Fusion knowledge heterogeneous graph neural network medical insurance fraud risk identification system
By constructing a heterogeneous graph neural network that integrates knowledge, and utilizing medical knowledge graphs and intra-batch negative sampling methods, the accuracy problem of medical insurance risk control systems in identifying new types of fraud was solved, achieving efficient identification and accurate prediction of new fraud patterns.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 上海金仕达卫宁软件科技有限公司
- Filing Date
- 2022-11-15
- Publication Date
- 2026-07-14
AI Technical Summary
Existing medical insurance risk control systems are unable to proactively detect new fraud methods. Traditional rule systems require a large number of experts to update them, and unsupervised models lose accuracy when there is an imbalance between positive and negative samples.
A heterogeneous graph neural network integrating knowledge is constructed. The graph neural network model learns the connection relationships between data nodes, utilizes rich features from medical knowledge graphs, and trains the model using an in-batch negative sampling method to improve the accuracy of fraud risk identification.
It achieves more efficient identification of medical insurance fraud risks, can discover new violation patterns, and improves the accuracy of the model under the condition of imbalance between positive and negative samples.
Smart Images
Figure CN115936893B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a data recognition technology, and more particularly to a medical insurance fraud risk recognition system that integrates knowledge through heterogeneous graph neural networks. Background Technology
[0002] Currently, rule-based systems based on medical insurance policies and clinical knowledge remain the mainstream approach to medical insurance risk control. Rule-based risk control engines can effectively filter predefined risk scenarios from data, building the first line of defense for the medical insurance fund. However, as risk control systems iterate, medical insurance fraud methods are constantly evolving. Rule-based systems struggle to proactively identify new fraud methods; therefore, after the initial system version is launched, a large number of experts are still needed to continuously update the rules to address newly exposed fraud scenarios.
[0003] With the rapid development of modern information technologies such as big data and artificial intelligence, traditional rule-based systems are struggling to meet all needs. Building a medical insurance risk control system centered on machine learning has become a major development direction. Deep learning models, combined with massive amounts of medical insurance settlement data, make it possible to efficiently and accurately identify new types of fraud. However, existing medical insurance data lacks positive and negative sample labels, making it impossible to directly locate violation characteristics through supervised learning methods. Therefore, much research focuses on unsupervised learning methods, primarily using clustering and outlier detection, to discover violation samples.
[0004] Traditional rule-based systems or unsupervised models mostly focus on a specific type of fraud. When new violation patterns emerge, they are difficult to detect. Summary of the Invention
[0005] To address the challenge of uncovering potential risks within medical insurance data, a heterogeneous graph neural network-based medical insurance fraud risk identification system integrating knowledge is proposed. By constructing a graph neural network and utilizing the connections between data nodes, multi-layer feature aggregation is performed to fully learn existing violation patterns, thereby discovering more potential fraud risks.
[0006] The technical solution of this invention is: a knowledge-integrated heterogeneous graph neural network medical insurance fraud risk identification system, comprising a relation-aware heterogeneous graph neural network for predicting the risk of each medical visit node, wherein the relation-aware heterogeneous graph neural network structure is as follows:
[0007] Patient visits generate various types of data, including visitation, diagnosis, medication, and institution. Diagnostic and medication data are constrained by knowledge graphs. Heterogeneous graphs are constructed using medical settlement data and medical knowledge graphs. A heterogeneous graph consists of a series of subgraphs, and each subgraph corresponds to a relationship.
[0008] On the constructed heterogeneous graph, a meta-path is used to randomly walk through the graph to train a downsampling skip character model to extract features from each node.
[0009] After connecting the meta-path features and the node's own attribute features, they are used as the node input features of the graph neural network model, and the medical visit node is used to construct the graph model.
[0010] A graph convolutional layer is used to learn a unique node representation in the relation graph for each heterogeneous graph. A cross-relation information passing module is used to interact across relation node representations. A semantic aggregation module is used to aggregate relation-aware node representations to obtain a representation with relation nodes, thus forming a relation-aware heterogeneous graph neural network.
[0011] Preferably, each subgraph corresponds to a relation, which is defined by a string triple, including the source node type, edge type, and target node type.
[0012] Preferably, the medical visit node graph is constructed as follows: with medical visit as the target node, a four-part graph is constructed with the patient, medical institution, time, and diagnosis as source nodes, and the edges correspond to the relationships of the patient's affiliation, the medical institution, the time period, and the diagnosis information, respectively.
[0013] Preferably, the diagnostic and drug data are constrained by a knowledge graph: the diagnostic and drug data are input into the knowledge graph in pairs. If a connection exists in the knowledge graph, the diagnosis is used as the source node, the drug is used as the target node, and the edge is a subgraph of the treatment drug used to diagnose the disease, which is then added to the heterogeneous graph.
