A method and computer equipment for intelligent claims risk control in internet health insurance based on multi-source data fusion

Abstract

This application relates to the fields of insurance technology and data processing technology, and discloses an intelligent claims risk control method and computer equipment for internet health insurance based on multi-source data fusion. The method includes responding to claims applications and calculating historical and recent behavioral risk scores based on a customer risk profile knowledge graph; reviewing claims materials using a medical knowledge graph and object detection model, and outputting scores for unreasonable medication use and invoice compliance; obtaining relevant clauses based on a retrieval-enhanced generation framework and outputting liability determination indicators; aggregating the above scores and indicators to construct multimodal features; using an ensemble learning classifier to determine the risk level and generate a risk warning summary. By integrating knowledge graph, computer vision, and large language model technologies, it achieves deep correlation analysis of structured profiles, unstructured documents, and insurance clauses, solving the problems of sparse and poor interpretability of traditional risk control features, and improving the accuracy of risk identification and the efficiency of automated processing in claims cases.

Description

Technical Field

[0001] This invention relates to the fields of insurance technology and data processing technology, specifically to an intelligent claims risk control method and computer equipment for internet health insurance based on multi-source data fusion. Background Technology

[0002] With the large-scale development of internet health insurance, claims are becoming increasingly frequent, fragmented, and non-standardized, posing challenges to insurance companies' claims processing efficiency and risk control capabilities. Traditional claims risk control models primarily rely on manual review or automated systems based on rule engines. While manual review possesses strong logical judgment capabilities, it is limited by labor costs and struggles to handle massive concurrent claims, and review standards are easily influenced by subjective factors. Rule-based systems, although faster, typically only perform hard filtering based on set thresholds or blacklists, making it difficult to address highly concealed fraudulent activities.

[0003] Existing intelligent risk control technologies still suffer from insufficient data utilization and weak logical reasoning capabilities when processing health insurance claims. Current models primarily focus on analyzing the applicant's basic structured data, such as historical payout amounts or policy status. However, they often lack effective deep integration methods for the large amount of unstructured data generated during the claims process, such as medical bill images, medical record texts, and complex insurance policy documents. This makes it difficult for the system to identify cross-modal risk characteristics, such as medically inappropriate treatment plans, minor alterations to invoice stamps, or deep semantic conflicts between case descriptions and insurance liability clauses. Furthermore, existing risk control models typically employ an offline training and online prediction model. Model parameters are relatively fixed after deployment, lacking dynamic adaptability to emerging risk patterns or manual correction feedback. This leads to performance degradation over time, making it difficult to balance the retention of historical knowledge with the learning of new risk characteristics. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent claims risk control method for internet health insurance based on multi-source data fusion. This method solves the problem that existing health insurance claims risk control technologies cannot effectively integrate heterogeneous data from multiple sources, such as applicant behavioral profiles, medical document images, and insurance clause texts, resulting in low risk identification accuracy, poor interpretability, and delayed model updates for complex claims cases.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an intelligent claims risk control method for internet health insurance based on multi-source data fusion, comprising the following steps: In response to a claim, based on a pre-built customer risk profile knowledge graph, the applicant’s historical behavioral risk score, recent medical treatment behavior risk score, and related network risk score are calculated respectively, and the aforementioned scores are merged to generate a comprehensive customer risk score. Acquire data on claims materials, use medical knowledge graphs and object detection models to review medical behaviors and invoice formats respectively, and output scores for irrational drug use risk and invoice stamp compliance. Relevant insurance clauses are obtained using vectorized retrieval, and the confidence level of liability determination and the degree of rule conflict are output through a fine-tuned large language model. The aforementioned scores and metrics are aggregated to construct a multimodal fusion feature vector, and an ensemble learning classifier is used to determine the risk level and generate a risk warning summary. By comparing the manual adjudication results of closed cases with the model prediction results, cases with biases are identified and corrected data after manual review is obtained. Incremental updates are then performed on the fine-tuned large language model and the ensemble learning classifier.

[0006] The calculation of the applicant's historical behavioral risk score involves: using the applicant's unique identifier as an anchor point to traverse the customer risk profile knowledge graph and extracting a local risk subgraph containing historical claim event nodes. This subgraph is then traversed to obtain the payout amount attribute and case closure time attribute for each historical claim event node, calculating the time interval between the case closure time attribute and the current claim application trigger time. An exponential decay function is used to calculate the time decay weight for each historical claim event node, which decreases non-linearly with increasing time interval. The payout amount attribute of each historical claim event node is multiplied by its corresponding time decay weight, and the results are summed to obtain the time-weighted total payout. This time-weighted total payout is then divided by the applicant's cumulative coverage duration to obtain the historical behavioral risk score.

[0007] The calculation of the applicant's recent medical visit risk score is as follows: A short-term risk observation window is set based on the current claim application trigger time. Medical visit event nodes falling within this window are retrieved from the customer risk profile knowledge graph. The total number of medical visit event nodes is counted to obtain the total number of visits. The logarithmic value of the total number of visits plus one is calculated as the logarithmic smoothing value of the medical visit density. The medical institution attributes associated with the medical visit event nodes are extracted, and an institution visit distribution sequence is constructed. The ratio of the standard deviation to the mean of the number of institutions in this sequence is calculated as the coefficient of variation. The coefficient of variation is summed with a baseline constant to obtain the medical institution dispersion factor. The logarithmic smoothing value of the medical visit density is multiplied by the medical institution dispersion factor to obtain the recent medical visit risk score.

