A method, device, equipment and storage medium for identifying medical insurance fraud

Abstract

The embodiment of the present invention provides a method, device, equipment and storage medium for identifying medical insurance fraud, which relates to the field of medical big data technology. Among them, this identification method includes steps S1 to S5. S1. Obtain medical data and construct a medical insurance heterogeneous graph based on the medical data. S2. Based on the multi-semantic meta-paths of different trajectories, obtain the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories, and fuse the neighbor nodes of each target node under different trajectories with each target node respectively to obtain the semantic representation of each target node under different trajectories. S3. Fuse the semantic representations of each target node under different trajectories to generate the spatial nodes of each target node. S4. Based on the time sequence, obtain the time series of the spatial nodes of each target node. S5. According to the time series, classify each target node through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is a medical insurance fraud.

Description

Technical Field

[0001] The present invention relates to the field of medical big data technology, and in particular to a method, device, equipment and storage medium for identifying medical insurance fraud. Background Art

[0002] Medical insurance generally involves stakeholders such as the insured, designated medical institutions (including hospitals and pharmacies), and doctors. There are many links, a long chain, and many risk points, which can easily breed medical insurance fraud.

[0003] Traditional medical insurance fraud detection methods such as machine learning and anomaly detection often only focus on feature attributes, ignoring the rich behavioral attributes in the medical process, which can also serve as important evidence for fraud detection. For example, a fraudulent patient steals a large number of the same drugs from multiple hospitals in the same period of time, or prescribes a large number of drugs that are not related to a department in a hospital. These behavioral attributes are difficult to reflect in features.

[0004] Therefore, how to identify fraudsters from complex medical insurance data is an urgent problem that needs to be solved in order to ensure the healthy operation of the medical insurance system and the rational use of medical insurance funds.

[0005] In view of this, the applicant filed this application after studying the existing technology. Summary of the invention

[0006] The present invention provides a medical insurance fraud identification method, device, equipment and storage medium to improve the above technical problems.

[0007] First,

[0008] An embodiment of the present invention provides a method for identifying medical insurance fraud, which includes steps S1 to S5.

[0009] S1. Obtain medical data and construct a medical insurance heterogeneous graph based on the medical data. The medical insurance heterogeneous graph uses entities as nodes and relationships as edges.

[0010] S2. Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the neighbor nodes of each target node under different trajectories are respectively merged with each target node to obtain the semantic representation of each target node under different trajectories.

[0011] S3. Fuse the semantic representations of each target node under different trajectories to generate the spatial nodes of each target node.

[0012] S4. Based on the time sequence, obtain the time series of the spatial nodes of each target node.

[0013] S5. According to the time series, each target node is classified through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0014] In an optional embodiment, step S2 specifically includes:

[0015] Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and steps S21 to S23 are performed on each target node under different trajectories respectively to obtain the semantic representation of each target node under different trajectories:

[0016] S21, obtain the attention correlation coefficient of the neighboring nodes of the target node in the current trajectory. is the attention correlation coefficient of the jth neighbor node of the ith target node under trajectory P, attnode is a neural network, W is a mapping matrix, hi and hj are the original features of target node i and neighbor node j respectively, and || represents concatenation.

[0017] S22. According to the attention correlation coefficients of the neighbor nodes, the neighbor nodes are activated by the LeakyReLU function and normalized by the softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory. is the attention weight of the jth neighbor node of the i-th target node under trajectory P, is the set of all neighbor nodes of the i-th target node under trajectory P, is the attention correlation coefficient of the sth neighbor node of the i-th target node under trajectory P.

[0018] S23. According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory. The node fusion model is is the semantic representation of the i-th target node under trajectory P, σ(·) represents the activation function, is the set of neighbor nodes of the i-th target node under trajectory P.

[0019] In an optional embodiment, step S3 specifically includes:

[0020] S31. Perform a nonlinear transformation on the semantic representation and then perform a dot multiplication with the attention vector q to obtain the node importance of each semantic representation.

[0021] S32. Take the average of the node importances of all semantic representations under the same trajectory to obtain the trajectory importance S of each trajectory. Pi .in, S Piis the trajectory importance of the multi-semantic path P of node i, V is the set of all target nodes under trajectory P, q is the attention vector, q T is the transformation rank of q, W is the mapping matrix, is the semantic representation of node i under trajectory P.

[0022] S33, normalize the trajectory importance of each trajectory through the softmax function to obtain the attention weight β of each trajectory Pi ,in, m is the number of target nodes under trajectory P,

[0023] S34, weighted fusion is performed on the semantic representations of each target node under different trajectories according to the attention weights of each trajectory to generate the spatial nodes of each target node. The fusion model of weighted fusion is:

[0024] In an optional embodiment, step S1 specifically includes:

[0025] S11. Obtain medical data.

