Medical insurance fraud detection method, device, computer equipment and storage medium

Abstract

This application relates to a medical insurance fraud detection method, apparatus, computer device, and storage medium, wherein the method comprises: obtaining heterogeneous medical data, performing standardized encoding and feature fusion on the heterogeneous medical data to generate a high-dimensional fused feature sequence; extracting time series features from the high-dimensional fused feature sequence based on claim time intervals using a long short-term memory network to generate a time series feature sequence; aggregating features of the time series feature sequence using a dynamic attention mechanism to generate a fixed-dimensional fraud representation vector; performing feature extraction processing on the fraud representation vector using a fully connected network to generate a high-density feature vector; generating a prediction result based on the high-density feature vector using a classifier, and conformally calibrating the prediction result to generate a fraud prediction set that meets a preset confidence level. This application improves the accuracy of medical fraud detection.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a medical insurance fraud detection method, apparatus, computer equipment, and storage medium. Background Art

[0002] With the rapid development of medical insurance in recent years, the amount of funds has expanded dramatically. Effectively managing these funds has become a key challenge. Furthermore, the surge in the number of insured persons and the increasing diversity of diagnosis and treatment models and medical programs have led to a proliferation of business relationships and a more complex medical insurance process. Therefore, it is necessary to focus on the act of fraud and attempt to identify it using advanced technologies.

[0003] To identify medical insurance fraud, we must first clarify the different types of fraud. The main entities involved in medical insurance are the insured, medical providers, and third-party payers. Based on the participating parties, fraud can be primarily categorized as single-entity fraud and multi-party fraud. However, fraud by insured individuals is typically small-scale and sudden, with a relatively limited impact on the medical insurance system. In contrast, fraud by medical providers is often more subtle and large-scale. However, most current research focuses on capturing fraudulent behavior by insured individuals, including identity theft, fictitious receipts, and drug abuse. Relatively little research has been conducted on fraud detection by medical providers, resulting in incomplete and inaccurate medical insurance fraud detection. Summary of the Invention

[0004] The purpose of the embodiments of the present application is to provide a medical insurance fraud detection method, apparatus, computer equipment and storage medium to improve the accuracy of medical insurance fraud detection.

[0005] To solve the above technical problems, the present invention provides a method for detecting medical insurance fraud, including:

[0006] Acquire heterogeneous medical data, and perform standardized encoding and feature fusion on the heterogeneous medical data to generate a high-dimensional fusion feature sequence;

[0007] Extracting time series features from the high-dimensional fusion feature sequence according to the claim time interval using a long short-term memory network to generate a time series feature sequence;

[0008] A dynamic attention mechanism is used to aggregate features of the temporal feature sequence to generate a fraud representation vector of fixed dimension;

[0009] Performing feature extraction processing on the fraud representation vector through a fully connected network to generate a high-density feature vector;

[0010] A prediction result is generated based on the high-density feature vector by a classifier, and the prediction result is conformally calibrated to generate a fraud prediction set that meets a preset confidence level.

[0011] To solve the above technical problems, the present invention provides a medical insurance fraud detection device, comprising:

[0012] A feature fusion module is used to obtain heterogeneous medical data, perform standardized encoding and feature fusion on the heterogeneous medical data, and generate a high-dimensional fusion feature sequence;

[0013] A time series feature extraction module is used to extract time series features from the high-dimensional fusion feature sequence according to the claim time interval through a long short-term memory network to generate a time series feature sequence;

[0014] A dynamic attention aggregation module, configured to perform feature aggregation on the temporal feature sequence using a dynamic attention mechanism to generate a fixed-dimensional fraud representation vector;

[0015] a feature refinement module, configured to perform feature refinement processing on the fraud representation vector through a fully connected network to generate a high-density feature vector;

[0016] A conformal calibration module is used to generate a prediction result based on the high-density feature vector through a classifier, and to conformally calibrate the prediction result to generate a fraud prediction set that meets a preset confidence level.

[0017] To solve the above technical problems, a technical solution adopted by the present invention is: providing a computer device, including one or more processors; a memory for storing one or more programs, so that the one or more processors can implement any one of the above-mentioned medical insurance fraud detection methods.

[0018] To solve the above technical problems, the present invention adopts a technical solution: a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, any one of the above-mentioned medical insurance fraud detection methods is implemented.

