Methods, devices, equipment, and media for identifying medical insurance fraud based on medical treatment behavior.

Abstract

This invention provides a method, apparatus, device, and medium for identifying medical insurance fraud based on medical treatment behavior, relating to the field of medical big data technology. The medical insurance fraud identification method includes steps S1 to S5. S1: Acquire medical data and construct a medical heterogeneous graph based on the medical data. S2: Sample target patient nodes in the medical heterogeneous graph according to behavioral patterns to obtain behavioral heterogeneous graphs for each behavioral pattern of the target patient node. S3: Obtain a global topological embedding based on the medical heterogeneous graph. S4: Using a two-step attention mechanism, first aggregate the target patient nodes within each behavioral heterogeneous graph according to their behavioral patterns to obtain behavioral feature embeddings for each behavioral pattern of the target patient node. Then, fuse the behavioral feature embeddings of each behavioral pattern guided by the global topological embedding to obtain the final embedding of the target patient node. S5: Classify the target patient node based on the final embedding to determine whether it is a medical insurance fraudster.

Description

Technical Field

[0001] This invention relates to the field of medical big data technology, and more specifically, to a method, apparatus, device, and storage medium for identifying medical insurance fraud based on medical treatment behavior. Background Technology

[0002] Medical fraud not only causes huge economic losses, but it is also a major factor affecting the quality of medical services and threatening patients' health.

[0003] Traditional methods for detecting medical insurance fraud include: 1. Using knowledge engineering to analyze medical insurance fraud and build knowledge models; 2. Detection methods based on data mining analysis; 3. Using statistical methods to build fraud identification models. These methods require selecting appropriate fraud assessment indicators, but commonly used medical insurance fraud assessment indicators may raise privacy concerns. Furthermore, traditional detection methods do not comprehensively mine medical data, neglecting the behavioral information implicit in the data.

[0004] Furthermore, existing graph neural network models are not designed for specific problems and are not suitable for solving healthcare fraud.

[0005] In view of this, the applicant hereby submits this application after studying the existing technology. Summary of the Invention

[0006] The present invention provides a method, apparatus, device and storage medium for identifying medical insurance fraud based on medical treatment behavior, in order to improve at least one of the above-mentioned technical problems.

[0007] First aspect

[0008] This invention provides a method for identifying medical insurance fraud based on medical treatment behavior, which includes steps S1 to S5.

[0009] S1. Acquire medical data and construct a medical heterogeneity graph based on the data. The medical heterogeneity graph uses patients, hospital departments, drugs, and dates as nodes.

[0010] S2. Sample the target patient nodes in the medical heterogeneous graph according to the behavior patterns to obtain the behavior heterogeneous graph of each behavior pattern of the target patient node.

[0011] S3. Obtain the global topology embedding based on the medical heterogeneous graph.

[0012] S4. Using a two-step attention mechanism, firstly, aggregate the target patient nodes within each behavioral heterogeneous graph to obtain the behavioral feature embeddings of each behavioral pattern of the target patient node. Then, use global topological embedding to guide the fusion of the behavioral feature embeddings of each behavioral pattern to obtain the final embedding of the target patient node.

[0013] S5. Based on the final embedding, classify and determine whether the target patient node is a medical insurance fraudster.

[0014] Based on the above embodiments, in an optional embodiment of the present invention, step S1 specifically includes steps S11 to S13.

[0015] S11. Obtain medical data.

[0016] S12. Extract medical records based on medical data.

[0017] S13. Based on the medical records, construct a medical heterogeneous graph with patients, hospital departments, dates, and medications as entities. Hospitals and departments are treated as a single entity, while departments with the same name in different hospitals are treated as different entities. Date entities are refined to the day level. Medication entities do not include medications with a unit price lower than a preset value.

[0018] Based on the above embodiments, in an optional embodiment of the present invention, step S2 specifically includes steps S21 to S23.

[0019] S21. Obtain the one-hop metapath, two-hop metapath, and metagraph corresponding to different behavior patterns.

[0020] S22. Sample the target patient node in the medical heterogeneous graph according to the one-hop meta-path, two-hop meta-path and meta-graph respectively, and obtain the initial heterogeneous graph of each behavior pattern of the target patient node.

[0021] S23. Preprocess the feature vectors of patient nodes in each initial heterogeneous graph to obtain behavioral heterogeneous graphs of each behavioral pattern of the target patient node. The preprocessing model is as follows: In the formula, h is the projected feature vector of the patient node after preprocessing, W is the projected feature vector of the patient node before preprocessing, λ is the hyperparameter, d is the degree of the patient node in the initial heterogeneous graph, and n is the total number of patient nodes in the initial heterogeneous graph.

[0022] Based on the above embodiments, in an optional embodiment of the present invention, step S3 specifically includes:

[0023] Based on the medical heterogeneity graph, the global topological embedding H is obtained through a graph neural network (GNN). t Among them, global topology embedding H tEmbedding behavioral features of different behavioral patterns of the target patient node in dimensions. They have the same dimensions.

[0024] Based on the above embodiments, in an optional embodiment of the present invention, step S4 specifically includes steps S41 to S44.

[0025] S41. Perform steps S42 to S44 on the behavioral heterogeneity graph of each behavioral pattern of the target patient node to aggregate the behavioral patterns of the target patient node and obtain the behavioral feature embedding of each behavioral pattern of the target patient node.

[0026] S42. Obtain the importance of each neighbor node in the behavioral heterogeneity graph through the attention mechanism. Among them, importance The formula for calculation is: In the formula, att intra It is a deep neural network that performs the attention mechanism, h u It is the projected feature vector of the target patient node u, || is the vector concatenation operator, h v It is the projected feature vector of neighbor node v, ρ i It is the i-th behavior pattern.

[0027] S43. Based on importance The weight coefficients of each neighbor node are obtained by normalizing using the softmax function. The formula for calculating the weighting coefficient is as follows: In the formula, It represents the importance of neighbor node v to target patient node u in the i-th behavioral pattern.

[0028] S44. Based on the weighting coefficients By aggregating behavioral patterns within the target patient node using activation functions and multi-head attention mechanisms, behavioral feature embeddings for each behavioral pattern of the target patient node are obtained. The formula for calculating the behavioral feature embedding is as follows: In the formula, It is the embedding of behavioral features of the i-th behavioral pattern of the target patient node u. It is a multi-head attention mechanism, σ is the activation function, and v is the neighboring patient nodes of the target patient node in the behavioral heterogeneity graph. It is the set of neighboring nodes of the i-th behavioral pattern of the target patient node u. It is the weight coefficient of neighbor node v to target patient node u in the i-th behavior pattern, h u It is the projected feature vector of the target patient node u.

