Insurance claim processing method and apparatus, electronic device, and computer program product

Abstract

The application provides a claim settlement processing method and device, electronic equipment and computer program product, and belongs to the field of data processing, to solve the problems of data processing isolation, static model rigidity, long-range dependence loss and insufficient interpretability of existing methods. The method comprises the following steps: preprocessing initial data associated with a claim request to obtain preprocessed data; generating a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on entities in the preprocessed data, and the edges between the nodes are obtained based on business logic associated with the claim request; based on the dynamic heterogeneous network, path exploration is performed through a meta-path constrained walk strategy; and performing target risk determination on the claim request based on the results of the path exploration.

Description

Technical Field

[0001] This application relates to the field of data processing, and more particularly to a method, apparatus, electronic device, and computer program product for processing insurance claims. Background Technology

[0002] Current insurance claims fraud prevention primarily relies on rule engines, traditional machine learning models, static graph analysis, and natural language processing. Rule engines, based on preset fixed thresholds and logic, lack flexibility and struggle to address constantly evolving fraud methods. Traditional machine learning models rely on local features in historical data, failing to effectively capture long-term dependencies and dynamic evolution patterns across time and institutions in medical practices. While static graph analysis can construct relationships between entities, it cannot reflect the dynamic changes of networks over time and lacks the ability to proactively detect new and covert fraud patterns. Natural language processing methods mainly focus on contradiction detection at the text level, making it difficult to deeply integrate and model multi-source, heterogeneous structured data. Therefore, existing solutions generally suffer from isolated data processing, static and rigid models, lack of long-range dependencies, and insufficient interpretability, resulting in a need to improve their accuracy, timeliness, and business adaptability when facing complex and ever-changing health insurance medical claims fraud scenarios. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and computer program product for processing insurance claims, which can solve the problems of isolated data processing, static and rigid models, lack of long-range dependencies, and insufficient interpretability in existing methods.

[0004] In a first aspect, embodiments of this application provide a method for processing insurance claims, which includes the following steps: The initial data associated with the claim request is preprocessed to obtain preprocessed data; A dynamic heterogeneous network is generated based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request; Based on the dynamic heterogeneous network, path exploration is performed using a meta-path constraint traversal strategy. Based on the results of the path exploration, the target risk of the claim request is determined.

[0005] Secondly, embodiments of this application provide an insurance claims processing apparatus, which includes the following: The data preprocessing module is used to preprocess the initial data associated with the claim request to obtain preprocessed data; A network construction module is used to generate a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request; The path exploration module is used to explore paths based on the dynamic heterogeneous network using a meta-path constraint traversal strategy; and, The risk assessment module is used to assess the target risk of the claim request based on the results of the path exploration.

[0006] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the insurance claim processing method as described in the first aspect.

[0007] Fourthly, embodiments of this application provide a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, the program instructions being executed by a computer to implement the steps of the insurance claim processing method as described in the first aspect.

[0008] This application's embodiments construct a dynamic heterogeneous network, unifying the representation of various dispersed and heterogeneous entities and their complex business relationships related to claims within a structured dynamic heterogeneous network. This breaks through the data silo limitations of existing methods, providing a solid data foundation for subsequent analysis. The introduction of a meta-path constraint traversal strategy for path exploration not only allows for targeted investigation using predefined risk patterns but also proactively searches for and discovers potential abnormal correlations and high-risk patterns within the dynamic heterogeneous network. This adapts to the evolution of fraud methods, proactively capturing novel or hidden risk paths that are difficult to detect using static rules or fixed models, thereby enhancing the model's generalization ability and proactive risk discovery capabilities. Risk judgment based on path exploration results does not rely on a single feature or black-box model output, improving the accuracy of the judgment results. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating an insurance claim processing method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating another insurance claim processing method provided in the embodiments of this application; Figure 3 This is a flowchart illustrating another insurance claim processing method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an insurance claims processing device provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating the specific process of a health insurance medical claims processing method provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] Current anti-fraud technology solutions in the insurance claims field mainly include the following: First, rule-driven engines are a widely used basic approach. They automatically perform preliminary screening of claims cases by pre-setting a series of fixed thresholds and logical conditions (such as the upper limit of a single claim amount, drug catalog restrictions, etc.), and mark suspicious cases that violate the rules and transfer them to manual review.

[0014] Secondly, traditional machine learning models utilize accumulated historical claims data to train classification models such as logistic regression and random forests. By learning the feature patterns in historical cases, they predict the probability of fraud in newly submitted cases, thereby assisting in risk decision-making.

[0015] In addition, static graph analysis technology attempts to take a relational approach, constructing a static knowledge graph of various entities (e.g., patients, hospitals, doctors, diagnoses, etc.) and their relationships in the claims chain. Then, it uses graph queries or simple graph algorithms to detect abnormal association patterns, such as suspicious behavior like the same doctor prescribing high-priced drugs to multiple patients in a short period of time.

[0016] Finally, Natural Language Processing (NLP) technology mainly focuses on unstructured text data. By performing keyword matching, contradiction detection, or descriptive analysis on texts such as medical records and diagnostic reports, it identifies potential risk signals at the text level, such as contradictory statements, unreasonable medical records, or descriptions of overtreatment.

[0017] These methods generally suffer from common problems such as insufficient adaptability to dynamic changes, limited ability to uncover complex relationships, and weak interpretability. Specifically: 1. Data silos and heterogeneity challenges: Medical data comes from scattered sources (e.g., hospitals, medical insurance bureaus, pharmacies, etc.) and has diverse structures (e.g., structured electronic medical records, unstructured image reports, time-series medication records, etc.), making it difficult to model uniformly using traditional methods.

[0018] 2. The contradiction between static rules and dynamic fraud: fraud methods iterate rapidly (e.g., falsifying chronic disease history, splitting hospitalizations, etc.), fixed rules are easily bypassed, and the model's generalization ability is insufficient.

[0019] 3. Lack of long-term dependence: Medical behavior is highly time-dependent (e.g., long treatment cycles for chronic diseases), seasonally fluctuates (e.g., peak flu season), and has evolutionary trends (e.g., the emergence of new fraud methods). Existing models focus more on the local characteristics of a single claim event and ignore long-term behavioral patterns across time and institutions (e.g., frequently changing hospitals to evade regulation).

[0020] 4. Poor interpretability: Although black-box models can output probability values, they cannot provide an intuitive chain of evidence to support underwriters' decisions.

[0021] Therefore, there is an urgent need for an insurance claims processing method, especially one that can efficiently integrate multi-source heterogeneous medical data to build a dynamically evolving global view, provide adaptive mechanisms to enable models to proactively discover new fraud patterns, and generate interpretable fraud evidence chains to meet regulatory and business needs.

[0022] This application utilizes a two-way collaborative mechanism of dynamic heterogeneous network-driven walking strategy and walking feedback to optimize network weights, thereby enabling the spatiotemporal correlation modeling of medical claims data and the proactive evolution of fraud patterns, in order to construct a closed-loop anti-fraud system from multi-source data fusion to intelligent judgment.

[0023] The following is in conjunction with the appendix Figures 1 to 6 This application provides a detailed description of an insurance claims processing method, apparatus, electronic device, and computer program product through specific embodiments and application scenarios.