[0019] Embodiments of the present invention provide a medical insurance fraud detection method, apparatus, computer device, and storage medium. The method includes: obtaining heterogeneous medical data, performing standardized encoding and feature fusion on the heterogeneous medical data to generate a high-dimensional fused feature sequence; extracting time series features from the high-dimensional fused feature sequence based on the claim time interval using a long short-term memory network to generate a time series feature sequence; aggregating features on the time series feature sequence using a dynamic attention mechanism to generate a fixed-dimensional fraud representation vector; performing feature extraction on the fraud representation vector using a fully connected network to generate a high-density feature vector; generating a prediction result based on the high-density feature vector using a classifier, and conformally calibrating the prediction result to generate a fraud prediction set that meets a preset confidence level. The embodiment of the present invention performs feature encoding on heterogeneous medical data, extracts features based on the claim time interval using a long short-term memory network, and uses conformally calibrated prediction calibration to perform prediction calibration. This allows the present application to detect medical fraud for medical providers, improves the comprehensiveness of detection, and captures the time series information of claims, which is beneficial for improving the accuracy of medical fraud detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the solutions in this application, a brief introduction will be given below to the drawings required for use in the description of the embodiments of this application. Obviously, the drawings described below are some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0021] Figure 1 This is a flowchart of the implementation process of the medical insurance fraud detection method provided in the embodiment of the present application;

[0022] Figure 2 This is a flowchart for implementing the first sub-process in the medical insurance fraud detection method provided in an embodiment of the present application;

[0023] Figure 3 This is a flowchart for implementing the second sub-process in the medical insurance fraud detection method provided in an embodiment of the present application;

[0024] Figure 4 This is a flowchart for implementing the third sub-process in the medical insurance fraud detection method provided in an embodiment of the present application;

[0025] Figure 5 This is a flowchart for implementing the fourth sub-process in the medical insurance fraud detection method provided in an embodiment of the present application;

[0026] Figure 6 This is a flowchart for implementing the fifth sub-process in the medical insurance fraud detection method provided in an embodiment of the present application;

[0027] Figure 7 is a schematic diagram of a medical insurance fraud detection device provided in an embodiment of the present application;

[0028] Figure 8 It is a schematic diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0030] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0031] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0032] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the medical insurance fraud detection method provided in the embodiment of the present application is generally executed by a server, and accordingly, the medical insurance fraud detection device is generally configured in the server.

[0034] See also Figure 1 , Figure 1 A specific implementation of a medical insurance fraud detection method is shown.

[0035] It should be noted that the method of the present invention is not limited to the method of Figure 1 The process sequence shown is limited to the following steps:

[0036] S1: Obtain heterogeneous medical data, perform standardized encoding and feature fusion on the heterogeneous medical data, and generate a high-dimensional fusion feature sequence.

[0037] Specifically, in the embodiments of the present application, a coding-based feature fusion layer is used to perform standardized coding and feature fusion on heterogeneous medical data, generating a high-dimensional fused feature sequence to achieve standardized expression of heterogeneous medical data. Heterogeneous medical data refers to a collection of data of various types, formats, or structures that exists in the medical field. These data originate from different medical systems, equipment, or departments and have diverse forms and characteristics. For example, heterogeneous medical data can be tabular data (patient ID, age, diagnosis code, medication code, etc.) in electronic health records (EHRs), laboratory test results (numeric data such as blood sugar levels and blood cell counts), medical reports in XML or JSON format, medical images, and electronic medical records.

[0038] See also Figure 2 , Figure 2 A specific implementation of step S1 is shown, which is described in detail as follows:

[0039] S11: Acquire the heterogeneous medical data;

[0040] S12: Mapping the discrete categorical data in the heterogeneous medical data into a continuous vector of fixed dimension through an embedding layer to obtain a first vector;

[0041] S13: extracting semantic embedding from the text data in the heterogeneous medical data to generate a second vector;

[0042] S14: performing normalization processing on the line-type numerical data in the heterogeneous medical data to generate a third vector;

[0043] S15: performing feature concatenation on the first vector, the second vector, and the third vector, and performing nonlinear fusion on the concatenated features to generate the high-dimensional fusion feature sequence.