[0026] S12. Based on the time sequence and medical data, construct multiple medical insurance heterogeneous graphs.

[0027] In an optional embodiment, step S4 is specifically used to:

[0028] Based on the time sequence, the spatial nodes of the target nodes in multiple medical insurance heterogeneous graphs are obtained to obtain the time series of the spatial nodes of each target node.

[0029] In an optional embodiment, step S5 specifically includes:

[0030] S51. According to the time series, the final embedding of each target node is obtained through the bidirectional long short-term memory model. The final embedding is expressed as: T represents the set of all subgraphs, x t It represents the spatial node obtained by spatial aggregation of the target node in the t subgraph. Represents a splicing operation, (for the model forward propagation, Backward propagation for the model.

[0031] S52: Input the final embedding into a multi-layer perceptron for classification to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0032] In an optional embodiment, the entities of the medical insurance heterogeneous graph include patients, hospital departments, drugs, and consultation dates.

[0033] Second,

[0034] An embodiment of the present invention provides a device for identifying medical insurance fraud, which includes:

[0035] The graph construction module is used to obtain medical data and construct a medical insurance heterogeneous graph based on the medical data. The medical insurance heterogeneous graph uses entities as nodes and relationships as edges.

[0036] The semantic module is used to obtain the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories based on the multi-semantic meta-paths of different trajectories, and fuse the neighbor nodes of each target node under different trajectories with each target node respectively to obtain the semantic representation of each target node under different trajectories.

[0037] The spatial constraint module is used to fuse the semantic representations of each target node under different trajectories and generate the spatial nodes of each target node.

[0038] The time constraint module is used to obtain the time sequence of the spatial nodes of each target node based on the time sequence.

[0039] The judgment module is used to classify each target node according to the time series through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0040] In an optional embodiment, the semantic module is specifically used to:

[0041] Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the following steps are performed on each target node under different trajectories to obtain the semantic representation of each target node under different trajectories:

[0042] Get the attention correlation coefficient of the neighboring nodes of the target node in the current trajectory. Among them, is the attention correlation coefficient of the jth neighbor node of the ith target node under trajectory P, attnode is a neural network, W is a mapping matrix, hi and hj are the original features of target node i and neighbor node j respectively, and || represents concatenation.

[0043] According to the attention correlation coefficients of neighbor nodes, the neighbor nodes are activated by LeakyReLU function and normalized by softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory. is the attention weight of the jth neighbor node of the i-th target node under trajectory P, is the set of all neighbor nodes of the i-th target node under trajectory P, is the attention correlation coefficient of the sth neighbor node of the i-th target node under trajectory P.

[0044] According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory. The node fusion model is is the semantic representation of the i-th target node under trajectory P, σ(·) represents the activation function, is the set of neighbor nodes of the i-th target node under trajectory P.

[0045] Thirdly,

[0046] An embodiment of the present invention provides a device for identifying medical insurance fraud, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the method for identifying medical insurance fraud as described in any paragraph of the first aspect.

[0047] Fourthly,

[0048] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the medical insurance fraud identification method as described in any paragraph of the first aspect.

[0049] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0050] By constructing the medical insurance dataset into a medical insurance heterogeneous graph, a foundation is laid for the subsequent use of graph neural network technology to solve the problem of medical insurance fraud detection. Through multi-semantic meta-paths, groups with the same behavior trajectory are mined from the heterogeneous graph, and the heterogeneous graph is further converted into a homogeneous graph, which reduces the complexity of calculation and facilitates the aggregation operation of the subsequent spatial constraint module. The information of neighbor nodes and meta-paths is aggregated through a double-layer attention mechanism, and weights are assigned to different neighbor nodes and meta-paths, reducing the interference of noise nodes and low-correlation meta-paths. The patient's individual context information is combined through the time constraint module, and the aggregated patient node is finally embedded in the representation, so as to capture anomalies and determine whether the patient is a fraudster.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 It is a flowchart of a method for identifying medical insurance fraud provided by the first embodiment of the present invention.

[0054] Figure 2 Schematic diagram of the process of sampling patient groups based on multi-semantic metapaths.

[0055] Figure 3 It is a schematic diagram of the space constraint process.

[0056] Figure 4 It is a time constraint flow chart.

[0057] Figure 5 It is a logic block diagram of a medical insurance fraud identification method provided by the first embodiment of the present invention.

[0058] Figure 6 It is a structural schematic diagram of a medical insurance fraud identification device provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0061] It should be clear that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0063] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0064] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0065] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0066] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0067] Embodiment 1:

[0068] See also Figures 1 to 5 The first embodiment of the present invention provides a method for identifying medical insurance fraud, which can be performed by a device for identifying medical insurance fraud. In particular, the method is performed by one or more processors in the device for identifying medical insurance fraud to implement steps S1 to S5.