[0024] Figure 1 This application illustrates an embodiment of an insurance claim processing method, which can be executed by an electronic device, including servers and / or terminal devices. In other words, the method can be executed by software or hardware installed on the server and / or terminal device, and includes the following steps: Step 110: Preprocess the initial data associated with the claim request to obtain preprocessed data.

[0025] This step involves data access and preprocessing upon receiving a claim request. The input is initial data from a multi-source heterogeneous system, and the output can be preprocessed data that has undergone cleaning, transformation, and standardization to provide high-quality data input for subsequent construction of dynamic heterogeneous networks.

[0026] The initial data can mainly include the following categories: core claims information, such as claims application forms, detailed expense lists, settlement documents, etc.; medical behavior data, such as electronic medical records (EMR), diagnostic codes (e.g., ICD-10), laboratory test reports, medical orders and prescription records, etc.; information of participating parties, such as patient basic information, attending physician and affiliated medical institution information, pharmacy information, etc.; spatiotemporal and cost data, such as consultation time, location latitude and longitude, various cost amounts, etc.; other related data, such as historical claims records, insurance product terms, medical policy catalogs, etc.

[0027] These initial data vary significantly in source, format, and quality, exhibiting multi-source heterogeneity. After preprocessing, they yield preprocessed data with a unified structure and controllable quality. Each entity in this dataset (e.g., a patient, a medical visit, a drug) has been transformed into a data object with a set of standardized feature vectors, while also clearly defining the relationships between entities (e.g., patient A received a diagnosis from doctor D at hospital H at time T). This dataset is used to construct a dynamic heterogeneous network, ensuring the accuracy and efficiency of network construction.

[0028] Step 120: Generate a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request.

[0029] The dynamic heterogeneous network transforms pre-processed, interconnected claims-related data into a structured network model, providing both a static snapshot of the data and the ability to evolve and adapt to new data over time. Nodes in a dynamic heterogeneous network can be heterogeneous, representing entities of different types and with different sets of attributes. These nodes originate directly from key business entities in the pre-processed data. Compared to homogeneous networks (where all nodes are of the same type), heterogeneous networks can more accurately model the complexities of the real world. For example, patients, doctors, and drugs are fundamentally different entities; classifying them into different node types allows subsequent analysis algorithms to distinguish and process the relationships and characteristics between different types of entities.

[0030] The edges between nodes represent specific, business-meaning relationships between them. Edge construction can be based on business logic and clinical pathways in the medical and insurance claims fields. For example, the edge "patient-visit-doctor" and the edge "diagnosis-matching-medication," although both are connections, represent different business meanings. Edge construction based on domain knowledge ensures that the network structure accurately reflects actual business interactions and helps in identifying abnormal patterns.

[0031] Dynamic heterogeneous networks are dynamic; their structure (including nodes and edges) grows and changes as new claims data is added, and the strength of edge relationships (usually quantified as weights) decays over time or is updated based on the latest interactions. For example, the weight of a "patient-doctor" edge gradually decreases as the patient stops seeing that doctor, reflecting changes in real-world doctor-patient relationships and making the network sensitive to new, progressively evolving fraud patterns.

[0032] Specifically, the construction process of dynamic heterogeneous networks may include the following: Node mapping and typification: Traverse the preprocessed data, map each unique business entity instance to a network node, and assign a node type label to it according to its entity type; the node type can include patient nodes (e.g., identified by anonymized customer ID), medical service provider nodes (e.g., doctors, medical institutions, etc.), medical item nodes (e.g., diagnosis, drugs, treatment items, etc.), and other auxiliary nodes (e.g., consultation time slices, geographical regions, etc.).

[0033] Edge relationship establishment based on business logic: Based on the business events and relationships recorded in the preprocessed data, directed or undirected edges are established between the corresponding nodes, and a relationship type is specified for each edge; the edge relationship type can include the primary diagnosis relationship (e.g., patient--[visited at]--doctor, doctor--[belongs to]--hospital, etc.), diagnosis association relationship (e.g., diagnosis A--[may be concurrent]--diagnosis B), medication / treatment relationship (e.g., diagnosis--[use of medicine]--medicine, diagnosis--[corresponding treatment]--treatment item, etc.), spatiotemporal proximity relationship (e.g., patient--[registered on the same day]--department A and patient--[registered on the same day]--department B).

[0034] Edge attribute initialization and dynamic weighting mechanism: Each edge is assigned an initial weight upon creation and its creation or last enhancement timestamp is recorded; the dynamic weighting mechanism can be implemented through a built-in time decay function (such as an exponential decay function); the current weight of an edge decays according to the difference between its timestamp and the current time, and the decay coefficient can be configured according to the edge type (e.g., spatiotemporal proximity relationships decay quickly, while primary relationships decay slowly) to match the natural change rate of different business relationships.

[0035] Step 130: Based on the dynamic heterogeneous network, path exploration is performed using a meta-path constraint walking strategy.

[0036] Metapaths are sequences of alternating node and edge types defined on a network schema, used to abstractly describe a class of semantically specific association paths. For example, the metapath "patient-(visit) doctor-(prescribe) medication" describes "a patient obtaining a certain medication through a certain doctor." Metapaths provide semantic guidance and structural constraints for roaming, limiting it to business-meaningful paths and avoiding ineffective random wandering in a large, heterogeneous network, thus improving the efficiency and focus of exploration. Different metapaths can correspond to different risk pattern assumptions (e.g., fraudulent prescriptions, impersonation, etc.).

[0037] The walking strategy determines how to select the next node from among the many neighboring nodes of the current node at each step, while satisfying the meta-path constraints. Specifically, based on an adaptive strategy, the selection probability can be determined by two parts: network weight guidance and reinforcement learning optimization.

[0038] Network weight guidance, based on the edge weights, reflects the recent strength or suspiciousness of the relationship. Walks tend to select edges with higher weights, thus naturally focusing on recently active or historically questionable associations.

[0039] Reinforcement learning optimization involves treating a complete walk as an interaction between the agent and the network environment. After the walk, a reward signal is given based on whether the agent successfully terminates in a high-risk region (e.g., near a labeled fraudulent subgraph). Through reinforcement learning algorithms such as policy gradient, the walk strategy is continuously fine-tuned, making it increasingly adept at discovering new and hidden risky paths, allowing the walk strategy to evolve as fraud methods change.

[0040] The path exploration process, based on the aforementioned constraints and strategies, involves a series of walks within a dynamic heterogeneous network, generating a large number of entity-level path instances that conform to specific semantics. These concrete path instances are the embodiment of meta-paths in real-world data. By analyzing the statistical characteristics of these paths (e.g., frequency of occurrence, node repetition, path length distribution, etc.) and the attributes of the path endpoints, abnormal and clustered behavioral patterns can be identified. For example, a large number of instances following the "patient-doctor A-high-priced drug X" path concentrated on a few doctor nodes can form a quantifiable high-risk signal.