[0029] Based on the above embodiments, in an optional embodiment of the present invention, step S4 further includes steps S45 to S47.

[0030] S45. Transform the behavioral feature embeddings of each behavioral pattern using a weight matrix, and then average the transformed embeddings to obtain the importance of each behavioral pattern. The calculation model for the importance of each behavioral pattern is as follows: In the formula, Let m represent the importance of the i-th behavior pattern, m be the number of behavior patterns, and q be the number of behavior patterns. T It is a learnable bias vector, W ρ For the weight matrix, It is the embedding of the behavioral features of the i-th behavioral pattern of the target patient node u, b ρ It is a parameterized attention vector.

[0031] S46. Obtain the importance of the global topology based on the global topology embedding. t The calculation model for the importance of the global topology is: S t =q T tanh(W t ·H t +b t In the formula, q T It is a learnable bias vector, W t It is the weight matrix, H t It is global topology embedding, b t It is the bias vector.

[0032] Based on the importance of each behavioral pattern The importance of global topology S t The weight coefficients of each behavior pattern are obtained by normalizing using the softmax function. And the weight coefficient β of the global topology t .

[0033] S47. Based on the weighting coefficients of each behavioral pattern Feature embedding of various behavioral patterns The weighting coefficient β of the global topology t and global topology embedding H t Obtain the final embedding H of the target patient node. The final embedding is calculated as follows: In the formula, H represents the final embedding of the target patient node. It is the weight coefficient of the i-th behavior pattern. It is the behavioral feature embedding of the i-th behavioral pattern of the target patient node u, β t The weight coefficients of the global topology, H tIt is a global topology embedding.

[0034] Based on the above embodiments, in an optional embodiment of the present invention, step S5 specifically includes:

[0035] The final embedded input is classified using a multilayer perceptron to determine whether the target patient node is a healthcare fraudster.

[0036] The second aspect

[0037] This invention provides a medical insurance fraud identification device based on medical treatment behavior, which includes:

[0038] The medical heterogeneity graph construction module is used to acquire medical data and construct a medical heterogeneity graph based on that data. The medical heterogeneity graph uses patients, hospital departments, drugs, and dates as nodes.

[0039] The behavioral heterogeneity graph construction module is used to sample the target patient node in the medical heterogeneity graph according to the behavioral patterns, and obtain the behavioral heterogeneity graph of each behavioral pattern of the target patient node.

[0040] The global topology acquisition module is used to obtain the global topology embedding based on the medical heterogeneous graph.

[0041] The vector fusion module uses a two-step attention mechanism to first aggregate the behavioral patterns of the target patient nodes in each behavioral heterogeneous graph, obtaining the behavioral feature embeddings of each behavioral pattern of the target patient node. Then, it fuses the behavioral feature embeddings of each behavioral pattern guided by global topological embedding to obtain the final embedding of the target patient node.

[0042] The vector classification module is used to classify based on the final embedding and determine whether the target patient node is a medical insurance fraudster.

[0043] Based on the above embodiments, in an optional embodiment of the present invention, the medical heterogeneity graph construction module specifically includes:

[0044] The medical data acquisition unit is used to acquire medical data.

[0045] The medical record extraction unit is used to extract medical records based on medical data.

[0046] The medical heterogeneity graph construction unit is used to build a medical heterogeneity graph based on medical records, using patients, hospital departments, dates, and medications as entities. Hospitals and departments are treated as a single entity, while departments with the same name in different hospitals are treated as different entities. Date entities are refined to the day level. Medication entities do not include medications with a unit price lower than a preset value.

[0047] Based on the above embodiments, in an optional embodiment of the present invention, the behavioral heterogeneous graph construction module specifically includes:

[0048] The behavior pattern acquisition unit is used to acquire one-hop meta-paths, two-hop meta-paths, and meta-graphs corresponding to different behavior patterns.

[0049] The sampling unit is used to sample the target patient node in the medical heterogeneous graph according to the one-hop meta-path, the two-hop meta-path and the meta-graph respectively, and obtain the initial heterogeneous graph of each behavior pattern of the target patient node.

[0050] The preprocessing unit is used to preprocess the feature vectors of patient nodes in each initial heterogeneous graph to obtain behavioral heterogeneous graphs of each behavioral pattern of the target patient node. The preprocessing model is as follows: In the formula, b is the projected feature vector of the patient node after preprocessing, W is the projected feature vector of the patient node before preprocessing, λ is the hyperparameter, d is the degree of the patient node in the initial heterogeneous graph, and n is the total number of patient nodes in the initial heterogeneous graph.

[0051] Based on the above embodiments, in an optional embodiment of the present invention, the global topology acquisition module is specifically used for:

[0052] Based on the medical heterogeneity graph, the global topological embedding H is obtained through a graph neural network (GNN). t Among them, global topology embedding H t Embedding behavioral features of different behavioral patterns of the target patient node in dimensions. They have the same dimensions.

[0053] Based on the above embodiments, in an optional embodiment of the present invention, the vector fusion module specifically includes:

[0054] The loop unit is used to execute the following unit on the behavioral heterogeneity graph of each behavioral pattern of the target patient node, respectively, to perform intra-behavioral aggregation of the target patient node and obtain the behavioral feature embedding of each behavioral pattern of the target patient node.

[0055] The neighbor importance calculation unit is used to obtain the importance of each neighbor node in the behavioral heterogeneous graph through an attention mechanism. Among them, importance The formula for calculation is: In the formula, att intra It is a deep neural network that performs the attention mechanism, h u It is the projected feature vector of the target patient node u, || is the vector concatenation operator, h v It is the projected feature vector of neighbor node v, ρ i It is the i-th behavior pattern.

[0056] The neighbor weight coefficient calculation unit is used to calculate the weight coefficient based on importance. The weight coefficients of each neighbor node are obtained by normalizing using the softmax function. The formula for calculating the weighting coefficient is as follows: In the formula, It represents the importance of neighbor node v to target patient node u in the i-th behavioral pattern.