[0069] S1. Obtain medical data and construct a medical insurance heterogeneous graph based on the medical data. The medical insurance heterogeneous graph uses entities as nodes and relationships as edges.

[0070] Specifically, the medical insurance fraud identification device can be a portable notebook computer, a desktop computer, a server, a smart phone, a tablet computer, or other electronic device with computing capabilities, and the present invention is not limited to this.

[0071] By constructing a medical insurance heterogeneous graph, we can model the real medical environment and convert the problem of how to identify fraud in medical insurance into the problem of how to classify nodes in a heterogeneous graph, so as to better solve the technical problems to be solved, which has great practical significance.

[0072] like Figure 5As shown, based on the above embodiment, in an optional embodiment of the present invention, step S1 specifically includes steps S11 to S12.

[0073] S11. Obtain medical data.

[0074] S12. Based on the time sequence and medical data, construct multiple medical insurance heterogeneous graphs.

[0075] Specifically, in the medical insurance data set, there are millions of transaction records from a large number of users. In order to better understand the patient's behavior, we construct it into a medical insurance heterogeneous graph. Preferably, we extract all the medical records of the selected patient and construct four entities from them: patient, hospital department, medical date, and drug. In other embodiments, the medical insurance heterogeneous graph may also include other entities, such as doctors, etc., which is not specifically limited in the present invention.

[0076] In order to further refine the spatial regions, we treat hospitals and departments as a whole, which means that even departments with the same name in different hospitals will be treated as different entities. The date entity is refined to the day, and for the drug entity, in order to avoid the graph being too dense, we remove the entities with a unit price of less than 20 yuan.

[0077] S2. Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the neighbor nodes of each target node under different trajectories are respectively merged with each target node to obtain the semantic representation of each target node under different trajectories.

[0078] Specifically, by extracting the neighbor nodes of the target node from the medical insurance heterogeneous graph through multi-semantic meta-paths, the complex medical insurance heterogeneous graph can be converted into a patient isomorphic graph containing only the nodes we care about most, thereby simplifying the calculation while retaining the topological information of the original heterogeneous graph as much as possible. Because the members of each group have similar medical behavior trajectories, information can be better mined from the perspective of the entire group. To this end, we let the target node combine the characteristics of other neighbor nodes in the group (ie: aggregation / fusion). Through the aggregation operation, the target node can fuse the characteristics of the entire group, thereby obtaining the behavioral attributes in the spatial dimension.

[0079] like Figure 2 , Figure 3 and Figure 5 As shown, based on the above embodiment, in an optional embodiment of the present invention, step S2 specifically includes:

[0080] Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and steps S21 to S23 are performed on each target node under different trajectories respectively to obtain the semantic representation of each target node under different trajectories:

[0081] S21, obtain the attention correlation coefficient of the neighboring nodes of the target node in the current trajectory. is the attention correlation coefficient of the jth neighbor node of the ith target node under trajectory P, attnode is a neural network, W is a mapping matrix, hi and hj are the original features of target node i and neighbor node j respectively, and || represents concatenation.

[0082] S22. According to the attention correlation coefficients of the neighbor nodes, the neighbor nodes are activated by the LeakyReLU function and normalized by the softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory. is the attention weight of the jth neighbor node of the i-th target node under trajectory P, is the set of all neighbor nodes of the i-th target node under trajectory P, is the attention correlation coefficient of the sth neighbor node of the i-th target node under trajectory P.

[0083] S23. According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory. The node fusion model is is the semantic representation of the i-th target node under trajectory P, σ(·) represents the activation function, is the set of neighbor nodes of the i-th target node under trajectory P.

[0084] It should be noted that the process of integrating neighboring nodes into a target node is called spatial constraint. The entire spatial constraint process is shown in the attached figure. Figure 2 shown.

[0085] First, in a medical insurance heterogeneous graph, multiple multi-semantic meta-paths are used to sample neighbors, and then the node features within each group are aggregated to the target node. In the node-level aggregation process, we may encounter additional Figure 2 The situation shown in the third part of group in the figure. That is, in a group with the same behavior trajectory, the neighbors of fraudulent patients are not all fraudulent patients. It can only be said that the neighbors of fraudulent patients are most likely fraudulent patients, and the neighbors of ordinary patients are most likely ordinary patients. If directly aggregated, the node representation performance will be reduced due to noise. Therefore, we use the attention mechanism to assign different weights to different node neighbors. Finally, we aggregate the target node representation through semantic level aggregation.

[0086] Finally, since the target node no longer has neighbors with the same behavior trajectory after node-level aggregation, we group the target nodes of other behavior trajectories into groups for aggregation. In this way, the target node representation can contain information with different semantics. Similarly, we will also use the attention mechanism to measure the importance of different behaviors.