[0041] By performing multiple walks, a set of paths is generated. The nodes and edges covered by these paths on the network are analyzed to identify densely connected regions frequently traversed by multiple high-weight paths; these regions are considered candidates for "high-risk subgraphs." Furthermore, the high-risk subgraphs discovered by the walks are compared with subsequent judgment results (step 140) or feedback from manual review. If the walk path effectively hits the risk, a positive reward is given to the policy network to reinforce the path exploration behavior; otherwise, a negative reward is given. This method continuously optimizes the walk strategy.

[0042] Step 140: Based on the results of the path exploration, determine the target risk of the claim request.

[0043] The input for target risk assessment comes directly from high-risk paths and subgraphs discovered through path exploration, rather than raw data features or single model scores. The assessment is based on real-world, business-interpretable abnormal association patterns in the network (e.g., a closed-loop path where a patient obtains the same controlled drug through multiple doctors), which can significantly improve the interpretability of the assessment and the ability to detect complex association fraud.

[0044] Specifically, the process of determining the target risk of a claim request may include the following: Feature extraction and representation: From the high-risk path set and subgraphs output in step 130, multi-dimensional feature vectors are extracted and used as input to the decision model. These features include graph structure features, semantic features, statistical features, and contextual features.

[0045] Multi-task decision model inference: The above feature vectors are input into a pre-trained multi-task decision model. Its main task includes risk level determination, outputting the overall risk probability or level label of the claim request (e.g., probability value 0.92, or label "high risk"). Its auxiliary task includes risk type identification, outputting a probability distribution of a risk type (e.g., falsely prescribed drugs: 0.85, hospitalization without a designated bed: 0.10, others: 0.05). During model training, historical data with risk level and risk type labels are used for supervised learning, so that the feature representation shared by the two tasks can simultaneously serve the identification of macro risks and micro causes.

[0046] In this embodiment, a dynamic heterogeneous network is constructed to unify the representation of various scattered and heterogeneous entities and their complex business relationships related to claims within a structured dynamic heterogeneous network. This breaks the data silo limitations of existing methods and provides a solid data foundation for subsequent analysis. The introduction of a meta-path constraint traversal strategy for path exploration not only allows for targeted investigation using predefined risk patterns but also proactively searches for and discovers potential abnormal correlations and high-risk patterns within the dynamic heterogeneous network. This adapts to the evolution of fraud methods and proactively captures novel or hidden risk paths that are difficult to detect using static rules or fixed models, thereby improving the model's generalization ability and proactive risk discovery capabilities. Risk judgment based on path exploration results does not rely on a single feature or black-box model output, thus improving the accuracy of the judgment results.

[0047] Figure 2 This illustration shows a flowchart of another insurance claim processing method provided by an embodiment of this application. This method can be executed by an electronic device, which may include a server and / or terminal devices, etc. In other words, the method can be executed by software or hardware installed on the server and / or terminal devices, and includes the following steps: Step 210: Based on step 110 of the above embodiment, the initial data associated with the claim request is preprocessed to obtain preprocessed data. The method of this embodiment may further include the following specific steps: The initial data is subjected to at least one of the following processes: de-identification, missing value imputation, and standardization.

[0048] De-identification involves using technical means to remove or replace information in the data that can directly or indirectly identify an individual. For example, sensitive fields such as names, ID card numbers, social security card numbers, detailed addresses, and mobile phone numbers can be hashed, generalized (e.g., retaining the geographical region but hiding the specific address), or completely deleted. Through de-identification, data can be used for analysis and modeling without infringing on personal privacy, meeting relevant legal and regulatory requirements, and ensuring that subsequent analysis focuses on behavioral patterns rather than individual identity.

[0049] Missing value handling involves detecting missing fields in the data and selecting appropriate strategies for imputation based on the field's business meaning and data type. Statistical imputation includes using the mean or median for numerical indicators (e.g., examination indicator values) and the mode for categorical fields (e.g., gender, department). Context-based imputation can be performed by inferring imputation using other relevant fields in the same medical record or other medical records of the same patient. For example, inferring the missing dosage unit based on the current diagnosis and medication information. Labeling and preservation involve specially labeling critical fields that cannot be reliably imputed, allowing subsequent models to treat them as special states. Missing value handling aims to address data incompleteness issues and prevent model training bias or failure due to missing data. A reasonable imputation strategy can maximize the preservation of data information, thereby improving the robustness of subsequent networks and models.

[0050] Standardization processes can include transforming data from different sources and formats into a unified, machine-readable numerical representation. Structured data standardization includes normalizing or standardizing numerical data to eliminate the influence of units, and performing one-hot encoding or embedding encoding on categorical data. Text data vectorization involves processing textual information such as chief complaints, diagnostic descriptions, and pathology reports. Specific steps typically include word segmentation, stop word removal, stemming / lexical reconstruction, and generating text vectors using Term Frequency-Inverse Document Frequency (TF-IDF) or word embedding models (e.g., Word2Vec, BERT). Temporal and spatiotemporal data encoding includes converting consultation times into periodic features (e.g., day of the week, holiday status) and geographical coordinates into regional codes or distance features to other locations. Standardization maps heterogeneous raw data to a unified mathematical space, enabling information from completely different types of sources, such as medical record texts, diagnostic codes, and cost figures, to be understood and computed equally and effectively by subsequent networks or machine learning models, thus facilitating multimodal information fusion.

[0051] Step 220: Generate a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request.

[0052] Step 230: Based on the dynamic heterogeneous network, path exploration is performed using a meta-path constraint walking strategy.

[0053] Step 240: Based on the results of the path exploration, determine the target risk of the claim request.

[0054] Steps 220-240 can be found above. Figure 1The specific descriptions of steps 120-140 in the illustrated embodiment are provided, and the same technical effects can be achieved. To avoid repetition, they will not be repeated here.

[0055] In this embodiment, de-identification removes sensitive personal information from the outset of data utilization, ensuring that all subsequent processing steps comply with privacy protection requirements. This allows for the full extraction of data value within a legal and compliant framework, avoiding the risk of privacy leaks. Missing value handling effectively addresses the common problems of incomplete and inconsistent data. It not only prevents model training bias or operational interruptions caused by missing data but also, through reasonable imputation and labeling, preserves the information content of the original data to the greatest extent possible, helping to improve the quality of input data. Standardization mapping transforms raw data from diverse sources and with different formats into a unified mathematical vector space, solving the data silo problem. This enables subsequent networks and machine learning models to understand and compute all types of features in a consistent manner, providing a foundation for deep multimodal information fusion and joint analysis.

[0056] In some embodiments, taking health insurance medical claims as an example, in the data access and preprocessing steps, multi-source claim request data such as hospital data, medical insurance settlement data, pharmacy purchase records, and physical examination reports can be integrated and preprocessed for de-identification, missing value imputation, and standardization. Specifically, as follows: (1) De-identification: Remove or obscure information that can directly or indirectly identify an individual, such as name, ID number, mobile phone number, social security card number, address and other personal privacy information, to ensure that the individual is not identifiable.

[0057] (2) Fill missing values: Select the appropriate statistical measure, mean / median / mode, according to business logic to fill missing items in the physical examination report and incomplete fields in the medical record. For example, fill the laboratory indicator with the mean and fill the gender field with the mode.