[0008] The output of the irrational medication risk score is as follows: Diagnostic information, drug lists, and item lists are extracted from the claim data. These drug and item lists are transformed into claim feature vectors. A set of standard treatment plans corresponding to the diagnostic information is retrieved from a medical knowledge graph, and graph embedding technology is used to transform this set into a standard plan vector. The cosine similarity between the claim feature vector and the standard plan vector is calculated, and the maximum similarity value is selected. Drugs or items not included in the standard treatment plan set are identified in the claim feature vector, and their total cost is calculated as the proportion of abnormal drug costs to total medical expenses. The product of the abnormal drug cost proportion and a preset penalty coefficient is subtracted from the maximum similarity value, and the difference between 1 and the difference obtained from this subtraction is calculated to obtain the irrational medication risk score.

[0009] The output of the document seal compliance score is as follows: An object detection model is used to identify the seal categories and corresponding detection confidence levels in the claims material data. Based on the document type corresponding to the claims material data, a preset required seal combination rule base is invoked, iterating through each required seal category. An indicator function is used to determine: when the detection confidence level of a required seal category is greater than a preset detection threshold, the detection confidence level is retained; when the detection confidence level is not greater than the preset detection threshold, the detection confidence level is recorded as zero. The average of the retained detection confidence levels for all required seal categories is calculated as the document seal compliance score.

[0010] Regarding the acquisition of relevant insurance clause fragments: A hybrid weighted matching strategy is employed to generate a case query vector. The basic semantic similarity between the case query vector and the pre-constructed insurance clause fragment vector is calculated. A position correction coefficient is calculated based on the normalized position index of the insurance clause fragment in the document; this coefficient decreases exponentially with increasing normalized position index. A type weighting coefficient is assigned to the insurance clause fragment based on its business type. The basic semantic similarity, position correction coefficient, and type weighting coefficient are then fused and calculated. Based on the calculation results, insurance clause fragments are selected.

[0011] Regarding the generation of risk warning summaries: An approximate risk probability function for the ensemble learning classifier is constructed, and the partial derivatives of this function with respect to each component of the multimodal fusion feature vector are calculated. The partial derivatives of each component are multiplied by the normalized feature values ​​to obtain the marginal contribution value of each feature. The top N features with the largest absolute values ​​of marginal contribution are selected as the dominant risk factors. Based on a pre-built interpretation template library, the feature values ​​of the dominant risk factors are mapped to natural language descriptions, and these descriptions are then concatenated to generate the risk warning summary.

[0012] Regarding incremental model updates: A parameter update strategy based on elastic weight consolidation is adopted, constructing a hybrid fine-tuning dataset containing corrected case data and historical core samples. During the model parameter update process, a regularization penalty term is added to the loss function. This regularization penalty term, calculated based on the Fisher information matrix, is used to constrain parameters highly important to historical tasks from significant changes. The loss function is expressed as the sum of the new task loss and the regularization penalty term, which is proportional to the product of the square of the parameter change and the corresponding Fisher information value.

[0013] A second aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0014] This invention provides an intelligent claims risk control method for internet health insurance based on multi-source data fusion. It has the following beneficial effects: 1. This invention constructs a multimodal fusion feature vector by aggregating customer risk profile knowledge graphs, medical material image data, and insurance clause text data. This effectively overcomes the problem that traditional claims risk control mainly relies on structured fields and has low utilization of unstructured medical documents and complex clause logic. By using a medical knowledge graph to verify the medical rationality of treatment plans and combining it with a target detection model to review the formal standardization of invoices and seals, it achieves comprehensive risk scanning from treatment behavior to claims materials, improving the accuracy of identifying fraudulent insurance claims and abuse of medical resources.

[0015] 2. This invention introduces time-dynamic weighting and a hybrid retrieval strategy in risk quantification calculation, improving the precision of feature engineering. By weighting historical payout amounts using an exponential decay function, the risk score more accurately reflects the applicant's recent risk status, reducing the interference of long-term historical data on current judgments. Simultaneously, in the clause matching stage, a hybrid weighting strategy integrating semantics, location, and type is employed, combined with conflict detection using a large language model, ensuring that the basis for liability determination accurately corresponds to the specific insurance clause details, reducing erroneous claims or denials due to misunderstandings of the clauses.

[0016] 3. This invention provides an interpretability mechanism based on feature attribution and an incremental update mechanism resistant to forgetting, enhancing the system's practicality. By calculating the marginal contribution values ​​of feature components to generate natural language risk summaries, complex model parameters are mapped to human-readable review suggestions, improving human-machine collaboration efficiency. The model is updated using a flexible weight consolidation strategy based on the Fisher information matrix. While absorbing data from newly amended documents, regularization constraints maintain the model's memory of historical core risk patterns, preventing catastrophic forgetting during continuous system iteration. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the historical behavior risk score calculation process of the present invention; Figure 3 This is a schematic diagram of the computer device structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a method for intelligent claims risk control of internet health insurance based on multi-source data fusion, comprising the following steps: S1. Obtain the claim application and determine the applicant's unique identifier. Use the applicant's unique identifier as the anchor point to traverse the pre-built customer risk profile knowledge graph, calculate the applicant's historical behavior risk score, recent medical treatment behavior risk score and related network risk score, and generate a weighted aggregated comprehensive customer risk score. S2. Extract the claim material data associated with the claim application, and review the rationality of medical behavior and the standardization of invoice form based on medical knowledge graph and object detection model, respectively, and output the risk score of unreasonable drug use and the standardization score of invoice seal; S3. Use vectorized retrieval to obtain insurance clause fragments associated with the claim application, perform clause matching with case details through fine-tuned large language model, and output liability determination results, liability determination confidence and rule conflict degree; S4. Aggregate the output results of steps S1 to S3 to construct a multimodal fusion feature vector, use an ensemble learning classifier to determine the risk level of the claim case, and allocate the claim case to the automatic claim process or the manual review process according to the risk level of the claim case, and generate a risk warning summary for the claim case allocated to the manual review process. S5. Use the intelligent quality inspection model to identify closed claims cases with potential adjudication biases, and incrementally update the models involved in the above steps based on the corrected case data.

