A medical insurance cost abnormality tracing method based on multi-view clustering and causal inference

Abstract

The present application relates to a kind of medical insurance expense abnormality tracing method based on multi-view clustering and causal inference, belong to medical big data analysis and artificial intelligence technical field, for medical insurance settlement data extraction heterogeneous feature view set;Subsequently, using consensus learning algorithm based on adaptive weight, each view is mapped to low-dimensional consensus representation space;Then, by calculating the view conflict score and outlier distance of sample, screening suspected abnormal settlement sample.For suspected abnormal sample, construct structured causal model, use causal effect Quantification to strip the interference of mixed factors such as severity of illness;Subsequently, by intermediary effect analysis, the path mechanism of abnormality is extracted, and the subject of responsibility is determined based on counterfactual reasoning.According to the attribution result, generate structured evidence chain, and combine the audit feedback mechanism to dynamically update clustering weight and causal graph topology, realize the classification and hierarchical supervision of medical insurance violation behavior.

Description

Technical Field

[0001] This invention belongs to the field of medical big data analysis and artificial intelligence technology, and relates to a method for tracing the source of abnormal medical insurance expenses based on multi-view clustering and causal inference. Background Technology

[0002] The operational security of medical insurance funds is directly related to the stability of the social security system. With the expansion of medical insurance coverage and the explosive growth of settlement data, the supervision of medical insurance funds faces severe challenges. Traditional medical insurance audits mainly rely on expert rule bases or manual spot checks, which are insufficient to cover the massive and often concealed abnormal settlement behaviors. Therefore, utilizing big data and machine learning technologies for abnormal expense identification has become the mainstream direction for intelligent medical insurance supervision.

[0003] Existing medical insurance anomaly detection technologies are mainly based on single-dimensional statistical analysis or basic machine learning models. However, in practical applications, these technologies have the following significant shortcomings: First, the data features are not fully utilized, resulting in information silos. Medical insurance data naturally possesses multi-view characteristics, including patient profile views, treatment behavior views, drug and consumable views, and relational network views. Existing technologies mostly involve simply stitching together features from different dimensions and then performing dimensionality reduction. This approach ignores the heterogeneity and complementarity between different views, leading to low sensitivity of the model in identifying complex collaborative fraud, and making it prone to false positives or false negatives.

[0004] Second, the models are black-boxed and lack interpretability. Although models such as deep learning are highly accurate in identifying anomalies, they often only provide a judgment of whether something is abnormal or a probability score. In the actual scenario of medical insurance auditing, regulatory authorities not only need to know who is abnormal, but also why they are abnormal. Due to the lack of in-depth analysis of the internal logic of abnormal behavior, auditors find it difficult to quickly pinpoint the core of the violation when conducting subsequent on-site verification or administrative penalties, leading to difficulties in evidence collection and low enforcement efficiency.

[0005] Third, correlation does not equal causation, making accurate attribution of responsibility difficult. Existing anomaly analyses are often based on the correlation between features, failing to distinguish whether abnormal costs are a reasonable increase caused by the complexity of the objective condition or by subjective malicious fraud. Without a causal inference framework, models cannot eliminate confounding factors such as the severity of the condition, making it difficult to scientifically explain the mechanisms behind abnormal behavior, and even more difficult to accurately determine responsibility among doctors, patients, and medical institutions.

[0006] Therefore, how to effectively integrate heterogeneous features from multiple views to accurately identify hidden anomalies, and combine causal inference technology to logically trace the source and attribute responsibility for abnormal behavior, is a key technical problem that urgently needs to be solved in the field of intelligent medical insurance supervision. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method for tracing the source of abnormal medical insurance expenses based on multi-view clustering and causal inference. This method identifies suspected violation samples by constructing a multi-view feature space and uses causal inference to remove confounding factors, thereby achieving accurate classification and attribution of responsibility for abnormal medical insurance settlement behavior.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for tracing the sources of medical insurance expense anomalies based on multi-view clustering and causal inference includes the following steps: S1: Multi-view feature construction and suspected anomaly identification: Select medical insurance settlement details data to construct multiple heterogeneous feature views, use an autoencoder to perform dimensionality reduction processing to obtain low-dimensional representations of each view; then, use a multi-view consensus learning algorithm based on adaptive weights to map each view to a unified low-dimensional consensus space, and determine suspected abnormal settlement samples by calculating view conflict scores and cluster outlier distances. S2: Causal Source Analysis: A structured causal model is constructed for the suspected abnormal samples. Propensity score matching is used to remove the interference of confounding factors on the severity of the illness. The individual treatment effect of the treatment behavior on the cost deviation is calculated. Subsequently, the mechanism of violation is extracted by mediation effect analysis, and the responsible party is determined by counterfactual reasoning. S3: Closed-loop supervision and model evolution: Generate a structured chain of evidence based on the causal tracing results, and dynamically update the multi-view weights and causal graph topology in conjunction with the audit feedback mechanism to achieve adaptive model evolution and subject credit rating early warning.