[0058] (3) Standardization: For textual data such as chief complaint descriptions and pathology reports, data standardization is carried out by combining stop word filtering, stemming and TF-IDF vectorization methods to highlight keywords with distinctiveness, retain key features and provide a high-quality data foundation for model input.

[0059] In yet another exemplary embodiment, based on step 120 of the above embodiment, a dynamic heterogeneous network is generated according to the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on entities in the preprocessed data, and the edges between the nodes are obtained based on business logic associated with the claim request. The method of this embodiment may further include the following specific steps: The newly added entity is embedded as a new node into the dynamic heterogeneous network, and the feature vectors of the affected nodes in the dynamic heterogeneous network are incrementally updated.

[0060] The newly added entities are identified as new nodes in the network when the system receives claims data containing new entities (e.g., first-time patients, newly signed doctors, newly introduced drugs, etc.). An initial feature vector representation is generated for each of these entities. Specifically, the initial feature vector is generated based on the entity's own attribute information (metadata). For example, for a new doctor node, it can be encoded based on its department, title, and institution. To avoid treating the new node as an isolated point, its association with existing nodes is utilized to aggregate the pre-trained feature information of its neighboring nodes (e.g., the first patient node seen by the new doctor, the hospital node to which it belongs, etc.). The initial features are then fused with the aggregated neighborhood features. Dimensionality reduction and normalization operations using fully connected layers can be used to generate the final embedded representation of the new node, enabling it to participate in subsequent graph computations.

[0061] When embedding new data or nodes, not only does the network size change, but the context of existing nodes may also be affected. Therefore, the feature vectors of nodes (and their neighbors within a certain order) affected by the establishment of new edges or changes in the weights of old edges are recalculated and updated. The system automatically detects local changes in the network structure caused by data updates (e.g., adding an edge) and determines the set of nodes that need updating (usually the two nodes connected by the edge and their neighbors). Preferably, an incremental learning strategy is used instead of retraining the entire network; feature propagation and aggregation calculations are performed only on the affected set of nodes for a limited number of rounds, thus efficiently updating the feature representations of these nodes. Versioning of the updated node features ensures that downstream tasks read the latest node state, avoiding subsequent judgment errors caused by feature asynchrony.

[0062] In this embodiment, the dynamic heterogeneous network can continuously and in real-time update newly generated entities and relationships in the business, ensuring that the knowledge graph on which risk identification is based is always synchronized with the actual business state. This overcomes the defect of static network models becoming rapidly invalid due to data expiration and enables timely capture of new risks based on the latest behavioral patterns. By incrementally updating affected nodes, the impact of new information can be reasonably propagated to relevant localities, allowing node features to dynamically reflect their latest structural roles and semantic context. This ensures that the feature representations of all nodes in the network are in the same temporal and contextual context, providing a consistent feature foundation for subsequent walks and judgment tasks, thereby improving the overall stability and reliability of the model. Furthermore, through localized embedding and incremental updates, network expansion and iteration are achieved with lower resource consumption, contributing to engineering feasibility and practicality.

[0063] Figure 3This illustration shows a flowchart of another insurance claim processing method provided by an embodiment of this application. The method can be executed by an electronic device, which may include a server and / or terminal devices. In other words, the method can be executed by software or hardware installed on the server and / or terminal devices, and includes the following steps: Step 310: Preprocess the initial data associated with the claim request to obtain preprocessed data.

[0064] Step 310 can be found above. Figure 1 The specific description of step 130 in the illustrated embodiment, which achieves the same technical effect, will not be repeated here to avoid repetition.

[0065] Step 320: Based on step 120 of the above embodiment, a dynamic heterogeneous network is generated according to the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request. The method of this embodiment may further include the following specific steps: The weights of the edges are dynamically adjusted based on a time decay function, which uses exponential decay and different decay coefficients are configured for different types of edges.

[0066] The time decay function associates a timestamp with each edge in the network, recording the time when it was last created or strengthened through business activities (e.g., a doctor's visit, a prescription). Periodically or under specific triggering conditions, the current weight of each edge is recalculated based on its age (i.e., the difference between the current time and the edge's timestamp) using the preset time decay function. Time decay can simulate the natural weakening of business relationship strength over time in a real-world environment, allowing the network to focus on recently active relationships.

[0067] Specifically, an exponential decay function is adopted. Exponential decay has the advantages of being mathematically simple to calculate, having a smooth decay, and being easy to configure. It can ensure that the weight continues to decrease over time, with a faster decay in the early stage and a gradual flattening in the later stage. This is consistent with the understanding that in most business relationships, recent behavior has the greatest impact, while the impact of history gradually fades.

[0068] Different types of edges are configured with different attenuation coefficients, including configuring different attenuation coefficients according to the business semantic type of the edge. For example, the primary care relationship edge (patient-doctor) is configured with a smaller attenuation coefficient because the doctor-patient relationship has a certain stability, and its historical records are still valuable for judging the rationality of the current visit. The spatiotemporal proximity edge (patient-multiple visits in a short period of time) is configured with a larger attenuation coefficient because abnormal clustering behavior in the short term is an important real-time risk signal, but its indicative significance decreases rapidly over time.

[0069] Step 330: Based on the dynamic heterogeneous network, path exploration is performed using a meta-path constraint walking strategy.

[0070] Step 340: Based on the results of the path exploration, determine the target risk of the claim request.

[0071] Steps 330-340 can be found above. Figure 1 The specific descriptions of steps 130-140 in the illustrated embodiment are provided, and the same technical effects can be achieved. To avoid repetition, they will not be repeated here.

[0072] In this embodiment, maintaining high weights for recently frequent edges automatically concentrates analytical resources and attention on the most recently occurring and active suspicious behavior patterns, helping to improve the detection sensitivity and response speed for ongoing or recently occurring fraudulent activities. Differential time decay makes the network's modeling of various relationships more realistic, providing a dynamic representation that aligns with business intuition and offering authentic and reasonable input to downstream algorithms. Adaptive filtering of historical noise enhances the model's ability to identify new patterns, reducing interference from historical random events or outdated fraudulent techniques on current judgments, and improving the model's adaptability and generalization ability. Weight decay eliminates the need for complex model retraining, facilitating low-overhead, routine dynamic maintenance of the network.

[0073] In some embodiments, a dynamic heterogeneous network is constructed to map entities such as patients, doctors, hospitals, diagnostic codes (e.g., ICD-10), and drugs / treatment items as heterogeneous nodes. Dynamic edge relationships, including primary care relationship, diagnostic association, medication association, and spatiotemporal proximity, are established, transforming fragmented medical data into a structured knowledge graph. This network introduces a time decay function to adjust historical edge weights, supports real-time embedding of new nodes and iterative updates of old node attributes, maintains network timeliness, and accurately depicts complex relationships between entities.

[0074] This embodiment uses a medical claims scenario as an example for specific explanation.

[0075] In complex medical claims scenarios, the first step is to map each of the diverse medical entities as differentiated nodes in a heterogeneous network. Based on this, and grounded in real-world medical behavior and business logic, a multi-dimensional, dynamically evolving edge relationship system is constructed. The specific steps are as follows: 1. Node Type Construction: Based on the core elements of medical claims, entities are mapped to differentiated nodes in the network, including: patient (uniquely identified by customer number), doctor (uniquely identified by employee number), hospital, ICD-10 diagnostic code, drug / treatment item, consultation time, geographical information, and report / indicator.