[0020] In step S1, the process of constructing a customer risk profile knowledge graph is the foundation for subsequent dynamic risk quantification. This step specifically includes the following sub-steps: A multi-dimensional data model of entities and relationships is established. The system predefines the ontology structure of the knowledge graph, with the customer as the core entity, extending to peripheral entities related to their claims behavior. Specifically, entity types include customer nodes, claims event nodes, medical institution nodes, disease diagnosis nodes, and associated person nodes. Among them, customer nodes store the customer's basic identity information and policy status; claims event nodes represent each historical claim record; medical institution nodes represent hospitals or pharmacies providing medical services; disease diagnosis nodes correspond to specific diseases in the International Classification of Diseases (ICD); and associated person nodes represent natural persons who have a family policy co-insurance relationship or beneficiary relationship with the customer.

[0021] Define semantic relationships between entities. Establish directed edges with business meaning between the aforementioned entity types. Specific connections include: a "claim initiation" relationship from a customer node to a claims event node; a "treatment at" relationship from a claims event node to a medical institution node; a "diagnosis" relationship from a claims event node to a disease diagnosis node; and a "family association" or "policy association" relationship from a customer node to an associated person node. This topology maps isolated data scattered across different business tables into a connected graph data structure, thereby supporting depth-first traversal and association analysis across nodes.

[0022] Feature filling is performed on the graph nodes and edge attributes. The system extracts structured data from the business database and maps it to the attribute values ​​of corresponding nodes or edges in the graph. For claims event nodes, the attributes include the payout amount, the time of the incident, and the claim status; for edges initiating claims, the attributes include the time interval between the claim and the current time; for medical institution nodes, the attributes include the hospital level, hospital type, and location; for disease diagnosis nodes, the attributes include the severity level of the disease or whether it is a chronic disease. For the storage and querying of graph data, those skilled in the art can use existing graph database technologies, and the underlying storage principles will not be elaborated here.

[0023] A subgraph extraction method using customer ID as the anchor point is implemented. When a new claim application is received, the system does not traverse the entire full graph. Instead, it uses the current applicant's customer ID as the starting node and performs a multi-hop traversal to retrieve all historical claim events, frequently visited medical institutions, and associated personnel directly or indirectly connected to the customer, constructing a local risk subgraph specific to that customer. This subgraph contains all the factual data required to calculate subsequent risk scores, achieving rapid dimensionality reduction from full data to sample data.

[0024] In step S1, the core of the quantitative calculation of customer historical behavioral risk lies in eliminating the assessment bias caused by a single cumulative compensation amount, and introducing correction factors based on the time dimension and the coverage period. This process is specifically implemented through the following sub-steps: The system iterates through the customer's risk subgraph and extracts all historical claim event nodes associated with that customer. For each historical claim event, the system obtains the corresponding payout amount and case closure timestamp. Simultaneously, the system reads the policy effective date attribute of the customer node and calculates the customer's cumulative coverage duration using the difference between the current time and the policy effective date.

[0025] The system calculates the time decay weight for each historical claim. To reflect the business logic that "recent behavior has a higher predictive value for current risk," the system calculates the time interval between the settlement time of each historical claim and the trigger time of the current claim application. This time interval is fed into an exponential decay function to generate a corresponding weight coefficient. This weight coefficient decreases non-linearly with increasing time interval, thus ensuring that recent claims dominate the overall risk score, while the impact of older claims is gradually diluted.

[0026] The system performs weighted aggregation and normalization on historical claims. For each historical claim, the system multiplies the claim amount by its corresponding time decay weight to obtain the discounted risk value for that claim. Then, it sums up the discounted risk values ​​of all historical claims to obtain the time-weighted total claim amount. To objectively evaluate customers with different insurance durations and avoid misjudging customers with naturally accumulated claims due to longer coverage periods as high-risk, the system divides the aforementioned time-weighted total claim amount by the customer's cumulative coverage duration to obtain the average risk contribution value per unit time, i.e., the historical behavioral risk score.

[0027] The above calculation logic is specifically manifested as follows: the system uses an exponential decay model to calculate historical behavior risk scores. The formula is as follows: ; in: For the first The amount of compensation paid in this historical case; This refers to the time interval between the settlement time of this claim and the time when the current claim application was triggered; This is the time decay coefficient, used to control the weight of recent claims. The cumulative coverage period for customers; This is the smoothing constant.

[0028] In step S1, the assessment of recent medical behavior risk aims to capture the applicant's unusually high frequency of medical visits or cross-institutional medical visits within a short period of time. This process is performed through the following steps: The system identifies a short-term risk observation window and extracts the medical visit sequence. Based on the current claim application time, the system traces back a preset number of days and retrieves all medical visit event nodes falling within that time window from the customer's risk profile knowledge graph.

[0029] The system calculates the log-smoothed value of the medical visit density. Considering the long-tailed distribution of medical visit frequency across different populations, directly using the raw numbers may distort the risk score due to extreme values. The system increments the total number of medical visits by one before calculating its logarithmic value. Through the mapping of the logarithmic function, the linear growth of medical visit frequency is transformed into a marginally decreasing growth of the risk score, thus smoothing the impact of extreme outliers on the overall model while preserving high-frequency features.