[0087] like Figure 2 , 5 As shown in the figure, the neighbor nodes of the target node are collected through the multi-semantic meta-path, specifically:

[0088] For the sake of convenience, we only show three entities: patients, hospital departments, and drugs in the heterogeneous graph. Figure 2 As shown in the first part of the Heterogeneous graph instance, patients P1, P2, and P3 were prescribed drug M at hospital H1, and P4 and P5 were also prescribed drug M, but at hospital H2.

[0089] The multi-semantic metapath PHMHP (patient-hospital department-drug-hospital department-patient) is as shown in the attached Figure 2 As shown in the second part of Mult-semantic metapath, using multi-semantic metapath to sample neighbors can be seen as a heterogeneous graph starting from the patient node, wandering between different types of nodes according to the order of the metapath, and finally returning to the patient node. For example: starting from the fraudster P2, passing through hospital H1, drug M and hospital H1 in turn, and finally returning to the fraudster P3. By analogy, we can get the following Figure 2 There are two groups shown in the third part group.

[0090] The semantic information of the multi-semantic metapath PHMHP itself can be understood as patients who are prescribed the same medicine in the same hospital department, and the patient groups finally obtained are all groups with the same behavioral attributes in the spatial dimension. In the medical insurance heterogeneous graph, due to the heterogeneity of the nodes, different types of nodes are in different feature spaces. Although we can use unique hot encoding to construct the features of date, medicine, and hospital department, we use the corresponding weight matrix in different node types to map them to a unified feature space. However, due to the characteristics of heterogeneous graphs, forcibly integrating different types of nodes together may cause "incompatibility" between the features. We construct the medical insurance heterogeneous graph in the hope of reflecting the behavioral attributes and learning useful information through the topological structure, while medical insurance fraud only cares about nodes of the patient type. Therefore, we only need to use the features of patients, and we do not need to use the features of other node types in the heterogeneous graph.

[0091] The use of multi-semantic meta-paths can transform a complex medical insurance heterogeneous graph into a patient homogeneous graph containing only the nodes we care about most, thereby simplifying the calculation while retaining the topological information of the original heterogeneous graph as much as possible. We can use multiple meta-paths at the same time, and the semantic information contained in each meta-path can be regarded as a behavioral attribute in a spatial dimension. In addition to PHMHP, there can also be PDMDP (patient-date-drug-date-patient), PHP (patient-hospital department-patient), PDP (patient-date-patient), etc. Among them, PDMDP represents patients who are prescribed the same drug on the same date, PHP represents patients who are treated in the same hospital department, and PMP represents patients who use the same drug. The present invention does not specifically limit the number and semantic information of meta-paths.

[0092] The reason for using multi-semantic meta-paths is that they contain richer information. For example, for PHMHP, PMP, and PHP, the relationship between departments and drugs added by PHMHP compared to single-semantic meta-paths is also very important. It can check the situation of random prescription of drugs by departments. Ultimately, using multiple meta-paths can capture groups with the same behavioral trajectory from multiple aspects.

[0093] like Figure 3 , 5 As shown in the figure, the specific process of fusing neighbor nodes is as follows:

[0094] After sampling the neighbors of the node through the multi-semantic meta-path, the new node representation of the patient is aggregated from these neighbor nodes. However, the contribution of each neighbor node is different and cannot be treated equally. For example, the proportion of fraudulent patients should be larger, while the proportion of general patients should be smaller.

[0095] Given a pair of nodes (i, j) sampled from a multi-semantic meta-path P, the attention correlation coefficient As shown in formula (1), it means the importance of node j to i. Where hi and hj represent the original features of node i and node j respectively, W∈R F′×F is the mapping matrix, attnode∈R 2F′ is a neural network that performs node-level attention, and || represents a concatenation operation. Since the weights between node pairs depend on their own characteristics, different importance can be assigned to different neighbors.

[0096]

[0097] In getting the mutual attention coefficient Finally, the LeakyReLU activation function is used, and then the softmax function is used to normalize all neighbors to obtain the normalized attention weights. It represents the weight of neighbor node j among all neighbors of node i. The specific details are shown in formula (2). Represents the set of all neighbor nodes of node i.

[0098]

[0099] Finally, node i is represented by the node-level embedding based on the multi-semantic meta-path P As shown in formula (3), where σ(·) represents the activation function. Since each node is aggregated by its neighbors, its embedding representation can well capture the behavior of the entire group.

[0100]

[0101] This attention mechanism can be extended to a multi-head mechanism, which helps the training process to be more stable. Specifically, we repeatedly perform K node-level aggregations and concatenate the learned embedding representations.

[0102]

[0103] After fusing neighbor nodes (node-level aggregation), different semantic representations of the target node are obtained. In order to obtain the final representation of the target node, different semantic representations need to be fused (semantic-level aggregation).