[0076] 2. Edge Relationship Construction: Based on medical business logic and clinical time patterns, four core edge relationships are involved, including: (1) Primary and secondary relationship: Patient (customer number) → Doctor (employee number) → Hospital, clearly defining the relationship of diagnosis and treatment responsibility.

[0077] (2) Diagnostic correlation: Disease A → Disease B, for example, diabetes → retinopathy, reflecting the co-occurrence pattern of complications and diseases, and the pathological evolution path.

[0078] (3) Medication association: diagnosis → drug / treatment item, such as hypertension → ACEI drugs, presenting a clinical decision-making path from diagnosis to prescription or specific drug / treatment item.

[0079] (4) Spatial and temporal proximity: Based on the principle of spatial and temporal proximity such as "registering in multiple departments on the same day" and "re-visiting across hospitals in a short period of time", we capture the abnormal medical treatment behavior of the same patient across institutions and departments within a specific time period.

[0080] 3. Dynamic Weight Updates: To adapt to the rapid evolution of fraud patterns, a time decay function is introduced into the network to dynamically adjust variable weights, focusing on recent high-frequency interaction behaviors. Simultaneously, it supports real-time embedding of new nodes and iterative updates of old node attributes, thus forming an intelligent network architecture that can reflect the spatiotemporal evolution of medical behavior and adapt to the dynamic changes in fraud patterns. The specific steps are as follows: (1) For each edge of the network Set timestamp , This indicates the last update time for this edge.

[0081] (2) A time decay function is used to dynamically adjust the weights of historical edges. Based on the interval between the current time and the last update time of the edge, decay coefficients are set for different types of edges to suit the characteristics of health insurance medical claims. This paper uses exponential decay:

[0082] in, This indicates the interval between the current time and the last update time of the edge. The attenuation coefficient is used to control the attenuation rate. Based on the characteristics of health insurance medical claims, different types of edge values ​​are set, and the attenuation coefficients are shown in the table below.

[0083]

[0084] (3) Combine (1) and (2) to dynamically adjust the edge weights and update the edge weights as follows: For frequently accessed recent edges, a memory cache is established to accelerate decay calculation. A scheduled task is set up at midnight to monitor AUC-ROC and recall rate, and the decay formula is immediately applied based on newly added and modified edges and business feedback, triggering a global update.

[0085] (4) Adding new node embedding: First, collect available metadata based on entity category as initial features. Initialize the embedding. Then, locally aggregate the neighboring nodes of the newly added nodes to generate pre-trained features. Finally, the initial features and pre-trained features are merged, and after dimensionality reduction and zero-mean normalization by a fully connected layer, they are used as the final node embeddings.

[0086] .

[0087] (5) Old Node Feature Update: As a key mechanism for the continuous evolution of dynamic heterogeneous networks, its core lies in accurately learning long-range dependencies between nodes by deeply mining and integrating the rich information contained in the k-order neighbors. For example, for the node feature update of a certain customer, the network will comprehensively consider the attending physician, the hospital visited, and other multi-hop neighbor information such as other patient groups and commonly used drug categories that are further connected to these entities. Whenever a new claim request arrives, the following process is triggered: a. Neighborhood change detection: Compare the old and new edge sets to identify the list of affected nodes.

[0088] b. Incremental feature refresh: Only the feature vectors of the affected nodes and their K-order neighbors are recalculated.

[0089]

[0090] in, The weight matrix represents the model's learning parameters. The optimal value of k is determined based on business scenario experience and a grid search method. Typically, in medical scenarios, k=2~3 is sufficient to cover most clinical pathways (e.g., diagnosis → treatment → rehabilitation), but for cross-institutional fraud, k=4~5 may be necessary. The loss function combines cross-entropy loss with a graph-based regularization term, minimizing the distance between connected nodes and maximizing the distance between disconnected nodes. The AUC-ROC curve on the validation set is monitored, and training automatically stops if there is no improvement for N consecutive rounds.

[0091] c. Version Management: Maintain the version number of the feature cache to ensure that downstream models read the latest valid features.

[0092] In yet another exemplary embodiment, based on step 130 of the above embodiment, and based on the dynamic heterogeneous network, path exploration is performed through a meta-path constraint traversal strategy. The method of this embodiment may further include the following specific steps: At each step of the meta-path constraint walk, the probability of walking to the next node is calculated based on the time decay weights of all adjacent edges of the current node.

[0093] The weights, as one of the inputs to this step, can be dynamic edge weights after time decay, or they can be derived from the previously mentioned steps (e.g., step 320), where the weights are dynamically calculated using a time decay function to reflect the current importance of the edge. These weights encode the strength of the relationship's timeliness. Higher weights indicate a more active and closer association between the two nodes connected by the edge in the recent past, or a more important relationship type itself. In the context of risk detection, high-weight edges may indicate frequently occurring suspicious associations in the recent past.

[0094] Another input to this step is all the adjacent edges of the current node. Based on the predefined meta-path, the type of the next node and edge to be visited in the current step is determined. First, from all the neighbors of the current node, those neighbor nodes connected by edges that meet the requirements of the meta-path are selected to form a candidate node set. This ensures that the traversal path always follows a business-meaning pattern and avoids getting lost in aimless random wandering in a large heterogeneous network. This is the key to achieving semantically guided search.

[0095] For each node in the candidate node set, the dynamic weight of the edge connecting it to the current node is used as the basis for calculating its selection probability. For example, normalization can be performed to transform all weights into a probability distribution. Nodes with higher edge weights have a greater probability of being selected. This guides the walk to naturally converge towards recently active or structurally important regions. At the same time, because it is a probabilistic selection, rather than simply selecting the node with the highest weight, it still retains a certain degree of random exploration capability, and has the opportunity to discover atypical but potentially high-risk paths.

[0096] In this embodiment, by directly converting dynamic edge weights into walk probabilities, the search focus can be automatically and in real time adjusted, favoring the exploration of associated regions with unusually high recent activity. These regions are often highly correlated with potential fraud clusters, giving the walk process an inherent risk orientation. Because edge weights undergo time decay, the weights of outdated and inactive historical relationships are reduced, thus weakening their potential to guide walks. This makes the walk strategy more focused on newly formed network structures and behavioral patterns, ensuring that discovered anomalies are timely and reflect current or evolving risks. Probabilistic selection follows meta-paths and weight guidance, introducing controllable randomness. This allows for the possibility of accidentally jumping into seemingly low-weight but actually novel paths, preserving the ability to discover hidden, new, or emerging fraud patterns and enhancing proactive detection capabilities.

[0097] In yet another exemplary embodiment, the method of this embodiment may further include the following specific steps: updating the weights of relevant edges in the dynamic heterogeneous network in reverse based on the results of path exploration.