[0044] Specifically, heterogeneous medical data is acquired, which is divided into different types of data, including discrete categorical features (such as diagnosis codes and drug codes), text features (such as case descriptions), and numerical features (such as costs and time intervals). An embedding layer is used to map the discrete categorical data (such as diagnosis codes and drug codes) into a continuous vector of fixed dimension (such as 128 dimensions) in this embodiment of the present application, generating a first vector. Semantic embedding is extracted from text features (such as medical record descriptions) using pre-trained word vectors (such as Word2Vec) or the BERT model to generate a second vector. Numerical feature normalization: Continuous numerical features (such as costs and time intervals) are normalized (Z-score) to generate a third vector. The first, second, and third vectors are concatenated along the feature dimensions to form a unified high-dimensional feature vector. Finally, the concatenated features are nonlinearly fused using a multi-layer fully connected network (such as a two-layer 512-dimensional network with a ReLU activation function). The normalized feature matrix is ​​output, resulting in a high-dimensional fused feature sequence.

[0045] S2: extracting time series features from the high-dimensional fusion feature sequence according to the claim time interval through a long short-term memory network to generate a time series feature sequence.

[0046] Specifically, existing research in the field of medical insurance fraud pays little attention to the sequence characteristics of claims. Even if there is, the characteristics contained in the sequence are basically treated as a covariate. However, such a method cannot directly embed the time characteristics of the claim sequence into the model for learning, resulting in an inability to extract features from the sequence well. When faced with sequence data with dynamic time dependencies, traditional machine learning methods can usually only capture static features. For long sequences, long short-term memory networks usually have better processing effects. Since the embodiments of the present application are aimed at long sequence data, most of which are more than 500 in length, the embodiments of the present application use long short-term memory networks for feature extraction. In the embodiments of the present application, the structure of the long short-term memory network is improved so that it can capture and selectively forget the time intervals in long sequences.

[0047] See also Figure 3 , Figure 3 A specific implementation of step S2 is shown, which is described in detail as follows:

[0048] S21: using the high-dimensional fusion feature sequence and the claim time interval as inputs of the long short-term memory network;

[0049] S22: Calculating a forget gate based on the high-dimensional fusion feature sequence and the claim time interval using a first preset formula;

[0050] S23: Dynamically adjust the information amount of the candidate memory unit according to the claim time interval, calculate the output hidden state of each time step, and splice the output hidden states to generate a time series feature sequence.

[0051] In the embodiment of the present application, the claim time interval (e.g., the number of days between two claims) is introduced into the input high-dimensional fusion feature sequence as the input of the long short-term memory network. The forget gate is calculated by introducing the claim time interval. The first preset formula (1) for its calculation is:

[0052] (1);

[0053] in, is the forget gate, 、 、 are randomly initialized weights, is the bias generated by random initialization, is the high-dimensional fusion feature sequence, for the claim time interval, 、 For sequence 、 The moment of claim.

[0054] The input gate is calculated by formula (2), the memory unit is calculated by formula (3) and formula (4), the output gate is calculated by formula (5), the hidden state is calculated by formula (6), and the final output is calculated by formula (7). Among them, formula (2)-formula (7) are as follows:

[0055]

[0056] in, is the input gate, candidate memory unit, memory unit, output gate, hidden unit and final output, are randomly initialized weights, It is the bias generated by random initialization. In an embodiment of the present application, the improved forget gate can capture information at the timestamp level. For long sequences, the forget gate can determine whether the interval between timestamps can be remembered during training. The function of the input gate is to decide whether to ignore the input data. The function of the output gate is to decide whether to use the hidden state. The hidden state and memory unit enter the next long short-term memory network for calculation of the next time step, which is the output of each time step collected. In an embodiment of the present application, the amount of information of the candidate memory unit is dynamically adjusted according to the claim time interval, and the output hidden state of each time step is calculated, and the output hidden state is spliced ​​to generate a time series feature sequence. The improved long short-term memory network can not only perform feature processing on long sequences, but also capture the interval characteristics of sequence timestamps, which is conducive to improving the accuracy of feature extraction and thus improving the accuracy of fraud detection.

[0057] S3: A dynamic attention mechanism is used to aggregate the temporal feature sequence to generate a fraud representation vector of fixed dimension.