[0030] The system calculates a healthcare institution dispersion factor to identify the risk of transient healthcare visits. It extracts the healthcare institution identifiers involved in each visit within the aforementioned time window, constructing an institution visit distribution sequence. The system calculates the statistical dispersion characteristics of the number of healthcare institutions involved in this sequence, specifically including calculating the standard deviation and mean of the institution number sequence. Subsequently, it calculates the coefficient of variation using the ratio of the standard deviation to the mean, and sums it with a baseline constant to obtain the healthcare institution dispersion factor. This factor quantifies the dispersion of customer healthcare behavior. When a customer frequently changes different healthcare institutions within a short period, the value of this factor increases significantly, effectively identifying abnormal behavioral patterns such as repeated treatments at multiple institutions or splitting costs to circumvent deductibles.

[0031] Generate a recent healthcare visit risk score. The system combines visit frequency and institution dispersion to calculate the recent healthcare visit risk score.

[0032] ; in: The total number of medical visits within a preset number of days prior to the current application time is smoothed using a logarithmic function. This represents the standard deviation of the sequence of the number of medical institutions involved in outpatient visits within this time window. This is the mean of the sequence; The coefficient of variation measures the degree of dispersion of a customer's medical locations.

[0033] In step S1, the calculation of associated network risk is based on social network analysis theory and aims to assess the risk transmission effect within the customer's social circle or community of interests. This step specifically includes the following sub-steps: The system constructs and traverses a customer relationship subgraph. Using the current claims applicant as the central node, the system expands the knowledge graph one- or multiple degrees based on predefined strong relationships.

[0034] Retrieve the independent risk status of related individuals. For each related individual in the set, the system queries their recent historical claims risk control records at the individual level. This record can be reflected as the risk score of the related individual's most recent claims application, or their status as marked on a blacklist / greylist. The system sets a high-risk judgment threshold and compares each related individual's independent risk score with this threshold.

[0035] The system calculates the proportion density of high-risk associates. An indicator function is introduced to binarize the comparison results: when an associate's independent risk score is higher than a threshold, the count is 1; otherwise, it is 0. The system sums the binarized counts of all associates to obtain the total number of high-risk associates. Subsequently, the system divides this total by the total number of associates in the network to calculate the proportion of high-risk associates.

[0036] Generate a risk score and comprehensive score for the associated network. The system retrieves a list of high-risk entities within the associated network and calculates the associated network risk score. ): ; in: This represents the set of related parties of this customer; For indicator functions, when associated people Independent risk score Greater than the preset threshold If the value is 1, then the value is 0; otherwise, the value is 0.

[0037] Finally, the system weights and aggregates the above three scores to generate a comprehensive customer risk score. ): ; in, These are the weighting coefficients for each dimension.

[0038] In step S2, the medical rationality review overcomes the limitation of traditional rule bases in handling complex medical logic by mapping treatment behaviors to a high-dimensional vector space for similarity measurement. This process specifically includes the following sub-steps: A pre-constructed medical knowledge graph of "disease-drug-treatment item" is built. The system integrates clinical treatment guidelines, drug instructions, and historical compliant claims data to construct a dedicated medical domain graph. The nodes of the graph cover all diseases, generic drug names, and medical service items corresponding to ICD codes. Edges in the graph represent treatment relationships; for example, the "indication" edge connects drugs to diseases, and the "routine treatment" edge connects treatment items to diseases. Furthermore, through graph embedding technology, the system transforms each standard treatment plan in the graph into a fixed-dimensional feature vector. A standard treatment plan typically consists of a disease node and a set of associated high-frequency drug and item nodes, and its corresponding vector is represented as a standard plan vector.

[0039] Structured extraction and vectorization of claim information. The system receives medical record text and expense list data from the claim application, and extracts the main diagnosis, medication list, and examination items for this medical visit using natural language processing technology. For the extracted diagnostic information, the system retrieves all standard treatment plans corresponding to the diagnosis from the medical knowledge graph. Simultaneously, the system uses the same embedding model to transform the list of medications and items in this claim application into a claim feature vector.

[0040] The system calculates the semantic similarity of the treatment plans. It compares the claim feature vector with each retrieved standard plan vector. The comparison uses a cosine similarity algorithm to calculate the cosine of the angle between the two vectors. A cosine value closer to 1 indicates a closer alignment between the claimed treatment and the standard plan recommended by clinical guidelines; a smaller cosine value indicates a greater deviation. The system selects the maximum similarity value from all comparisons as the fundamental indicator for assessing the reasonableness of the medical action.

[0041] An abnormal cost penalty is introduced to calculate the final score. To further identify the risk of "over-treatment" or "off-label medication," the system calculates the cost deviation. The system identifies drugs or items not included in the atlas recommendation set in this claim list (i.e., off-label medications) and sums up the total cost of these drugs or items. Subsequently, it calculates the proportion of this abnormal cost to the total medical expenses. The system subtracts the product of this cost proportion and a preset penalty coefficient from the maximum similarity value to obtain the final irrational medication risk score.

[0042] The above calculation logic is specifically manifested as follows: the system compares the characteristics of this claim with the standard plan, introduces a cost penalty item, and calculates an unreasonable medication risk score. ): ; in: This is the diagnostic and treatment feature vector for this claim; Standard treatment plan in knowledge graph eigenvectors; The cosine similarity function; This is a collection of abnormal drugs; This refers to the penalty coefficient for abnormal expenses. This represents the percentage of abnormal drug expenses.