[0104] S3. Fuse the semantic representations of each target node under different trajectories to generate the spatial nodes of each target node.

[0105] Specifically, different multi-semantic meta-paths can represent different semantic information. Therefore, the present invention uses multiple multi-semantic meta-paths to mine more information from different angles in the heterogeneous graph.

[0106] like Figure 3 , 5 As shown, based on the above embodiment, in an optional embodiment of the present invention, step S3 specifically includes:

[0107] S31. Perform a nonlinear transformation on the semantic representation and then perform a dot multiplication with the attention vector q to obtain the node importance of each semantic representation.

[0108] S32. Take the average of the node importances of all semantic representations under the same trajectory to obtain the trajectory importance S of each trajectory. Pi .in, S Pi is the trajectory importance of the multi-semantic path P of node i, V is the set of all target nodes under trajectory P, q is the attention vector, q Tis the transformation rank of q, W is the mapping matrix, is the semantic representation of node i under trajectory P.

[0109] S33, normalize the trajectory importance of each trajectory through the softmax function to obtain the attention weight β of each trajectory Pi ,in, m is the number of target nodes under trajectory P,

[0110] S34, weighted fusion is performed on the semantic representations of each target node under different trajectories according to the attention weights of each trajectory to generate the spatial nodes of each target node. The fusion model of weighted fusion is:

[0111] It can be understood that in step S2, the target node fuses the features of neighboring nodes under different semantic meta-paths respectively, thereby obtaining the semantic representations of different semantic meta-paths. In the current step, it is necessary to fuse the semantic representations of the target node under different semantic meta-paths to reduce the number of semantic representations of the target node to one, so that the semantic representation of the target node contains information from various angles.

[0112] like Figure 3 , 5 As shown in Figure 2, the specific process of semantic level aggregation is as follows:

[0113] Suppose we now have m multi-semantic meta-paths {P0, P1, …, Pm}, after the previous node-level aggregation we can obtain m semantic-level embedding representations {H0, H1, …, Hm}.

[0114] Just like node neighbors, the importance of the semantic-level embedding representations of nodes obtained from different multi-semantic meta-paths also varies. In order to obtain a more comprehensive embedding representation of a node, on the one hand, we need to fuse the embedding representations at different semantic levels. On the other hand, we need to distinguish the importance of different multi-semantic meta-paths. Therefore, semantic-level aggregation will also introduce an attention mechanism, which automatically learns the weights of different multi-semantic meta-paths.

[0115] First, the node vectors after node-level aggregation are nonlinearly transformed. Then, the transformed vector is multiplied with the trainable semantic-level attention vector q to obtain the importance of the semantic-level embedding representation. Finally, the importance of all semantic-level embedding representations is averaged as the importance of each multi-semantic meta-path, and the importance of each meta-path is S Pi As shown in formula (5):

[0116]

[0117] After obtaining the importance of each meta-path, the attention weight β of each multi-semantic meta-path is obtained through normalization by the softmax function. Pi .

[0118]

[0119] After obtaining the attention weight of each multi-semantic meta-path, the attention weight of each multi-semantic meta-path is weighted and summed with the node-level embedding representation to obtain the semantic-level embedding representation. If the attention weight corresponding to the multi-semantic meta-path is larger, the final semantic-level embedding representation will be closer to it.

[0120]

[0121] Spatial constraints use behavioral trajectories to mine information from a group perspective. After node-level aggregation and semantic-level aggregation, the new node representation already contains various behavioral attributes. Next, we will use time constraints to consider the abnormalities on the entire medical timeline from an individual perspective.

[0122] S4. Based on the time sequence, obtain the time sequence of the spatial nodes of each target node. Specifically, step S4 is specifically used to: based on the time sequence, obtain the spatial nodes of the target nodes in the plurality of medical insurance heterogeneous graphs to obtain the time sequence of the spatial nodes of each target node.

[0123] Specifically, in step S1, the heterogeneous graph has been divided into T subgraphs in chronological order, and each subgraph G t All of them represent heterogeneous graphs constructed based on medical records within the corresponding time. In this way, when each subgraph is spatially constrained to obtain a new node representation, all target node representations are connected in series to obtain time series data.

[0124] S5. According to the time series, each target node is classified through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0125] Specifically, we use the Bi-LSTM algorithm to mine anomalies on the individual timeline. It can combine the previous and next relationships for bilateral reasoning, thereby capturing deep feature interactions and obtaining more powerful embedding expression capabilities.

[0126] like Figure 4 As shown, based on the above embodiment, in an optional embodiment of the present invention, step S5 specifically includes:

[0127] S51. According to the time series, the final embedding of each target node is obtained through the bidirectional long short-term memory model. The final embedding is expressed as: T represents the set of all subgraphs, x t It represents the spatial node obtained by spatial aggregation of the target node in the t subgraph. Represents a splicing operation, (for the model forward propagation, Backward propagation for the model.