[0058] Specifically, each sample consists of several claim records, meaning that each sample input to the model has a different shape, which the model cannot handle. Therefore, the data needs to be aggregated based on a single sample dimension. In this embodiment of the present application, a dynamic attention mechanism is used to aggregate features of temporal feature sequences to generate a fixed-dimensional fraud representation vector. This allows for the generation of a weighted distribution for variable-length sequences and compression into a fixed-dimensional feature vector.

[0059] See also Figure 4 , Figure 4 A specific implementation of step S3 is shown, which is described in detail as follows:

[0060] S31: Mapping the aggregated target into a query vector through a fully connected layer, and constructing a key-value pair from the temporal feature sequence;

[0061] S32: Calculate a score for each value based on the query vector and the key-value pair using a scaled dot product function to obtain an initial score, and normalize the initial score to obtain a normalized score;

[0062] S33: Performing a weighted summation on the normalized score and the input value to perform feature aggregation on the time series feature sequence to generate the fraud representation vector of fixed dimension.

[0063] Specifically, three concepts are introduced in the attention mechanism: the random clue is called query (query, q), and each input is a key-value pair of a value (value, v) and a non-random clue (key, k). That is, in the embodiment of the present application, the aggregation target (such as the provider ID) is mapped to a query vector through a fully connected layer, and the time series feature sequence is constructed into a key-value pair. The score of each value is calculated based on the query vector and the key-value pair by adopting a scaled dot product function to achieve a targeted query of the value for each query vector, obtain an initial score, and normalize the initial score to obtain a normalized score. Finally, the normalized score and the input value are weightedly summed to perform feature aggregation on the time series feature sequence to generate the fraud representation vector of fixed dimension. Among them, the formulas corresponding to the scaled dot product function, normalization and weighted summation are formula (8), formula (9) and formula (10), respectively, as follows:

[0064]

[0065] in, is an introduced scaling factor. Because the attention mechanism focuses on the importance of the sequence, the embodiments of this application provide a certain degree of interpretability for the neural network. During the calculation process, the attention weight of each time step in the sequence is obtained, and this attention weight is used to locate the subsequent claim. The attention weight can reflect the importance of the position, which is an important manifestation of the interpretability of the attention mechanism.

[0066] S4: Perform feature extraction on the fraud representation vector through a fully connected network to generate a high-density feature vector.

[0067] Specifically, the fraud representation vector is subjected to feature extraction processing through a fully connected network to mine fraud pattern representations at the provider level and generate a high-density feature vector.

[0068] See also Figure 5 , Figure 5 A specific implementation of step S4 is shown, which is described in detail as follows:

[0069] S41: Performing a multi-layer nonlinear transformation on the fraud representation vector through the fully connected network to obtain a transformed feature vector;

[0070] S42: performing feature compression on the changed feature vector to generate an initial high-density feature vector;

[0071] S43: If static features exist, extract the static features, and concatenate the extracted static features with the dynamic features in the initial high-density feature vector to generate the high-density feature vector.

[0072] Specifically, a fully connected network performs multi-layer nonlinear transformations on the fraud representation vector. In one specific embodiment, the first fully connected layer (512 dimensions, ReLU activation function) is followed by batch normalization (BatchNorm) and dropout (with a probability of 0.3); the second fully connected layer (256 dimensions, ReLU activation function) is followed by batch normalization; and the final fully connected layer (128 dimensions, Tanh activation function) compresses the feature dimensions and outputs a high-density representation. If other static features exist (such as provider history), they are concatenated with the refined dynamic features to form the final feature vector, resulting in a high-density feature vector.

[0073] S5: Generate a prediction result based on the high-density feature vector through a classifier, and perform conformal calibration on the prediction result to generate a fraud prediction set that meets a preset confidence level.

[0074] Specifically, the embodiment of the present application adopts a multi-strategy conformal prediction layer to output a confidence calibration block through cascaded feature transfer to form a closed-loop optimization system, which provides a fraud determination set for suspicious claim positioning and review priority while ensuring detection accuracy.

[0075] See also Figure 6 , Figure 6 A specific implementation of step S5 is shown, which is described in detail as follows:

[0076] S51: Divide the label set corresponding to the high-density feature vector to generate multiple independent calibration samples;

[0077] S52: classifying each of the independent calibration samples based on the high-density feature vector using the classifier to generate the prediction result corresponding to each of the independent calibration samples;

[0078] S53: Calculating an inconsistency score based on the prediction result according to a preset conformal prediction method, generating a quantile threshold, and performing prediction screening based on the quantile threshold and the preset confidence level to generate the fraud prediction set. The preset conformal prediction method is either a coverage priority method or a prediction size priority method.