[0009] Furthermore, the heterogeneous feature view in S1 specifically includes: Behavioral logic view: Extracting spatiotemporal features including diagnosis and surgery matching degree, clinical pathway deviation, frequency of cross-hospital visits, and density of examination items; Resource consumption view: Extracts numerical features including the proportion of drug costs, fluctuations in the unit price of high-value consumables, the proportion of out-of-pocket expenses, and the total amount paid by the medical insurance fund.

[0010] Furthermore, the data preprocessing in S1 specifically involves: using linear interpolation to handle missing values, removing outliers based on the 3σ criterion, and using Z-score standardization and Min-Max normalization to eliminate dimensional differences.

[0011] Furthermore, the autoencoder in S1 comprises a three-layer encoder and a three-layer decoder. The encoder uses the ReLU activation function, and the decoder uses the Sigmoid activation function. The low-dimensional representation of the view is obtained by minimizing the reconstruction loss function. .

[0012] Furthermore, the multi-view consensus learning algorithm in S1 learns the consensus representation matrix Z by minimizing the following objective function.

[0013] in, Represents the projection matrix. For view to adapt weights, Let be the graph Laplace matrix.

[0014] Furthermore, the determination of suspected abnormal samples in S1 specifically includes: Calculate view conflict score ; Calculate the cluster outlier distance of the samples in the consensus space Z. ; Calculate the comprehensive anomaly score ,like Exceeding the dynamic threshold It is then judged as a suspected anomaly.

[0015] Furthermore, the confounding factors regarding the severity of the disease in S2 specifically include: using the severity of the disease as a covariate. A counterfactual control group was constructed using propensity score matching, and the individual treatment effect was calculated. ,in For medical treatment, This is for cost deviation.

[0016] Furthermore, the extraction of the violation path mechanism in S2 specifically includes: using mediation analysis to decompose the total causal effect into direct effects and indirect effects generated through mediating variables, including the density of auxiliary examinations, the amount of duplicate prescriptions, and the length of hospital stay.

[0017] Furthermore, the adaptive evolution of the model in S3 specifically includes: collecting feedback from audit experts on the evidence chain, mapping it to a loss function feedback term, and using gradient descent to update the view weight coefficients in S1 online. :

[0018] Furthermore, the tiered early warning and penalty system in S3 specifically includes: updating the credit profile of the subject based on the responsibility attribution results, and triggering real-time interception instructions from the medical insurance settlement system for high-risk subjects.

[0019] The beneficial effects of this invention are as follows: 1) Fully explore the heterogeneous features of multi-view medical insurance data, and integrate the information of each view through an adaptive weighted multi-view consensus learning algorithm. This solves the "information silo" problem of traditional technology, improves the sensitivity of identifying collaborative fraud and hidden abnormal settlement behavior, and reduces the false alarm and false alarm rates.

[0020] 2) Combining a causal inference framework to achieve interpretability analysis of medical insurance anomalies can not only determine "whether it is abnormal", but also accurately analyze the path of abnormal expenses, remove the interference of objective confounding factors such as the severity of the illness, solve the problem of "black box" in traditional models, and provide a clear and verifiable chain of evidence for medical insurance audits.

[0021] 3) Accurate attribution of responsibility based on counterfactual reasoning can clarify the responsibility for violations among medical institutions, doctors and other entities, distinguish between reasonable increases in medical expenses related to the illness and expenses incurred due to subjective malice, provide a scientific basis for determining medical insurance administrative penalties and improve law enforcement efficiency.