[0014] Preferably, the graph neural network model for relation-aware features consists of four models: relation-defined graph convolution, weighted residual connections, cross-relation message passing, and relation representation learning, forming a relation-aware heterogeneous graph neural network layer. By stacking L layers, information from L-order neighborhoods is aggregated. Each layer provides a relation-aware node representation for the target node v, as well as a relation representation associated with the source node u. A semantic fusion module is used to aggregate the relation-aware node representations, and a node representation is obtained by weighting importance to serve as the node representation for downstream tasks.
[0015] A training method for a relation-aware feature graph neural network model for identifying medical insurance fraud risks is proposed. The method involves constructing a relation-aware feature graph neural network model, selecting historically detected irregular medical visits (detected using rule-based and unsupervised models) as positive samples, and the remainder as negative samples. Negative samples represent irregular medical visits with severe class imbalance. The positive and negative samples are then fed into the relation-aware feature graph neural network model for training. During training, an in-batch negative sampling method is used, where P represents positive samples, N represents negative samples, a represents the number of positive samples, and b represents the number of negative samples, with a << b. In each training iteration, n positive samples are sampled from a without replacement, and 2n negative samples are sampled from b with replacement, ensuring relative class balance while fully utilizing all data.
[0016] A knowledge-integrated method for identifying medical insurance fraud risk relationships involves constructing a graph neural network model with relationship-aware features, feeding unlabeled medical data into the graph neural network model for identification, and obtaining violation patterns.
[0017] The beneficial effects of this invention are as follows: the heterogeneous graph neural network medical insurance fraud risk identification system of this invention integrates knowledge, has a more complete graph construction method, establishes the connection between diagnosis and drugs through medical knowledge graph, and has richer relational connection information; the heterogeneous graph neural network model based on relation perception has better accuracy than the homogeneous graph model; the negative sampling method solves the problem of decreased model accuracy caused by extreme imbalance of positive and negative samples. Attached Figure Description
[0018] Figure 1 This is a schematic diagram illustrating the construction of a heterogeneous graph using medical settlement data and medical knowledge according to the present invention;
[0019] Figure 2 This is a schematic diagram illustrating node feature acquisition in a heterogeneous graph according to the present invention.
[0020] Figure 3 The graph convolution graph is defined by the relationship of the present invention. Detailed Implementation
[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0022] I. Graph Neural Network Framework
[0023] 1. Construct a heterogeneous graph that integrates knowledge:
[0024] Patient visits generate various types of data related to consultations, diagnoses, medications, and medical institutions. The relationships between diagnoses and medications are constrained by knowledge graphs. Heterogeneous graphs are constructed using medical settlement data and medical knowledge graphs, such as... Figure 1 As shown.
[0025] 2. Node feature representation:
[0026] (1) As Figure 2 The diagram shown illustrates the node feature acquisition process. On the constructed heterogeneous graph, a meta-path is used to randomly traverse and train a skipping word model to extract the features of each node.
[0027] (2) The meta-path features and the node’s own attribute features are connected and used as the node input features of the graph neural network model.
[0028] 3. Graph Neural Network Model Design:
[0029] The graph neural network model considering relation-aware features is a fine-grained level learning approach for node representations on heterogeneous graphs. Specifically, it first uses a graph convolutional layer to learn unique node representations for specific relationships within each heterogeneous graph. Then, a cross-relation information passing module is used to improve the interaction between cross-relation node representations. Furthermore, a semantic aggregation module is designed to aggregate the relation-aware node representations, resulting in a representation with relational nodes.
[0030] 4. Risk identification methods:
[0031] (1) Use the trained graph neural network model to predict the risk of each medical visit node.
[0032] (2) Patients with high risk scores will be reviewed by a business expert.
[0033] II. A Heterogeneous Graph Neural Network System for Medical Insurance Fraud Risk Identification Integrating Knowledge:
[0034] 1. Data preprocessing:
[0035] 1.1 Input data:
[0036] (1) Personnel information: age, gender, type of insurance
[0037] (2) Medical information: medical category, length of hospital stay, disease diagnosis, total medical expenses, amount paid by the pooled fund, amount paid by the individual account, amount paid in cash, disease diagnosis, and admission time.
[0038] (3) Medical institutions: Medical institution level, medical institution category
[0039] (4) Medical order information: Item name
[0040] (5) Violation label: Positive samples are violations of medical treatment detected in the past using rules and unsupervised models, and the rest are negative samples.
[0041] 1.2 Data Processing:
[0042] (1) Age: an integer between 0 and 110
[0043] (2) Gender: Male, Female
[0044] (3) Insurance categories: Urban employees and rural residents
[0045] (4) Medical institution level: No level, Level 1, Level 2, Level 3
[0046] (4) Medical categories: outpatient, inpatient, and medication purchase
[0047] (5) Admission / Discharge Time: Break it down into four dimensions: year, month, day, and hour.