[0057] Aggregation unit, used to aggregate based on weight coefficients By aggregating behavioral patterns within the target patient node using activation functions and multi-head attention mechanisms, behavioral feature embeddings for each behavioral pattern of the target patient node are obtained. The formula for calculating the behavioral feature embedding is as follows: In the formula, It is the embedding of behavioral features of the i-th behavioral pattern of the target patient node u. It is a multi-head attention mechanism, σ is the activation function, and v is the neighboring patient nodes of the target patient node in the behavioral heterogeneity graph. It is the set of neighboring nodes of the i-th behavioral pattern of the target patient node u. It is the weight coefficient of neighbor node v to target patient node u in the i-th behavior pattern, h u It is the projected feature vector of the target patient node u.

[0058] Based on the above embodiments, in an optional embodiment of the present invention, the vector fusion module further includes:

[0059] The behavior pattern importance calculation unit is used to transform the behavior feature embeddings of each behavior pattern through a weight matrix, and then average the transformed embeddings to obtain the importance of each behavior pattern. The calculation model for the importance of each behavioral pattern is as follows: In the formula, Let m represent the importance of the i-th behavior pattern, m be the number of behavior patterns, and q be the number of behavior patterns. T It is a learnable bias vector, W ρ For the weight matrix, It is the embedding of the behavioral features of the i-th behavioral pattern of the target patient node u, b ρ It is a parameterized attention vector.

[0060] The global importance calculation unit is used to obtain the global topology importance S based on the global topology embedding. t The calculation model for the importance of the global topology is: S t =q T tanh(W t ·H t +b t In the formula, qT It is a learnable bias vector, W t It is the weight matrix, H t It is global topology embedding, b t It is the bias vector.

[0061] The behavior pattern weighting coefficient calculation unit is used to calculate the weighting coefficient of each behavior pattern based on its importance. The importance of global topology S t The weight coefficients of each behavior pattern are obtained by normalizing using the softmax function. And the weight coefficient β of the global topology t .

[0062] Feature fusion unit, used to determine the weight coefficients of each behavioral pattern. Feature embedding of various behavioral patterns The weighting coefficient β of the global topology t and global topology embedding H t Obtain the final embedding H of the target patient node. The final embedding is calculated as follows: In the formula, H represents the final embedding of the target patient node. It is the weight coefficient of the i-th behavior pattern. It is the behavioral feature embedding of the i-th behavioral pattern of the target patient node u, β t The weight coefficients of the global topology, H t It is a global topology embedding.

[0063] Based on the above embodiments, in an optional embodiment of the present invention, the vector classification module is specifically used for:

[0064] The final embedded input is classified using a multilayer perceptron to determine whether the target patient node is a healthcare fraudster.

[0065] Third aspect

[0066] This invention provides a medical insurance fraud detection device based on medical treatment behavior, comprising 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 detection method based on medical treatment behavior as described in any paragraph of the first aspect.

[0067] Fourth aspect

[0068] This invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the medical insurance fraud identification method based on medical treatment behavior as described in any paragraph of the first aspect.

[0069] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0070] The medical insurance fraud identification method based on medical treatment behavior in this invention can fully explore abnormal behavior patterns in medical data, greatly improving the accuracy and efficiency of medical insurance fraud identification. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a flowchart illustrating the methods for identifying medical insurance fraud.

[0073] Figure 2 This is a model structure diagram of a medical insurance fraud detection method.

[0074] Figure 3 This is a schematic diagram of medical heterogeneity and sampling strategies.

[0075] Figure 4 This is a schematic diagram of the medical insurance fraud detection device. Detailed Implementation

[0076] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0077] Example 1

[0078] Please see Figures 1 to 3 The first embodiment of the present invention provides a method for identifying medical insurance fraud based on medical treatment behavior, which can be executed by a medical insurance fraud identification device based on medical treatment behavior (hereinafter referred to as: identification device). In particular, it is executed by one or more processors in the identification device to implement steps S1 to S5.

[0079] S1. Acquire medical data and construct a medical heterogeneity graph based on the data. The medical heterogeneity graph uses patients, hospital departments, drugs, and dates as nodes.

[0080] Specifically, converting medical datasets into a heterogeneous medical graph can model the medical insurance fraud detection problem as a classification problem within the heterogeneous graph, providing a theoretical basis for subsequent steps to solve the medical insurance fraud detection problem, which has great practical significance.

[0081] Understandably, the identification device can be any electronic device with computing power, such as a laptop, desktop computer, server, smartphone, or tablet.

[0082] Based on the above embodiments, in an optional embodiment of the present invention, step S1 specifically includes steps S11 to S13.

[0083] S11. Obtain medical data.

[0084] S12. Extract medical records based on medical data.

[0085] S13. Based on the medical records, construct a medical heterogeneous graph with patients, hospital departments, dates, and medications as entities. Hospitals and departments are treated as a single entity, while departments with the same name in different hospitals are treated as different entities. Date entities are refined to the day level. Medication entities do not include medications with a unit price lower than a preset value.

[0086] Specifically, the medical insurance dataset contains millions of transaction records from a large number of users. To better understand patients' behavioral patterns, we construct a medical insurance heterogeneous graph, such as... Figure 3 As shown in section a on the left. We extracted all the medical records of the selected patients and constructed four entities from them: patient, hospital department, date, and medication. To further refine the spatial representation, we treated the hospital and department as a single entity, meaning that even departments with the same name in different hospitals would be treated as different entities. The date entity was refined to the day, while for the medication entity, to avoid an overly dense graph, we removed entities with a unit price of less than 20 yuan. Ultimately, we abstracted the patient's medical record on the heterogeneous graph as a patient obtaining a specific medication from a specific department at a specific hospital on a specific day.

[0087] In the problem of medical insurance fraud detection, a medical heterogeneous graph, G = {V, ε, X}, is modeled to represent different types of objects and their interactions in real-world medical scenarios. In this embodiment, we focus on detecting the patient node set U, which is a subset of the node set V, denoted here as... In the dataset, each patient u∈U has a label Yu∈{0,1} to indicate whether the patient is a healthcare fraudster. The ultimate goal is to predict whether patient u is a healthcare fraudster. t The probability of being a medical insurance fraudster in the test set.

[0088] S2. Sample the target patient nodes in the medical heterogeneous graph according to the behavior patterns to obtain the behavior heterogeneous graph of each behavior pattern of the target patient node.