[0128] S52: Input the final embedding into a multi-layer perceptron for classification to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0129] The entire time constraint process is as follows Figure 4 and Figure 5 As shown in Figure 1, we input the output of the target node in each subgraph into the Bi-LSTM model, concatenate the results obtained by forward propagation and backward propagation, and finally obtain the final embedding representation of the target node through the average pooling layer.

[0130] The final embedding representation x′ of the node is shown in Formula 8:

[0131]

[0132] Where T represents the set of all subgraphs, x t represents the embedding representation of the node in the t subgraph after spatial aggregation, represents the concatenation operation, and LSTM is shown in Formula 9:

[0133]

[0134] Specifically, step S1 constructs the medical insurance dataset into a medical insurance heterogeneous graph, which provides a basis for the subsequent use of graph neural network technology to solve the problem of medical insurance fraud detection. Step 2 uses multi-semantic meta-paths to solve how to mine groups with the same behavior trajectory from heterogeneous graphs, and further converts heterogeneous graphs into isomorphic graphs, reducing the complexity of calculations, and also facilitating the aggregation operation of subsequent spatial constraint modules. Steps S3 and S4 design a double-layer attention mechanism to aggregate the information of neighbor nodes and meta-paths, assign weights to different neighbor nodes and meta-paths, and reduce the interference of noise nodes and low-correlation meta-paths. Steps S4 and S5 design a time constraint module to capture anomalies in combination with the contextual information of individual patients, and finally embed the aggregated patient nodes into a representation for predicting whether the patient is a fraudster.

[0135] By constructing the medical insurance dataset into a medical insurance heterogeneous graph, a foundation is laid for the subsequent use of graph neural network technology to solve the problem of medical insurance fraud detection. Through multi-semantic meta-paths, groups with the same behavior trajectory are mined from the heterogeneous graph, and the heterogeneous graph is further converted into a homogeneous graph, which reduces the complexity of calculation and facilitates the aggregation operation of the subsequent spatial constraint module. The information of neighbor nodes and meta-paths is aggregated through a double-layer attention mechanism, and weights are assigned to different neighbor nodes and meta-paths, reducing the interference of noise nodes and low-correlation meta-paths. The patient's individual context information is combined through the time constraint module, and the aggregated patient node is finally embedded in the representation, so as to capture anomalies and determine whether the patient is a fraudster.

[0136] Embodiment 2:

[0137] See also Figure 6 The second embodiment of the present invention provides a medical insurance fraud identification device, which includes:

[0138] The graph construction module 1 is used to obtain medical data and construct a medical insurance heterogeneous graph based on the medical data. The medical insurance heterogeneous graph uses entities as nodes and relationships as edges.

[0139] Semantic module 2 is used to obtain the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories based on multi-semantic meta-paths of different trajectories, and merge the neighbor nodes of each target node under different trajectories with each target node respectively to obtain the semantic representation of each target node under different trajectories.

[0140] The spatial constraint module 3 is used to fuse the semantic representations of each target node under different trajectories to generate a spatial node of each target node.

[0141] The time constraint module 4 is used to obtain the time sequence of the spatial nodes of each target node based on the time sequence.

[0142] The judgment module 5 is used to classify each target node according to the time series through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0143] In an optional embodiment, the semantic module 2 is specifically used for:

[0144] Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and steps S21 to S23 are performed on each target node under different trajectories respectively to obtain the semantic representation of each target node under different trajectories:

[0145] S21, obtain the attention correlation coefficient of the neighboring nodes of the target node in the current trajectory. is the attention correlation coefficient of the jth neighbor node of the ith target node under trajectory P, attnode is a neural network, W is a mapping matrix, hi and hj are the original features of target node i and neighbor node j respectively, and || represents concatenation.

[0146] S22. According to the attention correlation coefficients of the neighbor nodes, the neighbor nodes are activated by the LeakyReLU function and normalized by the softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory. is the attention weight of the jth neighbor node of the i-th target node under trajectory P, is the set of all neighbor nodes of the i-th target node under trajectory P, is the attention correlation coefficient of the sth neighbor node of the i-th target node under trajectory P.

[0147] S23. According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory. The node fusion model is is the semantic representation of the i-th target node under trajectory P, σ(·) represents the activation function, is the set of neighbor nodes of the i-th target node under trajectory P.

[0148] In an optional embodiment, the space constraint module 3 specifically includes:

[0149] The node importance unit is used to perform nonlinear transformation on the semantic representation and then perform dot multiplication with the attention vector q to obtain the node importance of each semantic representation.