[0048] (6) Project Name: Standardization Mapping to the Three Major Medical Insurance Catalogs
[0049] 2. Using medical records to create a diagram:
[0050] Medical insurance settlement data contains various types of nodes. When building models using graph neural networks, relationships between visit nodes are typically established using intermediate nodes, and finally, only isomorphic graphs of visit nodes are used for modeling. To address the issue of insufficient information, a heterogeneous graph structure containing multiple types of nodes is introduced, while medical knowledge is used to supplement the missing relationships between diagnoses and medications in the data.
[0051] A heterogeneous graph consists of a series of subgraphs, and each subgraph corresponds to a relation. Each relation is defined by a string triple (source node type, edge type, target node type).
[0052] In the settlement record data, with the visit as the target node, it is easy to construct a four-part seed graph: visit (target node) - patient (source node), with the edge indicating that the visit belongs to the patient; visit (target node) - medical institution (source node), with the edge indicating that the visit occurred at the medical institution; visit (target node) - time (source node), with the edge indicating that the visit occurred during the time period; and visit (target node) - diagnosis (source node), with the edge indicating the diagnosis information of the visit.
[0053] Diagnosis (source node) - Drug (target node), with edges representing the subgraphs of treatment drugs used to diagnose diseases: Input the diagnosis and drug into the knowledge graph in pairs, and establish the connection if a connection exists in the knowledge graph.
[0054] 2.1 Node Feature Representation Learning:
[0055] 2.1.1 Node attribute characteristics:
[0056] (1) Medical treatment nodes: Medical category and out-of-town medical treatment mark are given unique hot codes, and the number of days of hospitalization, total medical expenses, amount of payment from the pooled fund, amount of payment from the personal account and amount of cash payment are standardized;
[0057] (2) Patient node: gender and insurance type are uniquely coded, and age is standardized;
[0058] (3) Medical institutions: The level and category of medical institutions are uniquely coded.
[0059] 2.1.2 Graph structural features:
[0060] Based on multi-semantic meta-paths with different trajectories, node sequences are sampled. The resulting sequences are used to obtain features of consultation, patient, medical institution, treatment items, and time nodes using a word-skipping model. Among these, consultation, patient, medical institution, and attribute features are connected and used as input features for a graph neural network model.
[0061] 2.2 Graph Neural Network Model Considering Relationship-Aware Features:
[0062] Existing heterogeneous graph learning methods primarily rely on the propagation of node features, with limited research on how different relation types enhance fine-grained node features. To address this, we propose a relation-aware heterogeneous graph neural network that learns fine-grained node features by considering relation awareness. First, graph convolutions are used to learn unique node representations from each relation-specific graph. Then, a cross-relation message passing module improves the interaction between node representations across different relations. Furthermore, a hierarchical relation representation approach guides node feature learning. Finally, a semantic fusion module aggregates the relation-aware node representations into a compact feature representation.
[0063] 2.2.1 Relationally Constrained Graph Convolution:
[0064] 1) such as Figure 3 As shown, the target node, source node, and relation are projected into their latent space according to the transformation matrix of a specific relation:
[0065]
[0066]
[0067]
[0068] in It consists of the source node u, the target node v, and the relation ψ at layer l. (e) The trainable matrix. It consists of the source node u, the target node v, and the relation ψ at layer l. (e) The representation of layer 0 Features are derived using initialized node features, and the initial representation of the relation is a one-hot encoding of the relation. This yields the normalized importance of source node u to target node v.
[0069]
[0070]
[0071] Where [·,·] represents the concatenation operation, N ψ(e) (v) represents the set of neighboring nodes of v under a specific relation. Finally, the neighborhood information of v under the specific relation is represented as follows:
[0072]
[0073] 2.2.2 Weighted Residual Linkage:
[0074]
[0075] When aggregating neighborhood information through specific relationships, the importance of the target node's features needs to be considered; simple summation cannot distinguish the importance of the target node and its neighborhood information. Therefore, a trainable parameter is added. Adaptive aggregation of target features and neighborhood information.
[0076] 2.2.3 Cross-relationship message passing:
[0077] Since the source node and the target node are related, simple pooling operations can confuse the node representations of different relations. Therefore, it is necessary to establish different node representations. Let R(v) be the set of relations involved in node v. Then, cross-relation information transfer can be represented as:
[0078]
[0079] in The importance of relation r at level l is calculated using the following formula:
[0080]
[0081] in Let r be the trainable attention vector of the l-th layer relation ψ(e) to control the information flow between ψ(e) and relation combination R(v). r and r' both belong to R(v), and the role of equation (9) is to map the probability between [0,1].