[0089] Based on the above embodiments, in an optional embodiment of the present invention, step S2 specifically includes steps S21 to S23.

[0090] S21. Obtain the one-hop metapath, two-hop metapath, and metagraph corresponding to different behavior patterns.

[0091] Different behavioral patterns of medical fraudsters have different implications. In this embodiment, to uncover potential behavioral patterns in a medical heterogeneous graph, a multi-behavioral pattern-based decomposition is performed. Given a node u in the medical heterogeneous graph that exhibits multiple behavioral patterns, the node first needs to be sampled in different ways. The sampling strategy focuses more on patient nodes that contain multiple behavioral patterns, as these nodes often provide more useful information for fraud detection.

[0092] To deeply uncover the behavioral patterns of medical fraudsters, this invention employs three different sampling strategies: meta-path based on one-hop neighbors, meta-path based on two neighbors, and sampling based on the metagraph. Figure 3 The middle part b is shown.

[0093] S22. Sample the target patient node in the medical heterogeneous graph according to the one-hop meta-path, two-hop meta-path and meta-graph respectively, and obtain the initial heterogeneous graph of each behavior pattern of the target patient node.

[0094] In this embodiment, through multi-behavior pattern decomposition, we can obtain nodes with similar trajectories under different behavior patterns, such as... Figure 3 As shown in section c on the right. By employing three sampling strategies to uncover the implicit behavioral patterns of medical fraudsters, we can effectively address the problem of abnormal behavioral characteristics that are overlooked by some current medical insurance fraud detection methods.

[0095] Specifically, Definition 1: A metapath can be represented as (abbreviated as A1A2A3…) l+1 ) in the form of a path, where Describes objects A1 and A l+1 The composite relationship between them. Definition 2: A metagraph can be represented as... (abbreviated as A1A2A3…) l+1 ) in the form of a path, where Describes objects A1 and A l+1 The compound relationship between them is ||>2.

[0096] Based on definitions 1 and 2 above, the following examples illustrate the meaning of meta-paths and meta-graphs in medical insurance heterogeneous graphs. For example... Figure 3 As shown in section a on the left, a medical heterogeneous graph was constructed to model medical treatment scenarios where medical insurance fraud occurs. The medical insurance attribute heterogeneous network contains multiple types of objects (i.e., patients (P), hospital departments (D), drugs (M), and dates (T)), with rich attributes and relationships.

[0097] In a heterogeneous graph, two users can be connected through multiple meta-paths, such as Patient-Department-Patient (PDP), Patient-Medication-Patient (PMP), and Patient-Date-Patient (PTP). Different meta-paths always express different semantics. For example, the path PDP indicates that the two patients connected by this meta-path have seen a doctor in the same department. The path PTP, on the other hand, indicates that the two patients connected by this meta-path have had medical records on the same day. Two users can be connected through multiple multi-semantic paths.

[0098] like Figure 3 As shown in section b in the middle, the path PDTP indicates that the two patients connected by this path visited the same department on the same day. The path PDTMP connects to the target patient, indicating that they received the same medication from the same department on the same day.

[0099] The key to learning the representation of a target node in a graph network lies in how to accurately propagate and aggregate information from its neighbors. Based on Definitions 1 and 2 above, we select neighboring patient nodes with similar behavioral patterns. For example... Figure 3 As shown in section c on the right, the graph structure after sampling via a one-hop meta-path (PDP) is such that some nodes are retained while others are discarded. The graph structure differs depending on the sampling method used. Therefore, subsequent steps require learning all these different graph structure information to obtain a more accurate embedding representation.

[0100] S23. Preprocess the feature vectors of patient nodes in each initial heterogeneous graph to obtain behavioral heterogeneous graphs of each behavioral pattern of the target patient node. The preprocessing model is as follows:

[0101]

[0102] In the formula, h is the projected feature vector of the patient node after preprocessing, W is the projected feature vector of the patient node before preprocessing, λ is the hyperparameter, d is the degree of the patient node in the initial heterogeneous graph, and n is the total number of patient nodes in the initial heterogeneous graph.

[0103] Specifically, considering that fraudsters often have more medical records, this manifests as more densely connected nodes in a heterogeneous graph. We process the feature vectors of sampled patient nodes so that subsequent input graph network models can pay more attention to this type of node when learning node embeddings. After preprocessing, more densely connected nodes receive greater attention in the next model component.

[0104] S3. Obtain the global topology embedding based on the medical heterogeneous graph.

[0105] Specifically, to study the complex and diverse behavioral patterns of medical fraudsters, we decompose them into multiple behavioral patterns. Step S2 completes the fusion of behavioral patterns, allowing us to learn the semantic information implicit in each behavioral pattern.

[0106] While sampling using three strategies can identify groups with similar behavioral trajectories, it also results in the loss of some topological information about the graph. This missing topological structure is crucial for connecting different behavioral patterns. The multi-behavioral pattern decomposition step is essentially similar to local clustering, making it difficult to acquire global knowledge during behavioral pattern merging. Therefore, this embodiment proposes a topology-aware module to guide pattern merging, enabling a deeper exploration of the structural relationships between behavioral patterns.

[0107] Based on the above embodiments, in an optional embodiment of the present invention, step S3 specifically includes:

[0108] Based on the medical heterogeneity graph, the global topological embedding H is obtained through a graph neural network (GNN). t Among them, global topology embedding H t Embedding behavioral features of different behavioral patterns of the target patient node in dimensions. They have the same dimensions.

[0109] Specifically, such as Figure 2 As shown in part c, a graph neural network (GNN) is used to learn global topological structure information.

[0110] H t =GNN(A,x)

[0111] Among them, H t Embeddings learned from dimensions and individual behavioral patterns Dimensions remain consistent. In subsequent steps, the global topology will be embedded into H. t Embedding in a single behavior pattern They are input together into the inter-pattern merging module for fusion learning.

[0112] S4. Using a two-step attention mechanism, firstly, aggregate the target patient nodes within each behavioral heterogeneous graph to obtain the behavioral feature embeddings of each behavioral pattern of the target patient node. Then, use global topological embedding to guide the fusion of the behavioral feature embeddings of each behavioral pattern to obtain the final embedding of the target patient node.