[0079] The embodiment of the present application expects to perform analysis in the post-classifier prediction stage. Conformal prediction is a confidence predictor that can calculate the confidence of the prediction results based on the machine learning algorithm. This framework outputs the P-value of each result to measure the degree of consistency between the prediction set and the true set. The size of the prediction set and the category coverage are two indicators for evaluating the prediction set, so there are two design ideas for conformal prediction: "prediction size priority" and "coverage priority", that is, the preset conformal prediction method in the embodiment of the present application is either the coverage priority method or the prediction size priority method.

[0080] The coverage priority approach focuses on covering as many true results as possible at a given confidence level. Its goal is to provide high coverage, even if the confidence interval may be large. After making a slight correction for the set size, formula (11) is the minimum confidence level that just meets 1-α calculated on the calibration data set. . It can help determine whether each prediction on the test set can enter the prediction set. Specifically, Representation and eigenvectors The corresponding label set, the sum of its probabilities is , n represents the sample size of the data set, and α is the confidence level. Formula (12) determines whether each predicted label of the test set can enter the prediction set. Among them, formula (11) and formula (12) are respectively:

[0081] ;

[0082] .

[0083] The prediction size priority method is to minimize the prediction interval size as much as possible under a given confidence level. Its goal is to minimize the prediction interval to improve the certainty of the prediction. As shown in formula (13), is the sum of the probabilities of labels with a greater probability than y, represents the probability of label y, and u is a randomized term. The last term is the regularization term. Formula (14) generates a prediction set on the test set. The setting adds a matching upper bound to the set-valued function to limit the size of the prediction set to a certain extent, so that the set will not expand unconstrainedly to meet the confidence level of 1-α. Among them, Formula (13) and Formula (14) are respectively:

[0084] .

[0085] In one specific embodiment, after step S5, the process further includes: combining attention weights and long short-term memory network hidden state analysis to output an interpretable fraud location report. Specifically, the process extracts the weight of each time step in the dynamic attention mechanism and identifies high-risk claim records with weights above a preset threshold. Furthermore, the process uses the long short-term memory network hidden state to reversely trace claims with abnormal time intervals, generate a visual analysis of abnormal time series patterns, and generate an interpretable fraud location report.

[0086] In an embodiment of the present application, heterogeneous medical data is obtained, and standardized encoding and feature fusion are performed on the heterogeneous medical data to generate a high-dimensional fused feature sequence; time series features are extracted from the high-dimensional fused feature sequence according to the claim time interval through a long short-term memory network to generate a time series feature sequence; a dynamic attention mechanism is used to perform feature aggregation on the time series feature sequence to generate a fixed-dimensional fraud representation vector; feature extraction is performed on the fraud representation vector through a fully connected network to generate a high-density feature vector; a classifier is used to generate a prediction result based on the high-density feature vector, and the prediction result is conformally calibrated to generate a fraud prediction set that meets a preset confidence level. The embodiment of the present invention performs feature encoding on heterogeneous medical data, extracts features from the claims time interval through a long short-term memory network, and uses conformal calibration to perform prediction calibration. This allows the present application to detect medical fraud for medical providers, improve the comprehensiveness of detection, and capture the time series information of claims, which is conducive to improving the accuracy of medical fraud detection.

[0087] In the embodiment of the present application, an improved recurrent neural network is used to capture the time series information of claims. The attention mechanism can aggregate the sequence so that the neural network can process it. At the same time, the weight analysis based on the time series attention mechanism can focus on suspicious claims in the time series. The embodiment of the present application provides confidence guarantee for the prediction results. By introducing conformal prediction, this framework can obtain a prediction set for each sample, providing priority for subsequent manual review and compensating for the probability bias caused by unbalanced data.