[0043] In step S2, the review of the formalities of the invoices no longer relies on manual visual verification, but instead utilizes deep learning vision algorithms to automatically verify the formal requirements of the invoices. This process specifically includes the following sub-steps: A target detection model for medical invoice stamps is configured. The system uses the YOLO series architecture as the basic network to build the stamp detection model. This model, trained on a large amount of labeled data, can identify various common stamps on medical invoices, expense lists, and medical records, including but not limited to hospital finance stamps, payment stamps, diagnostic certificate stamps, and medical insurance office approval stamps. The model output includes the bounding box coordinates, stamp category label, and corresponding detection confidence score for each detected stamp.

[0044] A rule base for defining required seal combinations is established. Based on the compliance requirements of different types of claim documents, the system pre-defines corresponding "required seal combinations." For example, for outpatient invoices, the rule requires both a "government finance seal" and a "hospital billing seal"; for prescriptions, the rule requires either a "doctor's signature seal" or a "hospital seal"; and for discharge diagnosis certificates, the rule requires a "discharge diagnosis certificate seal." These rules are stored in sets, each set defining the indispensable seal categories for a specific document type.

[0045] The system performs a joint verification of seal completeness and clarity. It receives the output of the seal detection model and calls the corresponding required seal combination rules based on the type of document being processed. The system iterates through each required seal category in the rule set, checking if a detection instance of that category exists in the model output. To eliminate interference from blurry or incomplete seals, the system introduces a confidence threshold filtering mechanism.

[0046] Calculate the standardization score of the document seal. The system uses the recognition results to calculate the standardization score of the document seal.

[0047] ; in: The set of seal categories required for the current document type; For the model to the first Confidence level for detecting seal-like objects; The detection threshold; This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise.

[0048] In step S3, building a terms knowledge base is the fundamental step in transforming unstructured insurance contract text into computer-processable semantic vectors. This process is specifically implemented through the following sub-steps: The system performs hierarchical parsing and segmentation of insurance contract text. It reads the original insurance terms document in PDF or Word format and uses document parsing technology to identify its chapter structure. Based on the natural paragraphs and sequence numbers of the terms, the system breaks down the long text into independent semantic segments. Each segment typically contains a complete clause rule. Simultaneously, the system extracts metadata for each segment, including the chapter title, clause type label, and the clause's relative position index within the full text. The position index is normalized to the [0,1] interval for subsequent position weight calculation.

[0049] A vectorized index of clause fragments is constructed. The system utilizes a pre-trained language representation model to map each of the segmented text fragments into a dense vector of fixed dimensions. This vector captures deep semantic information in the clause text, ensuring that expressions with different literal meanings but similar connotations are close together in the vector space. The system stores the generated vectors, their corresponding original text fragments, and metadata in a vector database. For the selection and construction of the vector database, those skilled in the art can use mature technologies such as Milvus or FAISS, which will not be elaborated upon here.

[0050] A keyword-based inverted index is established to assist in retrieval. To compensate for the shortcomings of pure vector retrieval in precisely matching proper nouns, the system builds an inverted index in parallel. This index records the mapping relationship between key business terms and the IDs of clause fragments containing those terms. By combining the vector index and the inverted index, the system constructs a clause knowledge base with a hybrid index structure, capable of simultaneously supporting semantic fuzzy matching and keyword precise matching.

[0051] In step S3, in order to accurately locate the rule fragment most relevant to the current claim from the massive clause library, the system adopts a hybrid retrieval strategy that integrates location information and clause type. This process is specifically executed through the following sub-steps: Generate a case query vector. The system extracts the core case description of the current claim, including the cause of the accident, the diagnosis results, and the reimbursement items, and inputs it into the same pre-trained language model to generate a case query vector.

[0052] Calculate basic semantic similarity. Utilize the retrieval function of the vector database to calculate the cosine similarity between the case query vector and the vectors of each clause fragment in the clause knowledge base. The system initially filters out several candidate clause fragment sets with the highest similarity ranking.

[0053] The system calculates position correction and type weighting coefficients. It reads the metadata of each candidate clause fragment. For the position attribute, the system introduces a position attenuation function. Insurance contracts typically place the core definitions of general principles, insurance liabilities, and major exclusions at the beginning of the document, while procedural explanations are placed at the end. Therefore, the system calculates a position attenuation coefficient, which decreases exponentially with the clause's position index value. For the type attribute, the system assigns fixed weights to different types of clauses according to preset business rules. For example, clauses related to "exclusions" are given the highest weight, followed by "insurance liabilities," and then "definitions," thereby strengthening the focus on points of risk for claim denial.

[0054] The final clause matching score is calculated. The system integrates the basic semantic similarity, position correction coefficient, and type weighting coefficient for calculation.

[0055] In step S3, the liability determination process utilizes the reasoning capabilities of the large language model, combined with the retrieved clause context, to output a clear compensation conclusion and a quantitative indicator measuring the reliability of that conclusion. This process is specifically executed through the following sub-steps: The system constructs prompts and performs reasoning. It uses the selected Top-K clause fragments as contextual information, combines them with the current case description and the claim request to be determined, and assembles them into input prompts according to a predefined template. These prompts instruct the large language model to perform "evidence-based reasoning," requiring the model to explicitly state the supporting clauses for the determination. The system then calls upon the large language model, fine-tuned with knowledge of the health insurance domain, and receives its generated natural language judgment text. This text contains the conclusion of "payment / rejection / partial payment" along with specific reasons.

[0056] Calculate the decision confidence score. To evaluate the model's certainty regarding the output conclusion, the system extracts the token probability distribution when the model generates core conclusion terms. The system calculates the logarithmic mean probability of these key tokens and maps them to the [0,1] interval to obtain the decision confidence score. Specifically, this confidence level is equal to the set of key tokens. The probability of each token The geometric mean of the products. This indicator reflects the model's confidence in its conclusions within the current context; a lower confidence level indicates a gray area in the matching between the case details and the terms.