[0113] In this embodiment, neighbor node information of patients is aggregated based on different behavioral patterns to integrate structural information from multiple aspects of the medical heterogeneous graph for better representation learning. Building upon this, we propose a topology-aware module that introduces global topological information and distinguishes differences in behavioral patterns through semantic attention, obtaining the optimal weighted combination of task-specific semantic node embeddings.

[0114] Observations of real-world medical insurance fraud cases and datasets reveal that fraudsters typically operate in teams, exhibiting similar or even identical behavioral patterns. They tend to work in groups, visiting the same hospital or leaving medical records within the same timeframe. Based on these characteristics, we construct... Figure 2 The network model shown.

[0115] Specifically, for patient behavior information and patient feature information mined from medical heterogeneous graphs using three sampling strategies, a two-step attention mechanism is employed. First, the behavior patterns within a pattern are aggregated, and then the behavior patterns between patterns are aggregated, thus obtaining the composite semantic information of the target patient node (i.e., the final embedding H). This can more accurately represent the target patient node, thereby providing a foundation for downstream task decomposition.

[0116] Based on the above embodiments, in an optional embodiment of the present invention, step S4 specifically includes steps S41 to S44.

[0117] S41. Perform steps S42 to S44 on the behavioral heterogeneity graph of each behavioral pattern of the target patient node to aggregate the behavioral patterns of the target patient node and obtain the behavioral feature embedding of each behavioral pattern of the target patient node.

[0118] Specifically, such as Figure 2 As shown in Part 1, step S4 first requires aggregating the behavior patterns within the behavior heterogeneous graph to obtain the behavior feature embeddings of each behavior pattern of the target patient node.

[0119] S42. Obtain the importance of each neighbor node in the behavioral heterogeneity graph through the attention mechanism.

[0120] Specifically, within a given behavioral pattern (i.e., different behavioral heterogeneous graphs), the degree of the same patient node varies because these nodes are of different types or have different local topologies. This means that each neighboring node plays a different role and has different importance in intra-pattern fusion. Therefore, it cannot be directly aggregated using a basic GNN.

[0121] In this embodiment, an attention mechanism is used to aggregate data during this process. Specifically, based on a certain behavioral pattern ρ... i For node pairs (u, v), the attention mechanism can learn the importance. This represents the contribution of node v to node u. The importance of node v. The formula for calculation is:

[0122]

[0123] In the formula, att intra It is a deep neural network that performs the attention mechanism, h u It is the projected feature vector of the target patient node u, || is the vector concatenation operator, h v It is the projected feature vector of neighbor node v, ρ i It is the i-th behavior pattern.

[0124] From the formula for calculating the importance of node pairs, the weights between node pairs depend entirely on their projection characteristics. Therefore, different neighboring nodes of node u can be assigned different contributions.

[0125] S43. Based on importance The weight coefficients of each neighbor node are obtained by normalizing using the softmax function.

[0126] Specifically, in obtaining the importance between node pairs Next, the softmax function is used to obtain the normalized weighting coefficients 'a', where 'a' represents the weight of neighbor node v among all the neighbors of node u. The formula for calculating the weighting coefficients is:

[0127]

[0128] In the formula, It represents the importance of neighbor node v to target patient node u in the i-th behavioral pattern.

[0129] S44. Based on the weighting coefficients By aggregating behavioral patterns within the target patient node using activation functions and multi-head attention mechanisms, behavioral feature embeddings for each behavioral pattern of the target patient node are obtained.

[0130] Specifically, based on the weighting coefficients obtained earlier... Aggregate node u within its behavioral pattern and output it using an activation function σ(·). Due to the weighting coefficients... It is generated specifically within a certain behavioral pattern, and we use this method to extract specific semantic information from different behavioral patterns. That is:

[0131]

[0132] The training process for heterogeneous graph data suffers from high variance. This embodiment of the invention employs a multi-head attention mechanism to address this issue. Specifically, k independent attention mechanisms are executed, and their outputs are concatenated. This method effectively stabilizes the entire training and learning process. Finally, the calculation formula for the behavioral feature embedding is:

[0133]

[0134] In the formula, It is the embedding of behavioral features of the i-th behavioral pattern of the target patient node u. It is a multi-head attention mechanism, σ is the activation function, and v is the neighboring patient nodes of the target patient node in the behavioral heterogeneity graph. It is the set of neighboring nodes of the i-th behavioral pattern of the target patient node u. It is the weight coefficient of neighbor node v to target patient node u in the i-th behavior pattern, h u It is the projected feature vector of the target patient node u.

[0135] In this embodiment, For our target patient node u in behavioral pattern ρ i The following is a summary. It shows one aspect of the semantic information contained in node u. Suppose we have m behavioral patterns ρ0, ρ1...ρ m After pattern fusion, we can obtain a sequence represented by m behavior pattern-specific vectors.

[0136] like Figure 2 As shown in Part 2, after completing the fusion within behavioral patterns, it is necessary to perform fusion between behavioral patterns to obtain a unique feature vector representing the target patient node. In this embodiment, the importance of different behavioral patterns is learned through multi-modal attention learning, and they are merged to generate the final multi-modal representation of the target patient node u.

[0137] Based on the above embodiments, in an optional embodiment of the present invention, step S4 further includes steps S45 to S48.

[0138] S45. Transform the behavioral feature embeddings of each behavioral pattern using a weight matrix, and then average the transformed embeddings to obtain the importance of each behavioral pattern.

[0139] Specifically, in order to learn the importance of different behavioral patterns, a weight matrix W is first used. ρ This involves transforming the specific representation for each behavioral pattern, converting the embedded representation of that pattern into a suitable dimension, preferably reducing the dimensionality. Understandably, excessively large dimensions lead to high computational complexity and may cause additional redundancy.

[0140] After transforming the dimensions, the importance of each behavioral pattern is summarized by averaging all transformed embedding representations.

[0141] The calculation model for the importance of each behavioral pattern is as follows:

[0142]

[0143] In the formula, Let m represent the importance of the i-th behavior pattern, m be the number of behavior patterns, and q be the number of behavior patterns. T It is a learnable bias vector, W ρ For the weight matrix, It is the embedding of the behavioral features of the i-th behavioral pattern of the target patient node u, b ρ It is a parameterized attention vector.

[0144] Specifically, after obtaining the embeddings of different patterns of the target node, a weight matrix is ​​used to transform them to a suitable dimension. Then, an average embedding is obtained by averaging all the embeddings. Finally, the similarity between the embedding of each pattern and this average embedding is evaluated as the importance of each behavioral pattern.