[0082] 2.2.4 Learning Relational Representations:
[0083] The propagation mechanism of relation representation:
[0084]
[0085] 2.2.5 Relationship-aware representation fusion:
[0086] A relation-aware heterogeneous graph neural network layer is constructed using four models: relation-constrained graph convolution, weighted residual connections, cross-relation message passing, and relation representation learning. By stacking L layers, information from the L-order neighborhood can be aggregated. These L layers can provide a relation-aware node representation for the target node v. and the relationship representation associated with the source node u Aggregation of relation-aware node representations is performed through a semantic fusion module:
[0087]
[0088]
[0089] Where γ u,v The learned relation r represents the node representation h. v Importance. V r E r Node representation Relational representation The transformation matrix is then used to obtain the node representation h through importance weighting. v This serves as a node representation for downstream tasks.
[0090] 2.2.6 Optimization Objective:
[0091] Loss = ylogP + (1-y)(1-P)
[0092] Where y is the label of the node, which comes from traditional rules and unsupervised scenario models, and P is the probability predicted by the model.
[0093] 3. Model pre-training:
[0094] Positive samples are those of irregular medical visits detected in the past using rules and unsupervised models, while the rest are negative samples, which show serious class imbalance.
[0095] Downsampling is a common method to solve class imbalance. Traditional downsampling discards some negative samples before the model starts training, which may result in some important samples not being trained by the model.
[0096] This method employs intra-batch negative sampling, first dividing the samples into P... a N b Let P be the number of positive samples, N be the number of negative samples, a be the number of positive samples, and b be the number of negative samples, where a << b. In each training iteration, n positive samples are sampled from a without replacement, and 2n negative samples are sampled from b with replacement. This method ensures both relative class balance and full utilization of all data.
[0097] 4. Model Effects:
[0098] New violation patterns were discovered: On the test set, the model's predictions were compared with the labels of traditional rule-based and unsupervised scenario models, revealing some new violation samples. These were subsequently identified as novel violation patterns by business experts.
[0099] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A heterogeneous graph neural network-based medical insurance fraud risk identification system integrating knowledge, characterized in that, This includes a heterogeneous graph neural network that senses relationships and predicts the risk at each medical visit node. The relation-aware heterogeneous graph neural network structure: Patient visits generate various types of data, including visit data, diagnosis data, medication data, and institution data. The diagnostic and medication data are constrained by a knowledge graph. A heterogeneous graph is constructed using medical settlement data and the medical knowledge graph. A heterogeneous graph consists of a series of subgraphs, and each subgraph corresponds to a relation. The relation is defined by a string triple, including the source node type, edge type, and target node type. With medical visit as the target node, a four-part subgraph is constructed, with patients, medical institutions, time, and diagnoses as source nodes and edges corresponding to the relationships of the patients, medical institutions, time periods, and diagnoses, respectively. Diagnoses and medications are paired and input into the knowledge graph. If a connection exists in the knowledge graph, the diagnosis is used as the source node, the medication as the target node, and edges are added to the heterogeneous graph as subgraphs of the treatment medications used to diagnose the disease. On the constructed heterogeneous graph, a random walk is performed using meta-paths to train a sampling skipping word model to extract features from each node; After connecting the meta-path features and the node's own attribute features, they are used as the node input features of the graph neural network model, and the medical visit node is used to construct the graph model. The relation-aware graph neural network model consists of four models: relation-defined graph convolution, weighted residual connections, cross-relation message passing, and relation representation learning. This forms a relation-aware heterogeneous graph neural network layer. By stacking L layers, information from L-order neighborhoods is aggregated, with each layer providing the target node. v Relationship-aware node representation, and its relationship with the source node. u The relation representation is aggregated through a semantic fusion module to obtain a relation-aware node representation. The node representation is obtained by weighting the importance and used as the node representation for downstream tasks, thus forming a relation-aware heterogeneous graph neural network.
2. A training method for a graph neural network model of relationship-aware features for identifying medical insurance fraud risks, characterized in that, A graph neural network model for relationship-aware features as described in claim 1 is constructed. Historically detected violations of medical records using rule-based and unsupervised models are selected as positive samples, while the remainder are negative samples. Negative samples represent violations with severe class imbalance. These positive and negative samples are then fed into the graph neural network model for training, employing an in-batch negative sampling method during training. P As a positive sample, N Let be the number of negative samples, 'a' be the number of positive samples, and 'b' be the number of negative samples. In each training cycle, n positive samples are sampled from a without replacement, and 2n negative samples are sampled from b with replacement, ensuring a relatively balanced class distribution while making full use of all the data.
3. A knowledge-integrated method for identifying medical insurance fraud risk relationships, characterized in that, A graph neural network model for relationship-aware features as described in claim 1 is constructed. Unlabeled medical data is fed into the graph neural network model for identification to obtain violation patterns.