[0098] In this embodiment, a closed-loop feedback mechanism is introduced to update the weights of the dynamic heterogeneous network based on the path exploration results. This mechanism continuously collects the access frequency of each edge and the dwell characteristics of related nodes during the walk, quantifies the importance score of each edge from the perspective of risk detection, and dynamically adjusts the weights of the corresponding edges in the network based on this score. For example, the weight of an edge that is frequently touched in recent explorations and whose associated nodes stay for a long time, representing a high-risk pattern, will be increased, thereby guiding subsequent walks to focus more on this area; conversely, the weight of edges that have not been paid attention to for a long time will be relatively reduced.

[0099] The above steps, in conjunction with the global time decay function, enable the dynamic evolution of the network structure. This allows for automatic identification and continuous focus on risk hotspots, accelerating the discovery and adaptation to new and evolving fraud patterns. Furthermore, the traceable evolution of weights enhances the consistency and interpretability of decisions, thereby improving the self-evolutionary capability and long-term effectiveness of the risk identification model as a whole.

[0100] In some embodiments, an adaptive meta-path constraint walking strategy is used to design meta-paths for different fraud types (e.g., fraudulent hospital admissions, fictitious projects, etc.). Based on the edge weights in the dynamic heterogeneous network after time decay, the walking path is constrained to jump between specific edge types. Reinforcement learning algorithms are used to dynamically optimize the walking path, actively capture high-risk subgraphs, and update network parameters through backpropagation to actively discover new high-risk fraud patterns.

[0101] This embodiment uses a medical claims scenario as an example for specific explanation.

[0102] The specific process of adaptive meta-path constraint walk is as follows: 1. Initial Meta-Path: Combining the knowledge of risk control experts, predefine meta-paths for each type of existing fraud scenario (e.g., patients corresponding to fraudulent projects → diagnosis → high-priced drugs), and compile the meta-paths into constraint masks in the form of adjacency matrices.

[0103] 2. Network Weight-Guided Walk Probability: The walk direction is guided by the edge weights in the dynamic heterogeneous network after time decay, constraining the walk to only jump between specific edge types and avoiding random and ineffective roaming. A reinforcement learning algorithm is used to dynamically optimize the walk path and capture high-risk subgraphs. The steps are as follows: (1) Starting point and probability calculation: Based on the current pending claim, select TOP-K suspicious nodes from the dynamic network as the starting point for the walk. In each step of the walk, calculate the walk probability of the next node based on the weight of the adjacent edges of the current node:

[0104] in, It is a connection node and candidate nodes The dynamic weights of edges are usually calculated using a time decay function; It is a meta-path constraint mask or indicator function, usually taking the value 0 or 1, only changing when the edge... When the meta-path constraint is satisfied ( Its dynamic weights Only then will it be included in the probability calculation; if it does not meet the following conditions ( Even if the weight of this edge is high, its contribution is zero, which is equivalent to being filtered out from the candidate set in the current step; Represents a node The set of all neighboring nodes of a node Each neighbor node Calculate their corresponding Sum the weighted scores of all candidate neighbors that satisfy the meta-path constraints to ensure that the final calculated probabilities of all candidates are obtained. The sum of these values ​​equals 1, forming a valid probability distribution.

[0105] (2) Adaptive Update: Medical fraud detection requires a focus on stability and real-time performance. Reinforcement learning algorithms such as Proximal Policy Optimization (PPO) are more stable in sparse reward environments and are suitable for high-risk scenarios. By limiting the policy update step size to balance exploration and utilization, the system needs to update immediately after each walk, and respond to the latest fraud patterns by fine-tuning the walk strategy in real time.

[0106] The current node, historical access paths, and time windows are integrated into a state vector. The next node is selected based on the probability constraints of the metapath. If the traversal termination node is determined to be fraudulent, a positive incentive of +1 is given; otherwise, a penalty of -0.1 is given. After each traversal, the policy network parameters are updated immediately to avoid decision delays caused by data backlog.

[0107] 3. The network parameters are updated in reverse order based on the traversal results. By fully recording the number of visits and dwell time of each traversal path, the importance of nodes is quantitatively evaluated, and the edge weights in the network are dynamically adjusted in reverse. This forms a closed-loop mechanism of "exploration-feedback-optimization".

[0108] (1) Collect the number of visits to each edge within a period (one week), calculate the average dwell time of each node, and record the last update time to construct the basic dataset. Then perform the following steps: (2) Access Frequency Weight Calculation: This measures the activity level of an edge and eliminates bias caused by differences in traffic volume between edges through normalization. The calculation is performed for each edge:

[0109] (2) Calculation of dwell time weight: This captures potential risk characteristics, such as prolonged stays within blacklisted institutions. Logarithmic transformation is used to compress extreme values ​​and prevent long-tail effects from compromising model stability. Calculations are performed for each node:

[0110] (3) Comprehensive score: weighted summation (1) (2) weights, where m is the weight coefficient. According to the characteristics of health insurance medical claims business, the frequency of business behavior is the main factor for fraud detection, and is assigned a higher weight, with m taking the value of 0.7.

[0111]

[0112] (4) Dynamically update network edge weights: Calculate the change in score = current score - previous score, where n is the learning rate, controlling the step size of each update to achieve dynamic adjustment of edge weights:

[0113] Based on this quantification, the edge weights in the network are dynamically and rationally allocated, enabling the network structure to be updated in real time. For example, in real-world business scenarios, for frequently accessed "blacklisted hospital" nodes, the system automatically identifies high access frequency of their associated edges and high node dwell time, thereby automatically increasing the weight of these associated edges and giving them higher priority in subsequent traversals, thus guiding the system to pay more attention to these high-risk areas. Conversely, edges that have not been accessed for a long time gradually lose their influence due to weight decay.

[0114] Meanwhile, this dynamic adjustment mechanism further strengthens the time decay effect, that is, as time goes by, the network always focuses on recent business dynamics, promptly captures the characteristic changes of new fraud patterns, and provides decision support for medical fraud detection that is both timely and accurate.

[0115] In yet another exemplary embodiment, based on step 140 of the above embodiment, and according to the result of the path exploration, the target risk of the claim is determined. The method of this embodiment may further include the following specific steps: The results of the path exploration are input into a multi-task learning model, wherein the first task of the multi-task learning model includes performing a binary classification of the claim request, and the second task of the multi-task learning model includes determining the risk type of the claim request that is determined to be the target risk.

[0116] Specifically, the input to the multi-task learning model can include the output of step 130, namely, the feature representation of one or more high-risk subgraphs. These features may include: topological indicators of the subgraph (e.g., density, diameter), statistical characteristics of paths (e.g., frequency of occurrence of specific meta-paths), key entities involved and their attributes (e.g., the level of the hospital involved, whether the drug is in the monitoring catalog), etc. These graph structure, semantic, and statistical features are fused and encoded into a unified feature vector, which serves as the input to the multi-task learning model.

[0117] Multi-task learning models can adopt a shared bottom layer-multi-task head structure. The shared bottom layer can be a deep neural network (such as a multi-layer fully connected network), which is responsible for learning from input features and extracting general, high-level abstract feature representations that are valuable for various risk assessment tasks.

[0118] The first task is basic binary classification, which can be achieved through a binary classifier. Based on the common features extracted from the shared underlying layer, it focuses on determining whether the current claim request has the target risk and outputs a fraud probability or a "risk / normal" label to achieve a high recall rate and ensure no false negatives.