[0145] S46. Obtain the importance of the global topology based on the global topology embedding. t .

[0146] Specifically, this embodiment incorporates topology awareness to guide the merging process. Therefore, the importance of global topology information in the aggregation process is evaluated using the following formula: The calculation model for the importance of global topology is:

[0147] S t =q T tanh(W t ·H t +b t )

[0148] In the formula, q T It is a learnable bias vector, W tIt is the weight matrix, H t It is global topology embedding, b t It is the bias vector.

[0149] S47. Based on the importance of each behavioral pattern The importance of global topology S t The weight coefficients of each behavior pattern are obtained by normalizing using the softmax function. And the weight coefficient β of the global topology t .

[0150] Specifically, in this embodiment, the importance is normalized using the softmax function to obtain the contribution of each behavioral pattern and the global topology, i.e., the weight coefficient.

[0151]

[0152]

[0153] The weighting coefficients can be used to interpret each behavioral pattern ρ. i Its contribution to fraud detection is obvious. The higher the value, the better the behavioral pattern ρ. i The more important it is.

[0154] S48. Based on the weighting coefficients of each behavioral pattern Feature embedding of various behavioral patterns The weighting coefficient β of the global topology t and global topology embedding H t Obtain the final embedding H of the target patient node.

[0155] Specifically, after obtaining the importance and embedding representation of each behavioral pattern and the global topological embedding, all behavioral patterns are merged, and the global topological embedding is used to guide the fusion to obtain the final embedding H. The final embedding is calculated as follows:

[0156]

[0157] In the formula, H represents the final embedding of the target patient node. It is the weight coefficient of the i-th behavior pattern. It is the behavioral feature embedding of the i-th behavioral pattern of the target patient node u, β t The weight coefficients of the global topology, H t It is a global topology embedding.

[0158] In this embodiment, a two-layer attention mechanism is used to aggregate information from neighboring nodes, and a topology-aware module is introduced to guide the fusion of information between different behavioral patterns, thereby reducing the impact of noisy nodes on the final prediction task.

[0159] S5. Based on the final embedding, classify and determine whether the target patient node is a medical insurance fraudster.

[0160] Specifically, the learned final embedding is used to detect whether a target patient is a medical insurance fraudster. In this embodiment, a multilayer perceptron is used to classify the final embedding. In other embodiments, other existing classification models can be used to classify the embedding representation of the target patient node in the heterogeneous graph, thereby determining whether the target patient node is a medical insurance fraudster. This invention does not limit the specific type of classification model.

[0161] The medical insurance fraud detection method of this invention uses cross-entropy as the loss function during training and optimizes the model through backpropagation. Cross-entropy is expressed as follows:

[0162]

[0163] In the formula, Y l and H l It consists of the labels of the labeled patient nodes and the learned embeddings, where M is the parameter of the classifier.

[0164] exist Figure 2 In the graph neural network model shown, we use these frameworks to learn weights and aggregate information layer by layer to obtain meaningful node embeddings. Finally, the node embeddings are used to predict whether a user is a healthcare fraudster.

[0165] This invention's medical insurance fraud detection method based on medical treatment behavior transforms the medical insurance fraud detection problem into a node classification problem within a medical heterogeneous graph. Three sampling strategies are selected to mine abnormal behavior information of fraudsters in the medical heterogeneous graph. A two-step attention mechanism aggregates and learns the patient behavior information and patient feature information mined from the medical heterogeneous graph, and then graph representation learning techniques are used to learn low-dimensional embedding representations of nodes in the graph. Finally, the low-dimensional embedding representations are fed into a multilayer perceptron to predict whether a target user has engaged in medical insurance fraud. The attention mechanism can learn the preferences of different sampling methods to obtain a good target node embedding.

[0166] The medical insurance fraud identification method based on medical treatment behavior in this invention solves the problem that traditional methods are limited to focusing only on abnormal characteristics of users (such as abnormal medical treatment amounts). At the same time, it takes into account the behavioral pattern attributes in the medical process, which can fully explore abnormal behavioral patterns in medical data and greatly improve the accuracy and efficiency of medical insurance fraud identification.

[0167] Example 2

[0168] Please see Figure 4The second embodiment of the present invention provides a medical insurance fraud identification device based on medical treatment behavior, which includes:

[0169] The Medical Heterogeneous Graph Construction Module 1 is used to acquire medical data and construct a medical heterogeneous graph based on the data. The medical heterogeneous graph uses patients, hospital departments, drugs, and dates as nodes.

[0170] Behavioral heterogeneity graph construction module 2 is used to sample the target patient node in the medical heterogeneity graph according to the behavioral pattern, and obtain the behavioral heterogeneity graph of each behavioral pattern of the target patient node.

[0171] Global topology acquisition module 3 is used to obtain global topology embedding based on the medical heterogeneous graph.

[0172] Vector fusion module 4 uses a two-step attention mechanism to first aggregate the behavioral patterns of the target patient nodes in each behavioral heterogeneous graph, obtaining the behavioral feature embeddings of each behavioral pattern of the target patient node. Then, it fuses the behavioral feature embeddings of each behavioral pattern guided by global topological embedding to obtain the final embedding of the target patient node.

[0173] Vector classification module 5 is used to classify based on the final embedding and determine whether the target patient node is a medical insurance fraudster.

[0174] Based on the above embodiments, in an optional embodiment of the present invention, the medical heterogeneity graph construction module 1 specifically includes:

[0175] The medical data acquisition unit is used to acquire medical data.

[0176] The medical record extraction unit is used to extract medical records based on medical data.

[0177] The medical heterogeneity graph construction unit is used to build a medical heterogeneity graph based on medical records, using patients, hospital departments, dates, and medications as entities. Hospitals and departments are treated as a single entity, while departments with the same name in different hospitals are treated as different entities. Date entities are refined to the day level. Medication entities do not include medications with a unit price lower than a preset value.

[0178] Based on the above embodiments, in an optional embodiment of the present invention, the behavioral heterogeneous graph construction module 2 specifically includes:

[0179] The behavior pattern acquisition unit is used to acquire one-hop meta-paths, two-hop meta-paths, and meta-graphs corresponding to different behavior patterns.