[0088] Please refer to Figure 7 , as a response to the above Figure 1 The present application provides an embodiment of a medical insurance fraud detection device. Figure 1 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0089] like Figure 7 As shown, the medical insurance fraud detection device of this embodiment includes: a feature fusion module 61, a temporal feature extraction module 62, a dynamic attention aggregation module 63, a feature refinement module 64 and a conformal calibration module 65, wherein:

[0090] A feature fusion module 61 is used to obtain heterogeneous medical data, perform standardized coding and feature fusion on the heterogeneous medical data, and generate a high-dimensional fusion feature sequence;

[0091] A time series feature extraction module 62 is configured to extract time series features from the high-dimensional fusion feature sequence according to the claim time interval using a long short-term memory network to generate a time series feature sequence;

[0092] A dynamic attention aggregation module 63 is configured to perform feature aggregation on the temporal feature sequence using a dynamic attention mechanism to generate a fixed-dimensional fraud representation vector;

[0093] a feature refining module 64 for performing feature extraction processing on the fraud representation vector through a fully connected network to generate a high-density feature vector;

[0094] The conformal calibration module 65 is configured to generate a prediction result based on the high-density feature vector using a classifier, and to perform conformal calibration on the prediction result to generate a fraud prediction set that meets a preset confidence level.

[0095] Furthermore, the time series feature extraction module 62 includes:

[0096] An input confirmation unit, configured to use the high-dimensional fusion feature sequence and the claim time interval as inputs to the long short-term memory network;

[0097] a calculation unit, configured to calculate a forget gate based on the high-dimensional fusion feature sequence and the claim time interval using a first preset formula;

[0098] A time series feature sequence generating unit is used to dynamically adjust the amount of information of the candidate memory unit according to the claim time interval, calculate the output hidden state of each time step, and splice the output hidden states to generate a time series feature sequence.

[0099] Furthermore, the first preset formula is:

[0100] ;

[0101] in, is the forget gate, 、 、 are randomly initialized weights, is the bias generated by random initialization, is the high-dimensional fusion feature sequence, for the claim time interval, 、 For sequence 、 The moment of claim.

[0102] Furthermore, the dynamic attention aggregation module 63 includes:

[0103] A key-value pair construction unit, configured to map the aggregated target into a query vector through a fully connected layer, and construct a key-value pair from the temporal feature sequence;

[0104] a normalization processing unit, configured to calculate a score for each value based on the query vector and the key-value pair using a scaled dot product function to obtain an initial score, and perform normalization processing on the initial score to obtain a normalized score;

[0105] A feature aggregation unit is configured to perform a weighted summation on the normalized score and the input value to perform feature aggregation on the time series feature sequence and generate the fraud representation vector of fixed dimension.

[0106] Furthermore, the conformal calibration module 65 includes:

[0107] A sample division unit, configured to divide the label set corresponding to the high-density feature vector to generate a plurality of independent calibration samples;

[0108] a sample classification unit, configured to classify each of the independent calibration samples based on the high-density feature vector using the classifier, and generate the prediction result corresponding to each of the independent calibration samples;

[0109] A prediction screening unit is used to calculate an inconsistency score based on the prediction result according to a preset conformal prediction method, generate a quantile threshold, and perform prediction screening according to the quantile threshold and the preset confidence level to generate the fraud prediction set, wherein the preset conformal prediction method is any one of a coverage priority method and a prediction size priority method.

[0110] Furthermore, the feature fusion module 61 includes:

[0111] A data acquisition unit, configured to acquire the heterogeneous medical data;

[0112] A first vector generating unit, configured to map the discrete categorical data in the heterogeneous medical data into a continuous vector of fixed dimension through an embedding layer to obtain a first vector;

[0113] A second vector generating unit, configured to extract semantic embedding from text data in the heterogeneous medical data to generate a second vector;

[0114] A third vector generating unit, configured to perform standardization processing on the line-type numerical data in the heterogeneous medical data to generate a third vector;

[0115] The nonlinear fusion unit is used to perform feature splicing on the first vector, the second vector and the third vector, and perform nonlinear fusion on the spliced ​​features to generate the high-dimensional fusion feature sequence.

[0116] Furthermore, the feature refining module 64 includes:

[0117] a nonlinear transformation unit, configured to perform a multi-layer nonlinear transformation on the fraud representation vector through the fully connected network to obtain a transformed feature vector;

[0118] A feature compression unit, configured to perform feature compression on the changed feature vector to generate an initial high-density feature vector;

[0119] The feature splicing unit is used to extract the static features if there are any, and to splice the extracted static features with the dynamic features in the initial high-density feature vector to generate the high-density feature vector.