[0150] The trajectory importance unit is used to average the node importance of all semantic representations under the same trajectory to obtain the trajectory importance S of each trajectory. Pi .in, S Pi is the trajectory importance of the multi-semantic path P of node i, V is the set of all target nodes under trajectory P, q is the attention vector, q T is the transformation rank of q, W is the mapping matrix, is the semantic representation of node i under trajectory P.

[0151] The attention weight unit is used to normalize the trajectory importance of each trajectory through the softmax function to obtain the attention weight β of each trajectory. Pi ,in, m is the number of target nodes under trajectory P,

[0152] The spatial node unit is used to perform weighted fusion on the semantic representations of each target node under different trajectories according to the attention weights of each trajectory, and generate the spatial nodes of each target node. The fusion model of weighted fusion is

[0153] In an optional embodiment, the composition module 1 specifically includes:

[0154] The acquisition unit is used to acquire medical data.

[0155] The graph construction unit is used to construct multiple medical insurance heterogeneous graphs based on time sequence and medical data.

[0156] In an optional embodiment, the entities of the medical insurance heterogeneous graph include patients, hospital departments, drugs, and consultation dates.

[0157] In an optional embodiment, the time constraint module 4 is specifically used for:

[0158] Based on the time sequence, the spatial nodes of the target nodes in multiple medical insurance heterogeneous graphs are obtained to obtain the time series of the spatial nodes of each target node.

[0159] In an optional embodiment, the judging module 5 specifically includes:

[0160] The classification unit is used to obtain the final embedding of each target node based on the time series through the bidirectional long short-term memory model. The final embedding is expressed as: T represents the set of all subgraphs, x t It represents the spatial node obtained by spatial aggregation of the target node in the t subgraph. Represents a splicing operation, (for the model forward propagation, Backward propagation for the model.

[0161] The judgment unit inputs the final embedding into a multi-layer perceptron for classification to judge whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

[0162] Embodiment three:

[0163] The third embodiment of the present invention provides a device for identifying medical insurance fraud, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the method for identifying medical insurance fraud as described in any paragraph of the first embodiment.

[0164] Embodiment 4:

[0165] A fourth embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the medical insurance fraud identification method as described in any paragraph of the first embodiment.

[0166] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0167] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0168] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention is essentially or partly contributed to the prior art or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk. It should be noted that, in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such a process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0169] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying medical insurance fraud, characterized in that: Include: Acquire medical data, and construct a medical insurance heterogeneous graph based on the medical data; wherein the medical insurance heterogeneous graph has entities as nodes and relationships as edges; Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the neighbor nodes of each target node under different trajectories are respectively merged with each target node to obtain the semantic representation of each target node under different trajectories; Fusing the semantic representations of the target nodes under different trajectories to generate spatial nodes of the target nodes; Based on the time sequence, obtaining the time sequence of the spatial nodes of each target node; According to the time series, each target node is classified by a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud; Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the neighbor nodes of each target node under different trajectories are respectively fused with each target node to obtain the semantic representation of each target node under different trajectories. Specifically, Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the following steps are performed on each target node under different trajectories respectively to obtain the semantic representation of each target node under different trajectories: Get the attention correlation coefficient of the neighboring nodes of the target node under the current trajectory; where, , For the The target node is the first The attention correlation coefficient of neighbor nodes is For neural networks, is the mapping matrix, and The target nodes and neighbor nodes The original characteristics of Indicates splicing; According to the attention correlation coefficient of neighbor nodes, the neighbor nodes are activated by LeakyReLU function and normalized by softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory; , For the The target node is the first The attention weights of neighbor nodes, For the The target nodes are in the trajectory The set of all neighbor nodes under For the The target node is the first The attention correlation coefficient of neighbor nodes; According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory; wherein the node fusion model is , For the The target nodes are in the trajectory The semantic representation of represents the activation function, For the The target nodes are in the trajectory The set of neighbor nodes under ; The semantic representations of the target nodes under different trajectories are fused to generate spatial nodes of the target nodes, specifically including: The semantic representation is transformed nonlinearly and then combined with the attention vector Perform dot product to get the node importance of each semantic representation; The node importance of all semantic representations under the same trajectory is averaged to obtain the trajectory importance of each trajectory. ;in, , For Node Multi-semantic metapath The importance of trajectory, For the trajectory The set of all target nodes, is the attention vector, for The transfer rank, is the mapping matrix, For Node On track The semantic representation below; The trajectory importance of each trajectory is normalized by the softmax function to obtain the attention weight of each trajectory ,in, , is the number of target nodes under trajectory P, According to the attention weights of the respective trajectories, the semantic representations of the respective target nodes under different trajectories are weightedly fused to generate the spatial nodes of the respective target nodes; wherein the fusion model of the weighted fusion is: ; According to the time series, each target node is classified through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud, specifically including: According to the time series, the final embedding of each target node is obtained through a bidirectional long short-term memory model; wherein the final embedding is expressed as: , represents the set of all subgraphs, Indicates that the target node is The spatial nodes obtained by spatial aggregation of subgraphs, Represents a splicing operation, For the model forward propagation, Backward propagation for the model; The final embedding is input into a multi-layer perceptron for classification to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