[0180] The sampling unit is used to sample the target patient node in the medical heterogeneous graph according to the one-hop meta-path, the two-hop meta-path and the meta-graph respectively, and obtain the initial heterogeneous graph of each behavior pattern of the target patient node.

[0181] The preprocessing unit is used to preprocess the feature vectors of patient nodes in each initial heterogeneous graph to obtain behavioral heterogeneous graphs of each behavioral pattern of the target patient node. The preprocessing model is as follows: In the formula, h is the projected feature vector of the patient node after preprocessing, W is the projected feature vector of the patient node before preprocessing, λ is the hyperparameter, d is the degree of the patient node in the initial heterogeneous graph, and n is the total number of patient nodes in the initial heterogeneous graph.

[0182] Based on the above embodiments, in an optional embodiment of the present invention, the global topology acquisition module 3 is specifically used to: acquire the global topology embedding H based on the medical heterogeneity graph using a graph neural network (GNN). t Among them, global topology embedding H t Embedding behavioral features of different behavioral patterns of the target patient node in dimensions. They have the same dimensions.

[0183] Based on the above embodiments, in an optional embodiment of the present invention, the vector fusion module 4 specifically includes:

[0184] The loop unit is used to execute the following unit on the behavioral heterogeneity graph of each behavioral pattern of the target patient node, respectively, to perform intra-behavioral aggregation of the target patient node and obtain the behavioral feature embedding of each behavioral pattern of the target patient node.

[0185] The neighbor importance calculation unit is used to obtain the importance of each neighbor node in the behavioral heterogeneous graph through an attention mechanism. Among them, importance The formula for calculation is: In the formula, att intra It is a deep neural network that performs the attention mechanism, h u It is the projected feature vector of the target patient node u, || is the vector concatenation operator, h v It is the projected feature vector of neighbor node v, ρ i It is the i-th behavior pattern.

[0186] The neighbor weight coefficient calculation unit is used to calculate the weight coefficient based on importance. The weight coefficients of each neighbor node are obtained by normalizing using the softmax function. The formula for calculating the weighting coefficient is as follows: In the formula, It represents the importance of neighbor node v to target patient node u in the i-th behavioral pattern.

[0187] Aggregation unit, used to aggregate based on weight coefficients By aggregating behavioral patterns within the target patient node using activation functions and multi-head attention mechanisms, behavioral feature embeddings for each behavioral pattern of the target patient node are obtained. The formula for calculating the behavioral feature embedding is as follows: In the formula, It is the embedding of behavioral features of the i-th behavioral pattern of the target patient node u. It is a multi-head attention mechanism, σ is the activation function, and v is the neighboring patient nodes of the target patient node in the behavioral heterogeneity graph. It is the set of neighboring nodes of the i-th behavioral pattern of the target patient node u. It is the weight coefficient of neighbor node v to target patient node u in the i-th behavior pattern, h u It is the projected feature vector of the target patient node u.

[0188] Based on the above embodiments, in an optional embodiment of the present invention, the vector fusion module 4 further includes:

[0189] The behavior pattern importance calculation unit is used to transform the behavior feature embeddings of each behavior pattern through a weight matrix, and then average the transformed embeddings to obtain the importance of each behavior pattern. The calculation model for the importance of each behavioral pattern is as follows: In the formula, Let m represent the importance of the i-th behavior pattern, m be the number of behavior patterns, and q be the number of behavior patterns. T It is a learnable bias vector, W ρ For the weight matrix, It is the embedding of the behavioral features of the i-th behavioral pattern of the target patient node u, b ρ It is a parameterized attention vector.

[0190] The global importance calculation unit is used to obtain the global topology importance S based on the global topology embedding. t The calculation model for the importance of the global topology is: S t =q T tanh(W t ·H t +b t In the formula, q T It is a learnable bias vector, W t It is the weight matrix, H t It is global topology embedding, b t It is the bias vector.

[0191] The behavior pattern weighting coefficient calculation unit is used to calculate the weighting coefficient of each behavior pattern based on its importance. The importance of global topology S tThe weight coefficients of each behavior pattern are obtained by normalizing using the softmax function. And the weight coefficient β of the global topology t .

[0192] Feature fusion unit, used to determine the weight coefficients of each behavioral pattern. Feature embedding of various behavioral patterns The weighting coefficient β of the global topology t and global topology embedding H t Obtain the final embedding H of the target patient node. The final embedding is calculated as follows: In the formula, H represents the final embedding of the target patient node. It is the weight coefficient of the i-th behavior pattern. It is the behavioral feature embedding of the i-th behavioral pattern of the target patient node u, β t The weight coefficients of the global topology, H t It is a global topology embedding.

[0193] Based on the above embodiments, in an optional embodiment of the present invention, the vector classification module 5 is specifically used to: classify the final embedded input multilayer perceptron to determine whether the target patient node is a medical insurance fraudster.

[0194] Example 3

[0195] A third embodiment of the present invention provides a medical insurance fraud identification device based on medical treatment behavior, 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 medical insurance fraud identification method based on medical treatment behavior as described in any paragraph of Embodiment 1.

[0196] Example 4

[0197] This invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the medical insurance fraud identification method based on medical treatment behavior as described in any paragraph of Embodiment 1.

[0198] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0199] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0200] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, 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 a process, method, article, or apparatus. Without further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0201] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0202] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0203] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0204] The use of "first" and "second" in the embodiments is merely to distinguish similar objects and does not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "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.