[0119] The second task is risk type determination, which can be achieved through a multi-classifier. Based on shared underlying features, it may also incorporate some features that are more sensitive to type differentiation. It is responsible for further subdividing the specific risk types of cases that have been determined to be risky by the first task. For example, it outputs categories such as "falsely issued medicines", "hospitalization without a hospital bed", "impersonation for medical treatment", and "over-treatment" and their corresponding confidence levels.

[0120] The model is trained end-to-end using historical data with both overall risk labels and specific risk type labels. The loss function is typically a weighted sum of the losses from two tasks (e.g., binary cross-entropy loss combined with multi-class cross-entropy loss), which optimizes both shared and task-specific parameters simultaneously through backpropagation.

[0121] The model can simultaneously output two structured results, including: a basic risk assessment result, such as a risk probability of 0.92 or a high risk assessment, and a fine-grained risk classification result, such as a risk type of fraudulent drug prescription (confidence level 0.85) or hospitalization without a confirmed bed (confidence level 0.10).

[0122] In this embodiment, two key judgments are completed in parallel by a multi-task model, which shortens the overall judgment time, reduces computational overhead, and helps cope with the processing pressure of massive, real-time claims requests, thereby improving the automation efficiency of underwriting operations. The two tasks share underlying feature learning, avoiding feature redundancy or insufficient information that may occur when modeling each task separately, thus improving the model's generalization ability and judgment accuracy. The model output includes structured intelligence with specific type labels and confidence levels, providing guidance for subsequent manual review and business processing, facilitating precise risk triage and differentiated strategy application, and improving targeting and effectiveness.

[0123] In yet another exemplary embodiment, based on step 140 of the above embodiment, and according to the result of the path exploration, the target risk of the claim is determined. The method of this embodiment may further include the following specific steps: An evidence chain is generated based on the target risk assessment results. The evidence chain includes a key subgraph and a text report. The key subgraph is obtained based on the path exploration and identification, and the text report is obtained by performing natural language processing on the assessment criteria.

[0124] The key subgraph is used to generate visual evidence. It is the high-risk subgraph that contributes the most to the current risk assessment, identified in step 130 (path exploration). This subgraph is a local structure in a dynamic heterogeneous network, containing entity nodes such as patients, doctors, hospitals, diagnoses, and drugs that are directly related to the current claim, as well as the suspicious association edges between them.

[0125] The text reports are generated through natural language and can be obtained based on the fusion of multiple information sources, including risk assessment results, semantic information of key subgraphs, and the triggering business rules and policy basis.

[0126] In this embodiment, the generated key subgraphs and text reports can form a complete and visualized chain of evidence, clearly describing the location, source, and cause of the risk, thus meeting the interpretability requirements of the risk control model. Presenting the most suspicious core associations graphically and transforming complex graph analysis conclusions into concise natural language descriptions can reduce manual verification costs and improve operational efficiency. The generated text reports can also serve as high-quality labeled data, feeding back into the model training phase to provide supervisory signals accurate to specific associations. These signals can be used to fine-tune the walk strategy, adjust network weights, or optimize the judgment model, contributing to the continuous evolution of the system.

[0127] In some embodiments, fraud determination and evidence generation are achieved by employing a multi-task learning framework to output fraud decision results and key evidence chains, visually presenting high-contribution subgraphs. Natural Language Generation (NLG) technology is used to automatically generate review opinions, assisting in manual review and incorporating them into model iteration and optimization, forming a closed-loop risk control system.

[0128] This embodiment uses a medical claims scenario as an example for specific explanation.

[0129] The primary task uses a binary classification (fraud / normal) as the basic criterion, while an auxiliary task is set up for fine-grained multi-classification of fraud types (e.g., fraudulent booking / false invoicing / impersonation), forming a two-layer judgment mechanism combining coarse screening and fine judgment. The generation of an interpretable chain of evidence includes the following two aspects: Decision path visualization: By analyzing the correlation graph of each pending business through GraphMask, the subgraph with the greatest contribution is visualized, and the risk transmission path is presented intuitively, enabling reviewers to quickly locate suspicious links and significantly reduce the cost of manual verification.

[0130] Intelligent Report Generation: Combining Natural Language Generation (NLG), the system transforms complex graph structure analysis results into standardized natural language statements. Based on healthcare industry terminology standards, the system automatically generates structured review opinion descriptions for different types of fraud, along with risk level indicators and policy basis clauses, meeting the comprehensive evidentiary requirements of regulatory requirements.

[0131] Corresponding to the insurance claim processing method provided in the above embodiments, based on the same technical concept, this application also provides an insurance claim processing apparatus. See [link to previous document]. Figure 4 The device 400 includes a data preprocessing module 410, a network construction module 420, a path exploration module 430, and a risk assessment module 440.

[0132] The data preprocessing module 410 is used to preprocess the initial data associated with the claim request to obtain preprocessed data; the network construction module 420 is used to generate a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request; the path exploration module 430 is used to explore paths based on the dynamic heterogeneous network using a meta-path constraint traversal strategy; and the risk assessment module 440 is used to assess the target risk of the claim request based on the results of the path exploration.

[0133] It should be noted that the insurance claim processing device and the insurance claim processing method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned insurance claim processing method, and the repeated parts will not be described again.

[0134] See Figure 5 A health insurance medical claims fraud prevention and identification method based on dynamic heterogeneous networks and adaptive walking strategies may include the following specific process: Claims Request Access: The starting point of the process, where the system receives claims applications and related data pending review.

[0135] Data preprocessing: Cleaning, transforming, and standardizing the raw, multi-source, heterogeneous claims data to prepare high-quality input for subsequent analysis.

[0136] Dynamic heterogeneous networks: Based on preprocessed data, a network knowledge graph is constructed that can characterize the complex relationships between entities and has the ability to evolve over time.

[0137] Adaptive random walk: Under the semantic constraints of the meta-path, it performs intelligent path exploration, proactively discovering high-risk patterns and subgraphs in the network. The network weight-guided walk probability includes utilizing the real-time weights of edges in the dynamic network to calculate and guide decision probabilities for subsequent path exploration. The walk results can also be used to back-update edge weights in dynamic heterogeneous networks.

[0138] Fraud determination: Based on the pattern of exploration and discovery, the claim request is determined to determine whether there is fraud or risk.

[0139] Fraud classifier: Based on the determination that there is a risk, it further classifies the risk into finer-grained categories.

[0140] Evidence generation: Automatically generate visual evidence (e.g., key sub-graphs) and textual evidence (e.g., review reports) that support the above determinations.

[0141] Evidence chain output: The end point of the process, outputting a structured, interpretable, and complete evidence package for manual review or use by subsequent business systems.