[0022] 4) Construct a closed-loop supervision and model evolution mechanism that integrates human and machine collaboration. By dynamically updating model parameters and causal structure through feedback from audit experts, the model can adapt to new medical insurance violations. At the same time, implement credit-based early warning for entities to achieve precise and dynamic supervision of medical insurance funds, reduce losses of medical insurance funds from the source, and improve the intelligence and refinement of medical insurance fund supervision.

[0023] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for tracing the source of abnormal medical insurance expenses based on multi-view clustering and causal inference; Figure 2 A flowchart illustrating the structure for screening suspected anomaly samples based on multi-view clustering; Figure 3 Generate a path topology map for abnormal costs based on mediation effect analysis. Detailed Implementation

[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0026] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0027] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0028] Example 1: The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference provided by this invention includes the following steps: S1, Multi-view Feature Construction and Suspected Anomaly Identification: Extracting a heterogeneous feature view set from medical insurance settlement data. Subsequently, a consensus learning algorithm based on adaptive weights is used to map each view to a low-dimensional consensus representation space Z; then, suspected abnormal settlement samples are screened by calculating the view conflict score and outlier distance of the samples.

[0029] S2, Mechanism Analysis and Responsibility Attribution Based on a Causal Framework. A structured causal model is constructed for suspected anomalous samples, and causal effects are used to quantify and eliminate confounding factors such as disease severity. Subsequently, mediation effect analysis is used to extract the path mechanism of the anomaly, and counterfactual reasoning is used to determine the responsible party for the violation.

[0030] S3, Closed-Loop Supervision and Model Evolution. A structured evidence chain is generated based on the attribution results, and the clustering weights and causal graph topology are dynamically updated in conjunction with an audit feedback mechanism to achieve classified and graded supervision of medical insurance violations.

[0031] Preferably, step S1, the construction of multi-view features and identification of suspected anomalies, specifically includes the following sub-steps: 1) Medical Insurance Settlement Data Collection and Preprocessing: Raw data containing patient basic information, diagnosis codes, surgical procedures, drug and consumable details, and settlement amounts are collected from the medical insurance settlement center and hospital information system (HIS); missing values ​​are filled using linear interpolation, extreme outliers are removed based on the 3σ criterion, and dimensional differences are eliminated through Z-score standardization. The formula is:

[0032] By combining Min-Max normalization to scale the features to the [0,1] interval, the data cleaning and standardization processes are completed.

[0033] 2) Construction of heterogeneous feature views: The preprocessed data is divided into two heterogeneous feature views according to business logic, forming a feature set. Among them, the behavioral logic view Extract spatiotemporal features such as diagnosis and surgery matching degree, clinical pathway deviation, frequency of cross-hospital visits, and density of examination items to generate a resource consumption view. The data includes numerical characteristics such as the proportion of drug costs, fluctuations in the unit price of high-value consumables, the proportion of out-of-pocket expenses, and the total amount paid by the medical insurance fund.

[0034] 3) Feature Dimensionality Reduction and Denoising Based on Autoencoders: Constructing a Three-Layer Autoencoder (Encoder: ReLU activation; Decoder: The sigmoid activation function is used to reduce the dimensionality of each heterogeneous feature view, and the encoding function is used to perform dimensionality reduction. Mapping the original view data to the embedding space yields a low-dimensional feature representation. By minimizing reconstruction loss The model was optimized to ensure that more than 95% of the original diagnosis and treatment information was retained and random noise was filtered out.

[0035] 4) Multi-view consensus space learning based on adaptive weights: Construct an objective function and minimize the following loss function through joint matrix factorization:

[0036] Automatically assign adaptive weights to each view. To suppress the interference of noisy views on decision-making, the projection matrix is ​​obtained through iterative calculation. And the consensus representation matrix Z.

[0037] 5) Multi-dimensional suspected anomaly judgment: from visual... Figure 1 Anomaly indices are calculated using two dimensions: consistency and spatial outlier. First, a view conflict score, reflecting the degree of deviation in the performance of the sample across different views, is calculated. Then calculate the cluster outlier distance of the sample to the nearest cluster center in the consensus space Z. The comprehensive anomaly score is obtained through weighted calculation. ,like Exceeding the preset dynamic threshold If so, the sample is marked as a suspected abnormal settlement sample.

