the project enhanced feature Set is input into a Set transform module to obtain a project enhanced feature Set characterization ; Will aggregate features Pooling features of mean Maximum pooling feature Item enhanced feature set characterization Splicing to obtain a pattern book Is a scale feature representation of (c).
  10. 10. The method for detecting medical charging violations based on knowledge-graph and neural network of claim 7, wherein a loss function is used Wherein, the method comprises the steps of, Respectively the weight of the two materials is respectively given, For a classification loss based on the probability distribution of the offending detection results, For supervised contrast loss based on enhanced features of the nodes, For cross-modal alignment loss based on node enhancement features, To combat regularization losses.

Description

Medical charging violation detection method based on knowledge graph and graph neural network Technical Field The invention belongs to the field of medical charge supervision, and particularly relates to a medical charge violation detection method based on a knowledge graph and a graph neural network. Background With the increasing level of medical informatization, the size and complexity of medical charging data is increasing. In order to ensure safe and reasonable use of the medical insurance foundation, the regulatory authorities need to establish an effective violation detection mechanism. Medical charging violations mainly include types of recurring charges, over-standard charges, and mutually exclusive charges. Repeated charging means that the same medical service item is repeatedly charged within a prescribed time period (e.g., daily or every 24 hours) beyond a maximum number of times or duration allowed. The over-standard charging means that the amount of the medical service item charged exceeds a prescribed standard upper limit. Mutually exclusive charges mean that certain medical service items cannot be charged simultaneously, or the cumulative time of simultaneous charges cannot exceed a prescribed limit. Existing medical charging violation detection methods rely mainly on rule matching and statistical analysis. These methods suffer from the following disadvantages: 1. the rule expression is incomplete, and the traditional rule matching method is difficult to process complex rule relations and the relativity between items, so that the semantic relations between the rules can not be effectively utilized. 2. The feature extraction is insufficient, and the existing method mainly depends on simple statistical features (such as charging times and charging amounts) and cannot capture complex relations and context information among projects. 3. The detection of the violation types is single, most methods can only detect single-type violations, and are difficult to simultaneously process multiple violation types such as repeated charging, over-standard charging, mutually exclusive charging and the like. 4. The method is lack of interpretability, the existing method can only give out rule violation judgment results, cannot provide detailed rule violation reason analysis, and is not beneficial to the understanding and processing of supervisory personnel. In recent years, knowledge graph and graph neural network technology has achieved significant results in a number of fields. The knowledge graph can effectively organize the relation between the expression entity and the knowledge graph, and the graph neural network can learn the representation of the nodes and the edges and capture the complex graph structure information. However, the application of knowledge maps and graph neural networks to medical charging violation detection has been relatively few and the accuracy of violation detection is not high enough. Disclosure of Invention In order to solve the technical problems, the medical charging violation detection method based on the knowledge graph and the graph neural network comprises the following steps of obtaining medical charging data, inputting the medical charging data into the trained graph neural network to obtain violation detection results, wherein the training process of the graph neural network comprises the following steps: s1, acquiring a medical charging rule base and medical charging data, and preprocessing the medical charging data according to the medical charging rule base to obtain preprocessed medical charging data; s2, constructing a multi-mode heterogeneous knowledge graph according to the medical charging rule base and the preprocessed medical charging data; S3, generating a pattern book set containing violation labels according to the medical charging rule base, the preprocessed medical charging data and the multi-mode heterogeneous knowledge graph; and S4, training the graphic neural network by using the graphic book set containing the violation labels to obtain a trained graphic neural network. The invention has the beneficial effects that: 1. The method and the device can simultaneously detect rule types and violation types in unstructured texts by constructing a multi-modal knowledge graph to uniformly express the relation between the text semantics and entities of each rule, item, pricing unit and unstructured text, improve the comprehensiveness and accuracy of detection, 2, the method and the device fuse the static characteristics of the item and case-level dynamic characteristics and semantic embedments, more accurately express the semantics and context information of the item, and further improve the accuracy of detection, 3, the method and the device adopt an abnormal composition neural network, can effectively utilize complex relations between rule-item-pricing unit-text semantics-entities, introduce semantic gating mechanism learning node representation, improve th