[0057] The system calculates rule conflict degree to identify cases with conflicting judgments. It utilizes a large model to perform multiple inference sampling or employs a consistency check mechanism to detect logical mutual exclusion between different clause fragments. The system independently generates intermediate judgment results for each retrieved highly relevant clause. Subsequently, the proportion of contradictory intermediate results is statistically analyzed. Rule Conflict Degree It is defined as the ratio of the number of inconsistent logs among all intermediate decision pairs to the total number of logs, or the entropy value based on the conflict probability distribution of the model output.

[0058] The system calculates the confidence score and rule conflict degree. It uses the output probability of the large language model to calculate the confidence score of the responsibility determination. ): ; At the same time, calculate the rule conflict degree ( To identify mutual exclusion rules: ; in: Output probabilities for large models; It is a set of mutually exclusive rules; The score is triggered by the rule; This is an indicator function.

[0059] Calculate the final clause matching score. The system integrates semantic similarity with clause position / type weights to calculate the clause matching score.

[0060] ; in: This is a vector for case information retrieval; This is a fragment of a candidate clause; For semantic similarity; For the normalized location index of the clauses; For position attenuation parameters; Weight for the type of clause.

[0061] In step S4, the multi-source feature aggregation and ensemble learning discrimination step is the core step that transforms the quantitative indicators calculated by the preceding independent modules into the final business decision. This process is specifically implemented through the following sub-steps: Construct a multimodal fusion feature vector. The system defines a unified feature container to concatenate the heterogeneous scoring metrics generated in previous steps. Specifically, this feature vector... The system includes the following components: the customer's comprehensive risk score and its sub-items from step S1; the irrational medication risk score and the invoice / stamp compliance score from step S2; and the confidence level of liability determination and the degree of rule conflict from step S3. In addition, the system will also incorporate the basic discrete features of the claims case into the feature vector after one-hot encoding. This forms a high-dimensional dense numerical vector.

[0062] Perform risk probability prediction based on ensemble learning. The system will use the constructed feature vectors mentioned above. The input is fed into a pre-trained ensemble learning classifier. This classifier preferably employs a random forest or gradient boosting decision tree architecture. The model contains multiple base classifiers (i.e., decision trees), each independently outputting a prediction based on the input feature vector. The system performs a weighted average or voting on the outputs of all base classifiers, ultimately outputting a value normalized to the [0,1] interval; this value represents the overall risk probability of the case. The specific training process and parameter tuning of ensemble learning models can be implemented using existing machine learning frameworks by those skilled in the art, and will not be elaborated here.

[0063] Case triage decision-making based on dual thresholds. The system constructs a multimodal fusion feature vector and inputs it into an ensemble learning model to obtain the high-risk probability. Case risk classification ( The decision logic is as follows: ; Where δ is the automatic claims acceptance threshold. Tolerance for rule conflicts.

[0064] In step S4, in order to address the issue of unintuitive claims review criteria caused by the nonlinear integration model, the system executes a risk warning summary generation process based on marginal contribution.

[0065] In the above calculation logic, the first Risk warning value for each feature ( The calculation is as follows: ; in: For the risk probability function with respect to features The partial derivatives (or approximate gradients) of ; The standardized value for the feature.

[0066] System selection The most significant feature is the generation of natural language interpretations.

[0067] The marginal contribution value of each risk factor is quantified. Simple partial derivatives only reflect sensitivity and do not consider the numerical value of the feature itself. Therefore, the system further calculates the marginal contribution value of each feature. This calculation is performed by multiplying the partial derivative of the feature with its normalized feature value. The system sorts all the calculated feature marginal contribution values ​​and selects the top values ​​with positive values ​​and the largest absolute values. These characteristics were identified as the "dominant risk factors" that led to the current case being classified as high-risk.

[0068] Natural Language Mapping and Summary Assembly of Risk Factors. The system pre-configures a risk factor explanation template library, which defines a natural language description template for each feature dimension. The system iterates through the dominant risk factors selected in step S421, indexes the corresponding text templates based on their feature IDs, and dynamically fills the templates with specific feature values. For example, if the "institutional dispersion factor" is a dominant factor with a high score, the system calls the template to generate the text: "The number of medical institutions involved in recent medical visits is abnormal, and the dispersion score is a specific value." Finally, the system concatenates the generated explanation texts from highest to lowest contribution to form a concise risk warning summary, which is highlighted in the manual review interface to help reviewers quickly locate risk points.

[0069] In step S5, the intelligent quality inspection model aims to monitor the operational behavior patterns of human auditors, preventing abnormally rapid reviews or unauthorized approvals due to auditor fatigue or negligence. This process is not merely a review of the audit results, but an analysis of the behavioral trajectory of the audit process itself, specifically including the following steps: The system collects terminal interaction trajectory data streams. When a case is routed to the manual review interface, a lightweight monitoring service is initiated on the front end. This service captures the coordinate position and status of the input device at a preset sampling frequency. The collected data specifically includes a two-dimensional coordinate sequence of mouse movement. The characteristics of this trajectory include the vertical displacement of the scroll wheel, the trigger position of click events, and the keystroke interval of keyboard input. For mobile devices, these trajectory characteristics also include the pressure value of the touch screen and the swipe path. Those skilled in the art will understand that these raw interaction data constitute the physical layer characteristics describing the reviewer's behavioral patterns.