2. The method for identifying medical insurance fraud according to claim 1, characterized in that: The obtaining of medical data and constructing a medical insurance heterogeneous graph according to the medical data specifically includes: Access to medical data; Based on the time sequence, construct multiple medical insurance heterogeneous graphs according to the medical data; Based on the time sequence, the time sequence of the spatial nodes of each target node is obtained, specifically including: Based on the time sequence, the spatial nodes of the target nodes in the multiple medical insurance heterogeneous graphs are obtained to obtain the time sequence of the spatial nodes of the target nodes.

3. The method for identifying medical insurance fraud according to any one of claims 1 to 2, characterized in that: The entities of the medical insurance heterogeneous graph include patients, hospital departments, drugs and consultation dates.

4. A device for identifying medical insurance fraud, characterized in that: Include: A graph construction module, used to obtain medical data and construct a medical insurance heterogeneous graph based on the medical data; wherein the medical insurance heterogeneous graph has entities as nodes and relationships as edges; A semantic module, for obtaining neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories based on multi-semantic meta-paths of different trajectories, and fusing the neighbor nodes of each target node under different trajectories with each target node, respectively, to obtain semantic representations of each target node under different trajectories; A spatial constraint module, used for fusing the semantic representations of the target nodes under different trajectories to generate spatial nodes of the target nodes; A time constraint module, used for obtaining the time sequence of the spatial nodes of each target node based on the time sequence; A judgment module, used to classify each target node according to the time series through a bidirectional long short-term memory model to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud; The semantic module is specifically used for: Based on the multi-semantic meta-paths of different trajectories, the neighbor nodes of each target node in the medical insurance heterogeneous graph under different trajectories are obtained, and the following steps are performed on each target node under different trajectories respectively to obtain the semantic representation of each target node under different trajectories: Get the attention correlation coefficient of the neighboring nodes of the target node under the current trajectory; where, , For the The target node is the first The attention correlation coefficient of neighbor nodes is For neural networks, is the mapping matrix, and The target nodes and neighbor nodes The original characteristics of Indicates splicing; According to the attention correlation coefficient of neighbor nodes, the neighbor nodes are activated by LeakyReLU function and normalized by softmax function to obtain the attention weights of the neighbor nodes of the target node under the current trajectory; , For the The target node is the first The attention weights of neighbor nodes, For the The target nodes are in the trajectory The set of all neighbor nodes under For the The target node is the first The attention correlation coefficient of neighbor nodes; According to the attention weight, all neighbor nodes of the target node in the current trajectory are fused to the target node to obtain the semantic representation of the target node in the current trajectory; wherein the node fusion model is , For the The target nodes are in the trajectory The semantic representation of represents the activation function, For the The target nodes are in the trajectory The set of neighbor nodes under ; The space constraint module specifically includes: The node importance unit is used to perform nonlinear transformation on the semantic representation and then add it to the attention vector Perform dot product to get the node importance of each semantic representation; The trajectory importance unit is used to average the node importance of all semantic representations under the same trajectory to obtain the trajectory importance of each trajectory. ;in, , For Node Multi-semantic metapath The importance of trajectory, For the trajectory The set of all target nodes, is the attention vector, for The transfer rank, is the mapping matrix, For Node On track The semantic representation below; The attention weight unit is used to normalize the trajectory importance of each trajectory through the softmax function to obtain the attention weight of each trajectory. ,in, , is the number of target nodes under trajectory P; The spatial node unit is used to perform weighted fusion on the semantic representations of the target nodes under different trajectories according to the attention weights of the trajectories to generate spatial nodes of the target nodes; wherein the fusion model of weighted fusion is ; The judgment module specifically includes: The classification unit is used to obtain the final embedding of each target node through a bidirectional long short-term memory model according to the time series; wherein the final embedding is expressed as: , represents the set of all subgraphs, Indicates that the target node is The spatial nodes obtained by spatial aggregation of subgraphs, Represents a splicing operation, For the model forward propagation, Backward propagation for the model; The judgment unit is used to input the final embedding into a multi-layer perceptron for classification to determine whether each target node in the medical insurance heterogeneous graph is medical insurance fraud.

5. A device for identifying medical insurance fraud, characterized in that: It comprises a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement the medical insurance fraud identification method as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the medical insurance fraud identification method as described in any one of claims 1 to 3.

Images