[0205] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A medical insurance fraud identification method based on medical behavior, characterized in that, Include: Acquire medical data and construct a medical heterogeneity graph based on the medical data; wherein the medical heterogeneity graph uses patients, hospital departments, drugs, and dates as nodes; The target patient nodes in the medical heterogeneous graph are sampled according to the behavioral patterns to obtain the behavioral heterogeneous graph of each behavioral pattern of the target patient node. Based on the aforementioned medical heterogeneity graph, obtain the global topology embedding; Through a two-step attention mechanism, the target patient nodes in each behavioral heterogeneous graph are first aggregated within behavioral patterns to obtain the behavioral feature embeddings of each behavioral pattern of the target patient node; then, the behavioral feature embeddings of each behavioral pattern are fused with the global topology embedding to obtain the final embedding of the target patient node. Based on the final embedding, classification is performed to determine whether the target patient node is a medical insurance fraudster; For each target patient node in the behavioral heterogeneous graph, aggregation within behavioral patterns is performed to obtain the behavioral feature embeddings of each behavioral pattern of the target patient node, specifically including: performing the following steps on the behavior heterogeneous graph of each behavior pattern of the target patient node to obtain the behavior feature embedding of each behavior pattern of the target patient node : The importance of each neighbor node in the behavioral heterogeneity graph is obtained through an attention mechanism. Among them, importance The formula for calculation is: In the formula, It is a deep neural network that performs the attention mechanism. Target patient node The projected eigenvectors, || is the vector concatenation operator. Neighboring nodes Projected eigenvectors, It is the first A behavioral pattern; Based on the aforementioned importance The weight coefficients of each neighbor node are obtained by normalizing using the softmax function. The formula for calculating the weighting coefficient is as follows: In the formula, It is the first Neighbor nodes in each behavioral pattern For target patient nodes The importance of; According to the weighting coefficients By aggregating behavioral patterns within the target patient node through activation functions and multi-head attention mechanisms, behavioral feature embeddings of each behavioral pattern of the target patient node are obtained. The formula for calculating the behavioral feature embedding is as follows: In the formula, Target patient node The behavioral feature embedding of individual behavioral patterns It is a multi-head attention mechanism. It is an activation function. It is the neighboring patient nodes of the target patient node in the behavioral heterogeneity graph. Target patient node The The set of neighbor nodes for each behavioral pattern It is the first Neighbor nodes in each behavioral pattern For target patient nodes Weighting coefficients Target patient node ; The global topology embedding guides the fusion of behavioral feature embeddings of each behavioral pattern to obtain the final embedding of the target patient node, specifically including: The behavioral feature embeddings of each behavioral pattern are transformed using a weight matrix, and the importance of each behavioral pattern is obtained by averaging the transformed embeddings. The calculation model for the importance of each behavioral pattern is as follows: In the formula, Indicates the first The importance of individual behavioral patterns It is the number of behavioral patterns, It is a learnable bias vector. For the weight matrix, Target patient node The behavioral feature embedding of individual behavioral patterns It is a parameterized attention vector; The importance of the global topology is obtained based on the global topology embedding. The calculation model for the importance of the global topology is as follows: In the formula, It is a learnable bias vector. It is a weight matrix. It is global topology embedding, It is the bias vector; According to the importance of the respective behavior patterns and the global topology , the weight coefficients of the respective behavior patterns and the global topology are obtained by normalization through a softmax function Based on the weighting coefficients of each behavioral pattern Feature embedding of various behavioral patterns Weight coefficients of the global topology and the global topology embedding Obtain the final embedding of the target patient node. The final embedded calculation formula is: In the formula, H represents the final embedding of the target patient node. It is the first Weighting coefficients of each behavioral pattern Target patient node The behavioral feature embedding of individual behavioral patterns These are the weighting coefficients of the global topology. It is a global topology embedding.

2. The medical service behavior-based medical insurance fraud identification method according to claim 1, characterized in that, Based on behavioral patterns, target patient nodes in the medical heterogeneous graph are sampled to obtain behavioral heterogeneous graphs of various behavioral patterns of the target patient nodes, specifically including: Obtain the one-hop metapath, two-hop metapath, and metagraph corresponding to different behavior patterns; The target patient nodes in the medical heterogeneous graph are sampled according to the one-hop meta-path, two-hop meta-path and meta-graph respectively to obtain the initial heterogeneous graph of each behavior pattern of the target patient node. The feature vectors of patient nodes in each initial heterogeneous graph are preprocessed to obtain behavioral heterogeneous graphs of each behavioral pattern of the target patient node; wherein, the preprocessing model is... In the formula, For the preprocessed patient node's projected feature vector, For the projected feature vector of the patient node before preprocessing, For hyperparameters, The degree of the patient node in the initial heterogeneous graph. It is the total number of patient nodes in the initial heterogeneous graph.

3. The method of claim 1 or 2, wherein, Acquire medical data and construct a medical heterogeneity graph based on the medical data, specifically including: Obtaining medical data; Extract medical records based on the aforementioned medical data; Based on the medical records, the medical heterogeneous graph is constructed with the patient, hospital department, date, and medication as entities; wherein, the hospital and department are a whole, and the same department in different hospitals are different entities; the date entity is refined to the day; the medication entity does not include medications with a unit price less than a preset value.

4. The medical service behavior based medical insurance fraud identification method according to any one of claims 1 to 2, characterized in that, Based on the aforementioned medical heterogeneity graph, a global topology embedding is obtained, specifically including: Based on the aforementioned medical heterogeneity graph, a global topological embedding is obtained using a graph neural network (GNN). ; wherein, the global topology embedding Embedding behavioral features of different behavioral patterns of the target patient node in the dimensions of the target patient node The dimensions are the same; Based on the final embedding, classification is performed to determine whether the target patient node is a medical insurance fraudster, specifically including: The final embedded input multilayer perceptron is used for classification to determine whether the target patient node is a medical insurance fraudster. 5.A device for identifying health insurance fraud based on medical behavior, characterized by, A method for identifying medical insurance fraud based on medical treatment behavior, used to perform any one of claims 1 to 4; The medical insurance fraud detection device includes: A medical heterogeneity graph construction module is used to acquire medical data and construct a medical heterogeneity graph based on the medical data; wherein, the medical heterogeneity graph uses patients, hospital departments, drugs, and dates as nodes; The behavior heterogeneity graph construction module is used to sample the target patient node in the medical heterogeneity graph according to the behavior pattern, and obtain the behavior heterogeneity graph of each behavior pattern of the target patient node. A global topology acquisition module is used to acquire a global topology embedding based on the medical heterogeneous graph; The vector fusion module is used to first aggregate the target patient nodes in each behavioral heterogeneous graph through a two-step attention mechanism to obtain the behavioral feature embeddings of each behavioral pattern of the target patient node; then, the global topological embedding guides the fusion of the behavioral feature embeddings of each behavioral pattern to obtain the final embedding of the target patient node. The vector classification module is used to classify based on the final embedding and determine whether the target patient node is a medical insurance fraudster. 6.A medical insurance fraud identification device based on medical behavior, characterized by, It 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 medical insurance fraud identification method based on medical treatment behavior as described in any one of claims 1 to 4.

7. 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, it controls the device where the computer-readable storage medium is located to perform the medical insurance fraud identification method based on medical treatment behavior as described in any one of claims 1 to 4.

Images