[0038] Preferably, step S2, the mechanism analysis and responsibility attribution based on the causal framework, specifically includes the following sub-steps: 1) Construction of Structured Causal Model (SCM): Combining prior medical knowledge with historical audit data, a directed acyclic graph (DAG) is constructed for suspected anomalous samples; key diagnostic and treatment behaviors are defined as processing variables. Settlement fee deviation or fund expenditure exceeding the limit are the outcome variables. The severity of the patient's condition (comorbidity index, age, past medical history, etc.) is used as a covariate. Establish causal dependency , Isolating confounding factors from causal pathways Interference.

[0039] 2) Propensity score matching (PSM) to remove confounding factors: Logistic regression is used to calculate the propensity score of each sample for a given condition. The propensity score for accepting medical treatment (T) was used; a control group sample was matched from the normal settlement sample database with the propensity score closest to that of the suspected abnormal sample to ensure statistical consistency of the severity of the disease between the two groups; the individual treatment effect was calculated through the matched samples. This represents the net increase in costs resulting from medical treatment after removing the interference of the patient's condition.

[0040] 3) Mediation effect analysis to analyze abnormal paths: The density of auxiliary examinations, the amount of duplicate prescriptions, and the length of hospital stay are introduced as mediating variables M. The total causal effect is decomposed into the direct effect (DE) that directly leads to abnormal costs due to medical behavior and the indirect effect (IE) that indirectly leads to abnormal costs through mediating variables. By minimizing the regression residuals to optimize the path coefficients, key violation paths with a contribution rate greater than 60% are identified, and the specific mechanisms of abnormal cost generation are clarified.

[0041] 4) Counterfactual reasoning to achieve responsibility attribution: Calculating the counterfactual conditional probability The simulation explores a counterfactual scenario: "Would the cost still exceed the limit if the treatment was not performed?" Based on the mediation path and the counterfactual simulation results, it determines the medical institution's responsibility (such as splitting hospitalizations or using inflated diagnostic codes) and the doctor's subjective responsibility (over-treatment deviating from the clinical pathway). If the value is close to 0, it is considered a "pseudo-abnormality" caused by the complexity of the condition and is removed.

[0042] Preferably, in step S3, the closed-loop supervision and model adaptive evolution specifically include the following sub-steps: 1) Automatic generation of structured audit evidence chains: integrating the view conflict characteristics of S1 and the individual processing effects of S2. Combined with counterfactual path diagrams and other multi-dimensional data, a standardized audit report is generated. The report includes a description of abnormal behavior, causal path analysis, attribution of responsibility, and recommended penalty scores, forming a logically closed-loop audit evidence chain. An example of an audit report is shown in Table 1.

[0043] Table 1

[0044] 2) Human-machine collaborative audit feedback collection: The audit report is pushed to the medical insurance audit expert's end, and the expert's feedback on "whether it is confirmed as a violation" and "whether the attribution logic is accurate" is collected. The annotation results are divided into three categories: complete match, accurate identification but incorrect attribution, and false alarm. The annotation results are then mapped to loss function feedback terms. .

[0045] 3) Dynamic adaptive update of model parameters: Using gradient descent, based on the loss function feedback term... Online update of view adaptive weights in S1 :

[0046] Improve the model's sensitivity to new and covert violations; if experts determine that the attribution logic is incorrect, automatically adjust the cause-effect graph (DAG). The causal strength coefficient of the path, or the introduction of new confounding variables. Optimize the accuracy of causal inference.

[0047] 4) Credit rating and precise supervision of entities: Based on the attribution of responsibility and historical data on violations, update the credit profiles of regulatory entities such as medical institutions and doctors, and classify entities into three levels: high-risk, medium-risk, and low-risk, according to the frequency of violations, the cumulative amount of violations, and the degree of subjective malice. Real-time interception instructions are triggered in the medical insurance settlement system for high-risk entities, key audits are implemented for medium-risk entities, and routine monitoring is carried out for low-risk entities, thereby upgrading the supervision of medical insurance funds from "post-event audit" to "in-event control".