[0070] The system maps interaction trajectories to regions of interest (ROIs) on the page. It acquires the layout structure of the current review interface and identifies the coordinate range of key information display areas. These key areas are called ROIs, specifically including the generated risk warning summary display area, the medical document image display area, and the historical claims list area. The system performs spatial overlay analysis on the collected trajectory coordinate sequence and these ROIs, calculating the cursor or touchpoint dwell time and path coverage within each key area.

[0071] The system calculates a due diligence score for auditors. Based on trajectory data, it calculates multi-dimensional behavioral indicators to quantify the auditor's level of engagement. The system detects unnatural human-computer interaction characteristics, such as a cursor moving in an absolutely straight line or a page scrolling from top to bottom in an extremely short time. Subsequently, the system combines dwell time and coverage to calculate the final quality inspection score.

[0072] Calculate the suspicion level of manual decisions. The system combines case risk characteristics with the trajectory characteristics of the manual review process (such as dwell time and click path) to calculate the suspicion level of manual decisions. ): ; in: This represents the risk feature vector of the preceding steps; For manual review of trajectory feature vectors; Characteristics of the final ruling; This indicates vector concatenation.

[0073] Cases exceeding the threshold will be pushed to the quality inspection queue.

[0074] Implement real-time quality inspection and blocking strategies. The system will then calculate the audit due diligence score. The score is compared with a preset compliance threshold. If the score is below the threshold, the system determines the current operation as a low-quality review, temporarily freezes the submit button, and displays a mandatory reading prompt, requiring the reviewer to re-examine the risk summary or document image until the trajectory characteristics meet compliance requirements before the case can be processed. This mechanism transforms post-event accountability into in-process intervention, effectively ensuring the final execution quality of the human-machine collaborative risk control system.

[0075] In step S5, to ensure that the risk control model can continuously adapt to evolving fraud methods and prevent the "catastrophic forgetting" phenomenon of forgetting old fraud features due to only learning new data, the system implements an incremental update mechanism based on regularization constraints. This process specifically includes the following sub-steps: A high-value feedback dataset is constructed. The system tracks the final processing results of all cases diverted to manual review in real time. The system focuses on filtering out two types of samples: the first type is "missed reports," which are cases that the model judges as low-risk but are confirmed as violations by manual review; the second type is "false positives," which are cases that the model judges as high-risk but are confirmed as normal by manual review. The system assembles the feature vectors of these two types of samples with the manually determined true labels into an incremental training set. For data storage and cleaning, those skilled in the art can use distributed file systems and ETL tools, which will not be elaborated here.

[0076] Maintaining a core sample replay pool. To retain old knowledge when training on new data, the system does not directly discard historical data but maintains a fixed-capacity experience replay pool. This replay pool is not simply randomly sampled but maintained based on the information entropy or feature space distribution of the samples. The system periodically calculates the feature centers of historical samples, retaining the typical samples closest to the feature centers and the difficult samples near the decision boundary. When updating the model, the system mixes the collected incremental training set with the historical core samples in the replay pool according to a preset ratio to construct the current hybrid fine-tuning dataset.

[0077] The system performs parameter updates based on elastic weight consolidation. When updating model parameters via backpropagation using a mixed fine-tuning dataset, parameter importance constraints are introduced. The system calculates the importance of each parameter in the model to past tasks, approximating this importance using the diagonal elements of the Fisher information matrix. During optimization, the system modifies the loss function by adding a regularization penalty term, aiming to limit significant changes to parameters crucial to past tasks while allowing less important parameters to adapt freely to new data.

[0078] The above calculation logic is specifically manifested as follows: The loss function is updated using parameters based on Elastic Weight Consolidation (EWC). ): ; in: The loss function on the new data; These are the parameters for the old model; For the Fisher information matrix elements, the representative parameters are... The importance of old tasks; is the regularization coefficient.

[0079] Model shadow deployment and dynamic switching. The updated model does not directly replace the online model, but is first deployed to the shadow environment. The system sends real-time claims requests to both the old online model and the new shadow model, but only uses the output of the old model for business decisions. The system monitors the performance metrics of the new shadow model over a period of observation. Only when the new model achieves a preset standard in recognizing incremental samples while maintaining its historical accuracy will the system perform a blue-green deployment switch, officially pushing the new model into the production environment.

[0080] This application also provides a computer device, including: a processor and a memory, the memory storing a computer program executable by the processor, the computer program performing the method described above when executed by the processor.

Claims

1. A method for intelligent claims risk control in internet health insurance based on multi-source data fusion, characterized in that, Includes the following steps: S1: In response to a claim, based on a pre-built customer risk profile knowledge graph, calculate the applicant's historical behavioral risk score, recent medical treatment behavior risk score, and related network risk score, and generate a comprehensive customer risk score based on the aforementioned scores; S2: Obtain the image and text data of the claim materials associated with the claim application, and review the rationality of the medical behavior and the standardization of the invoice form based on the medical knowledge graph and the target detection model, and output the risk score of unreasonable drug use and the standardization score of the invoice seal; S3: Based on the retrieval enhancement generation framework, vectorized retrieval is used to obtain insurance clause fragments related to the claim application, and the clauses are matched with the case through a fine-tuned large language model, outputting the liability determination result, determination confidence and rule conflict degree; S4: Obtain the basic attribute data of the claim application, and concatenate the basic attribute data with the scores generated in step S1, the scores output in step S2, and the judgment indicators output in step S3 to construct a multi-dimensional structured feature vector. An ensemble learning classifier is used to determine the risk level of a claim case based on the multidimensional structured feature vector. The claim cases are then assigned to an automated claims process or a manual review process based on the risk level. A risk warning summary is generated for the claim cases assigned to the manual review process. S5: Based on the intelligent quality inspection model, back-scan the closed claims cases to identify claims cases with potential adjudication bias, obtain corrected case data for claims cases with potential adjudication bias, and use the corrected case data to incrementally update the customer risk profile knowledge graph, target detection model, fine-tuned large language model and ensemble learning classifier involved in the above steps.

2. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S1, calculating the applicant's historical behavioral risk score specifically includes: Using the applicant's unique identifier as an anchor point, the customer risk profile knowledge graph is traversed to extract a local risk subgraph containing historical claim event nodes; Traverse the local risk subgraph to obtain the payout amount attribute and case closure time attribute of the historical claims event node, and calculate the time interval between the case closure time attribute and the current claims application trigger time; The time decay weight of each historical claims event node is calculated using an exponential decay function, and the time decay weight decreases non-linearly as the time interval increases. The payout amount attribute of each historical claim event node is multiplied by the corresponding time decay weight and then summed to obtain the time-weighted total payout. The time-weighted total payout is then divided by the applicant's cumulative coverage period to obtain the historical behavior risk score.

3. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S1, calculating the applicant's recent medical treatment behavior risk score specifically includes: A short-term risk observation window is set based on the current claim application trigger time, and medical visit event nodes that fall within the short-term risk observation window are retrieved from the customer risk profile knowledge graph; The total number of visits is obtained by counting the number of the medical visit event nodes. The logarithm of the total number of visits plus one is calculated and used as the logarithmic smoothing value of the medical visit density. Extract the medical institution attributes associated with the medical visit event nodes, construct an institution visit distribution sequence, calculate the ratio of the standard deviation to the mean of the number of institutions in the institution visit distribution sequence, and add the coefficient of variation to the benchmark constant to obtain the medical institution dispersion factor; The recent medical visit behavior risk score is obtained by multiplying the log-smoothed value of the medical visit density with the medical institution dispersion factor.

4. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S2, the output of the irrational drug use risk score specifically includes: Diagnostic information, drug list, and item list are extracted from the claim material data. The drug list and item list are transformed into claim feature vectors. The standard treatment plan set corresponding to the diagnostic information is retrieved from the medical knowledge graph. The standard treatment plan set is transformed into a standard plan vector using graph embedding technology. Calculate the cosine similarity between the claim feature vector and the standard solution vector, and select the maximum cosine similarity value; Identify drugs or items in the claim feature vector that are not included in the standard treatment set, and calculate the proportion of the total cost of drugs or items not included in the standard treatment set to the total medical expenses. The proportion of the total cost to the total medical expenses is the abnormal drug cost ratio. The difference is obtained by subtracting the product of the abnormal drug cost ratio and the preset penalty coefficient from the maximum cosine similarity value, and then subtracting the difference from 1 to obtain the irrational drug use risk score.

5. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S2, the process of scoring the standardization of the output document seal is as follows: The target detection model is used to identify the type of seal and the corresponding detection confidence level in the claims material data; Based on the document type corresponding to the claim material data, the preset required seal combination rule library is called, and each required seal category in the required seal combination rule library is traversed. When the detection confidence level of the required seal category is greater than the preset detection threshold, the detection confidence level is retained; When the detection confidence of the required seal category is not greater than the preset detection threshold, the detection confidence is recorded as zero; Calculate the mean of the retention detection confidence scores for all the required seal categories, and use it as the standardization score for the document seal.

6. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S3, the step of using vectorized retrieval to obtain relevant insurance clause fragments specifically includes: Generate a case query vector and calculate the basic semantic similarity between the case query vector and a pre-constructed insurance clause fragment vector; A position correction factor is calculated based on the normalized position index of the insurance clause fragment in the document, and the position correction factor decreases exponentially as the normalized position index increases; Assign a type-weighted coefficient based on the business type of the insurance clause fragment; The basic semantic similarity, the position correction coefficient, and the type weighting coefficient are fused and calculated, and the insurance clause fragments are selected based on the calculation results.

7. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S3, the confidence level of the responsibility determination is obtained by generating the geometric mean of the key token probability distribution of the core conclusion words through the fine-tuned large language model; The calculation process of the rule conflict degree is as follows: using the fine-tuned large language model, multiple intermediate judgment results are generated for the same case based on different insurance clause fragments. The proportion frequency of the mainstream judgment conclusion in the intermediate judgment results is counted, and the difference between 1 and the proportion frequency is calculated to obtain the rule conflict degree.

8. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S4, the process of generating the risk warning summary is as follows: Construct a risk probability approximation function for the ensemble learning classifier, and calculate the partial derivatives of the risk probability approximation function with respect to each component of the multimodal fusion feature vector; Multiply the partial derivatives of each component by the normalized eigenvalues ​​to obtain the marginal contribution value of each feature. The top N features with the largest absolute values ​​of marginal contribution are selected as the dominant risk factors. Based on a pre-set interpretation template library, the feature values ​​of the dominant risk factors are mapped to natural language descriptions and then concatenated to generate the risk warning summary.

9. The intelligent claims risk control method for internet health insurance based on multi-source data fusion according to claim 1, characterized in that, In step S5, when performing incremental updates, a parameter update strategy based on elastic weight consolidation is adopted: Construct a hybrid fine-tuning dataset that includes the revised case data and historical core samples; During the process of updating model parameters, a regularization penalty term is added to the loss function. The regularization penalty term is calculated based on the Fisher information matrix and is used to constrain parameters that are highly important to historical tasks from undergoing significant changes. The loss function is expressed as the sum of the new task loss and the regularization penalty term, which is proportional to the product of the square of the parameter change and the corresponding Fisher information value.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent claims risk control method for Internet health insurance based on multi-source data fusion as described in any one of claims 1 to 9.

Images