[0142] The insurance claims processing method and apparatus provided in this application include a dynamic heterogeneous network architecture for constructing a multi-source heterogeneous network for medical claims, introducing a dynamic edge weight update mechanism, which can improve the accuracy of cross-institutional fraud identification by 9% and reduce the false positive rate by 12%; an adaptive meta-path constraint roaming strategy that integrates risk control expert knowledge and dynamic network features to reduce manual intervention, proactively discover new fraud patterns, and improve the speed of detecting new fraud claims by 3 times; a network and roaming strategy collaborative closed-loop mechanism to achieve the co-evolution of network structure and roaming strategy, effectively preserving key features, with model update delay not exceeding 90 seconds; and an interpretable evidence chain generation that provides visual review basis by defining a multi-classification tasker, meeting the requirements of regulatory penetration management, reducing the workload of manual review, and improving operational efficiency.

[0143] Corresponding to the insurance claim processing method provided in the above embodiments, based on the same technical concept, this application also provides an electronic device for executing the above method. Figure 6 To illustrate the structure of an electronic device according to various embodiments of this application, as shown in the following diagrams... Figure 6 As shown. Electronic device 500 can vary considerably due to differences in configuration or performance, and may include one or more processors 510 and memory 520. Memory 520 may store one or more application programs or data. Memory 520 may be temporary or persistent storage. The application programs stored in memory 520 may include one or more modules (not shown), each module may include a series of computer-executable instructions for the electronic device. Furthermore, processor 510 may be configured to communicate with memory 520 and execute the series of computer-executable instructions stored in memory 520 on the electronic device.

[0144] This application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, implement the steps of the insurance claim processing method described above.

[0145] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0146] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems, devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0147] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0148] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] In a typical configuration, an electronic device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0150] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0151] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0152] It should also be noted that 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 limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0153] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0155] It should be understood that the training and prediction processes of the artificial intelligence (AI) models involved in the various embodiments of this specification all adhere to multiple legal and compliant principles, including legal data sources, compliant data content, compliant data governance, compliant training objectives and schemes, compliant training processes, compliant training environments and tools, and compliant ethical verification of training results, and comply with the requirements of Article 5 of the Patent Law. Among them: Data source legitimacy: All datasets used for AI model training were obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, with complete copyright attribution and authorization scope clearly marked, and no unauthorized open-source code or data reuse. Data authorized by partners has been subject to formal data usage agreements, clearly defining the scope, duration, and confidentiality obligations, and possessing a complete authorization chain. For self-collected data involving personal information, strict informed consent procedures have been followed, and anonymization processes (including but not limited to field masking, feature anonymization, and differential privacy technology applications) have been implemented to remove personally identifiable information, fully complying with the requirements of relevant laws and regulations such as the "Interim Measures for the Administration of Generative Artificial Intelligence Services" and the "Personal Information Protection Law."

[0156] Data content compliance: The AI ​​model's dataset undergoes multiple screenings and cleaning processes to remove all content that may violate social morality or harm public interests. It contains no obscene, pornographic, violent, discriminatory, or information that endangers national or public safety, nor does it involve the illegal acquisition or use of genetic resources. For data in sensitive fields (such as healthcare and finance), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technologies) ensures that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0157] Data governance norms: A complete data traceability system is established during the AI ​​model training process to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure that the data is verifiable throughout its entire lifecycle. The dataset annotation process for AI models is completed by a professional human R&D team, clearly defining the proportion of human creative contributions and avoiding reliance on AI-generated data that has not undergone substantial human modification, thus meeting the examination requirements for "human main contributions" in AI patent applications.

[0158] Training objectives and plans are compliant: The training objective of the AI ​​model focuses on [specific technical scenarios that can be supplemented, such as intelligent driving decision optimization, multimodal information interaction, etc., and replaced based on specific content]. The training scheme and the final output results do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risks of being used for illegal activities, infringing on privacy, or disrupting public safety. The model strictly adheres to the ethical principle of "intelligent for good".

[0159] Training process compliance: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision-making logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on expert system feedback to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop control of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0160] Training environment and tool compliance: AI model training is implemented using nationally licensed chips and a compliant training platform. All open-source frameworks and components used in the training process have obtained their corresponding licenses, and copyright statements and patent citation information are fully retained, with no instances of infringement or reuse. The training environment is built using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. Furthermore, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0161] Training results ethical verification compliance: After the model is trained, it undergoes additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism is established to ensure that the model always complies with Article 5 of the Patent Law and relevant laws and regulations in practical applications.

[0162] In summary, the data and training process used in the AI ​​model of this specification strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there are no violations of laws, social ethics, public interests, or illegal use of genetic resources. It fully meets the compliance requirements for patent authorization.

Claims

1. A method for processing insurance claims, characterized in that, The method includes the following steps: The initial data associated with the claim request is preprocessed to obtain preprocessed data; A dynamic heterogeneous network is generated based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request; Based on the dynamic heterogeneous network, path exploration is performed using a meta-path constraint traversal strategy. Based on the results of the path exploration, the target risk of the claim request is determined.

2. The method of claim 1, wherein, The process of preprocessing the initial data associated with the claim request to obtain preprocessed data includes the following steps: The initial data is subjected to at least one of the following processes: de-identification, missing value imputation, and standardization.

3. The method of claim 1, wherein, The step of generating a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on entities in the preprocessed data, and the edges between the nodes are obtained based on business logic associated with the claim request, further includes the following steps: The newly added entity is embedded as a new node into the dynamic heterogeneous network, and the feature vectors of the affected nodes in the dynamic heterogeneous network are incrementally updated.

4. The method of claim 1, wherein, The step of generating a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on entities in the preprocessed data, and the edges between the nodes are obtained based on business logic associated with the claim request, includes the following steps: The weights of the edges are dynamically adjusted based on a time decay function, which uses exponential decay and different decay coefficients are configured for different types of edges.

5. The method according to claim 1, characterized in that, The path exploration based on the dynamic heterogeneous network and the meta-path constraint walking strategy includes the following steps: At each step of the meta-path constraint walk, the probability of walking to the next node is calculated based on the time decay weights of all adjacent edges of the current node.

6. The method according to claim 1, characterized in that, The step of determining the target risk of the claim request based on the results of the path exploration includes the following steps: The results of the path exploration are input into a multi-task learning model, wherein the first task of the multi-task learning model includes performing a binary classification of the claim request, and the second task of the multi-task learning model includes performing a risk type determination on the claim request determined as the target risk.

7. The method according to claim 1, characterized in that, The step of determining the target risk of the claim request based on the results of the path exploration also includes the following steps: An evidence chain is generated based on the target risk assessment results. The evidence chain includes a key subgraph and a text report. The key subgraph is obtained based on the path exploration and identification, and the text report is obtained by performing natural language processing on the assessment criteria.

8. An insurance claim processing device, characterized in that, The device includes the following: The data preprocessing module is used to preprocess the initial data associated with the claim request to obtain preprocessed data; A network construction module is used to generate a dynamic heterogeneous network based on the preprocessed data, wherein the nodes in the dynamic heterogeneous network are obtained based on the entities in the preprocessed data, and the edges between the nodes are obtained based on the business logic associated with the claim request; The path exploration module is used to explore paths based on the dynamic heterogeneous network using a meta-path constraint traversal strategy; and, The risk assessment module is used to assess the target risk of the claim request based on the results of the path exploration.

9. An electronic device, characterized in that, It includes a processor, a memory, a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, implement the steps of the method as described in any one of claims 1 to 7.

Images