[0048] Example 2: This embodiment further illustrates the invention in conjunction with the accompanying drawings, such as... Figure 1 As shown, the method for tracing the source of abnormal medical insurance expenses based on multi-view clustering and causal inference specifically includes the following steps: S1: Multi-view Feature Construction and Suspected Anomaly Identification S11: Data Acquisition and Preprocessing Step 1: Collect 50,000 sets of inpatient settlement details from a municipal medical insurance settlement center.

[0049] Step 2: Extract 24 original features including patient age, ICD-10 diagnostic code, surgical procedure code, drug cost details, examination item density, and length of hospital stay.

[0050] Step 3: Data cleaning. Missing values ​​are handled using linear interpolation, outliers (such as negative costs or logically incorrect hospital stays) are removed based on the 3σ criterion, and noise is reduced using moving average filtering.

[0051] Step 4: Standardization. Z-score standardization is used to eliminate dimensional differences between different features. The calculation formula is: The Min-Max method is used to scale the features to the [0,1] interval.

[0052] S12: Multi-view Feature Construction and Preliminary Dimensionality Reduction Step 1: Construct a heterogeneous feature view set

[0053] See Figure 1 (Behavioral Logic View) ): Focuses on clinical diagnosis and treatment logic, including the matching degree between diagnosis and medication, and the deviation of clinical pathways; See Figure 2 (Resource Consumption View) ( ): Focuses on economic characteristics, including the proportion of drugs, fluctuations in high-value consumables, and total amount.

[0054] Step 2: Perform preliminary dimensionality reduction and noise reduction on the features using an autoencoder. Construct a three-layer autoencoder (encoder: 16-8-4, ReLU activation; decoder: 4-8-16, Sigmoid activation). Optimize by minimizing the reconstruction loss, calculated as follows:

[0055] in This represents the reconstructed data. The autoencoder is trained 1000 times with a learning rate of 0.001. This yields a low-dimensional embedding feature representation. .

[0056] S13: Multi-view Consensus Learning and Anomaly Detection Step 1: Use consensus learning algorithms to... and Mapped to the consensus representation matrix Z. View weights are automatically assigned through iterative optimization of the following objective function. :

[0057] The weight smoothing coefficient γ is set to 2, and the graph regularization term λ is set to 0.1.

[0058] Step 2: Calculate the anomaly detection index.

[0059] View Conflict Rating This reflects the degree of deviation between the diagnostic logic view and the economic view of the sample.

[0060] outlier distance : The Euclidean distance of a sample from the consensus space to the cluster center.

[0061] Step 3: Identify suspected abnormal samples. For example... Figure 2 As shown, the comprehensive anomaly score is calculated. Set dynamic threshold This represents the 95th percentile of the overall sample distribution. 12.50 potentially abnormal settlement statements were identified.

[0062] S2: Responsibility Attribution Analysis Based on a Causal Framework Step 1: For suspected anomalous samples, use "key diagnostic and treatment behaviors" as the processing variable. "Settlement amount" is the result variable "Severity of illness" as a covariate Construct a cause-effect graph.

[0063] Step 2: Use propensity score matching (PSM) to isolate disease factors.

[0064] Through the Perform a balanced matching and calculate the individual treatment effect. :

[0065] like If the value is significantly greater than 0 and passes the significance test (p<0.05), then the sample is suspected of subjective violation.

[0066] Step 3: Analyzing the mechanism of the mediation effect. For example... Figure 3 As shown, a mediating variable M (e.g., auxiliary examination density) is introduced. The proportion of indirect effects is calculated.

[0067] For example, a sample incurred 3,000 yuan in overspending due to hospitalization for a "cold". Analysis revealed that 80% of the cost was generated through the intermediary path of "repeated CT scans", thus identifying the violation mechanism.

[0068] Step 4: Determining liability.

[0069] Medical institution responsibility: The path points to the breakdown of hospitalization, bed-occupancy fraud, etc.

[0070] Physician responsibility: The pathway points to overtreatment that deviates from the clinical pathway.

[0071] S3: Closed-Loop Supervision and Model Update Step 1: Automated generation of evidence chain report.

[0072] The report includes: anomaly scores, causal path topology, and an estimated amount of pure violation after removing disease-related conditions.

[0073] Step 2: Closed-loop feedback.

[0074] Upon receiving feedback from auditors, if the feedback is marked as "Accurate Identification," the model parameters are maintained; if it is marked as "False Alarm," the view weights are updated via feedback gradients. .

[0075] Step 3: Hierarchical management.

[0076] The credit score of the entity is updated based on the attribution results. Real-time medical insurance settlement is blocked for high-risk entities.

[0077] Example 3: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0078] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0079] Example 5: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0080] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0081] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0082] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0083] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0084] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0085] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0086] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0087] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for tracing the source of abnormal medical insurance expenses based on multi-view clustering and causal inference, characterized in that: Includes the following steps: S1: Multi-view feature construction and suspected anomaly identification: Select medical insurance settlement details data to construct multiple heterogeneous feature views, and use an autoencoder to perform dimensionality reduction processing to obtain low-dimensional representations of each view; Subsequently, using a multi-view consensus learning algorithm based on adaptive weights, each view is mapped to a unified low-dimensional consensus space, and suspected abnormal settlement samples are identified by calculating view conflict scores and cluster outlier distances. S2: Causal source analysis: A structured causal model is constructed for the suspected abnormal samples. Propensity score matching is used to remove the interference of confounding factors on the severity of the illness, and the individual treatment effect of the treatment behavior on the cost deviation is calculated. Subsequently, the mechanism of violation was extracted by analyzing the mediation effect, and the responsible party was determined by counterfactual reasoning. S3: Closed-loop supervision and model evolution: Generate a structured chain of evidence based on the causal tracing results, and dynamically update the multi-view weights and causal graph topology in conjunction with the audit feedback mechanism to achieve adaptive model evolution and subject credit rating early warning.

2. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The heterogeneous feature view in S1 specifically includes: Behavioral logic view: Extracting spatiotemporal features including diagnosis and surgery matching degree, clinical pathway deviation, frequency of cross-hospital visits, and density of examination items; Resource consumption view: Extracts numerical features including the proportion of drug costs, fluctuations in the unit price of high-value consumables, the proportion of out-of-pocket expenses, and the total amount paid by the medical insurance fund.

3. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The data preprocessing in S1 specifically involves: using linear interpolation to handle missing values, removing outliers based on the 3σ criterion, and using Z-score standardization and Min-Max normalization to eliminate dimensional differences.

4. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The autoencoder in S1 comprises a three-layer encoder and a three-layer decoder. The encoder uses the ReLU activation function, and the decoder uses the Sigmoid activation function. The low-dimensional representation of the view is obtained by minimizing the reconstruction loss function. .

5. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The multi-view consensus learning algorithm in S1 learns the consensus representation matrix Z by minimizing the following objective function. in, Represents the projection matrix. For view to adapt weights, Let be the graph Laplace matrix.

6. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The determination of suspected abnormal samples in S1 specifically includes: Calculate view conflict score ; Calculate the cluster outlier distance of the samples in the consensus space Z. ; Calculate the comprehensive anomaly score ,like Exceeding the dynamic threshold It is then judged as a suspected anomaly.

7. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The confounding factors for the severity of the disease identified in S2 specifically include: using disease severity indicators as covariates. A counterfactual control group was constructed using propensity score matching, and the individual treatment effect was calculated. ,in For medical treatment, This is for cost deviation.

8. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The extraction of the violation path mechanism in S2 specifically includes: using mediation analysis to decompose the total causal effect into direct effects and indirect effects generated through mediating variables, including the density of auxiliary examinations, the amount of duplicate prescriptions, and the length of hospital stay.

9. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The adaptive evolution of the model in S3 specifically includes: collecting feedback from audit experts on the evidence chain, mapping it to a loss function feedback term, and using gradient descent to update the view weight coefficients in S1 online. : 。 10. The method for tracing the source of medical insurance expense anomalies based on multi-view clustering and causal inference according to claim 1, characterized in that: The tiered early warning and penalty system in S3 specifically includes: updating the credit profile of the subject based on the responsibility attribution results, and triggering real-time interception instructions from the medical insurance settlement system for high-risk subjects.

Images