[0120] To solve the above technical problems, the present application also provides a computer device. Figure 8 , Figure 8 This is a basic structural block diagram of the computer device in this embodiment.

[0121] The computer device 7 includes a memory 71, a processor 72, and a network interface 73 that are interconnected through a system bus. It should be noted that Figure 8 Only a computer device 7 having three components, memory 71, processor 72, and network interface 73, is shown. However, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead. It should be understood by those skilled in the art that a computer device herein is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to a microprocessor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc.

[0122] Computer devices can be desktop computers, laptops, PDAs, cloud servers, etc. Computer devices can interact with users through keyboards, mice, remote controls, touchpads, or voice control devices.

[0123] Memory 71 includes at least one type of readable storage medium, including flash memory, hard disks, multimedia cards, card-type memories (e.g., SD or DX memories), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, magnetic disks, optical disks, and the like. In some embodiments, memory 71 may be an internal storage unit of computer device 7, such as the hard disk or internal memory of computer device 7. In other embodiments, memory 71 may also be an external storage device of computer device 7, such as a plug-in hard disk, a SmartMedia Card (SMC), a Secure Digital (SD) card, or a flash memory card. Of course, memory 71 may also include both internal and external storage devices of computer device 7. In this embodiment, memory 71 is typically used to store the operating system and various application software installed on computer device 7, such as the program code for a medical insurance fraud detection method. Furthermore, memory 71 may also be used to temporarily store various types of data that has been or will be output.

[0124] In some embodiments, processor 72 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. Processor 72 is generally used to control the overall operation of computer device 7. In this embodiment, processor 72 is used to execute program code stored in memory 71 or process data, such as executing the program code of the aforementioned medical insurance fraud detection method to implement various embodiments of the medical insurance fraud detection method.

[0125] The network interface 73 may include a wireless network interface or a wired network interface. The network interface 73 is generally used to establish a communication connection between the computer device 7 and other electronic devices.

[0126] The present application also provides another embodiment, namely, providing a computer-readable storage medium, which stores a computer program. The computer program can be executed by at least one processor to enable the at least one processor to perform the steps of a medical insurance fraud detection method as described above.

[0127] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of this application, or the part that contributes to the existing technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of each embodiment of this application.

[0128] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of protection of the present application.

Claims

1. A medical insurance fraud detection method, characterized in that: include: Acquire heterogeneous medical data, and perform standardized encoding and feature fusion on the heterogeneous medical data to generate a high-dimensional fusion feature sequence; Extracting time series features from the high-dimensional fusion feature sequence according to the claim time interval using a long short-term memory network to generate a time series feature sequence; A dynamic attention mechanism is used to aggregate features of the temporal feature sequence to generate a fraud representation vector of fixed dimension; Performing feature extraction processing on the fraud representation vector through a fully connected network to generate a high-density feature vector; generating prediction results based on the high-density feature vectors using a classifier, and conformally calibrating the prediction results to generate a fraud prediction set that meets a preset confidence level; The step of acquiring heterogeneous medical data and performing standardized coding and feature fusion on the heterogeneous medical data to generate a high-dimensional fusion feature sequence includes: Acquiring the heterogeneous medical data; Mapping the discrete categorical data in the heterogeneous medical data into a continuous vector of fixed dimension through an embedding layer to obtain a first vector; Extracting semantic embedding from text data in the heterogeneous medical data to generate a second vector; performing normalization processing on the continuous numerical features in the heterogeneous medical data to generate a third vector; Performing feature splicing on the first vector, the second vector, and the third vector, and performing nonlinear fusion on the spliced ​​features to generate the high-dimensional fusion feature sequence; The generating of prediction results based on the high-density feature vector by a classifier and conformal calibration of the prediction results to generate a fraud prediction set that meets a preset confidence level includes: Dividing the label set corresponding to the high-density feature vector to generate multiple independent calibration samples; classifying each of the independent calibration samples based on the high-density feature vector by the classifier, and generating the prediction result corresponding to each of the independent calibration samples; An inconsistency score is calculated based on the prediction result according to a preset conformal prediction method to generate a quantile threshold, and prediction screening is performed based on the quantile threshold and the preset confidence level to generate the fraud prediction set, wherein the preset conformal prediction method is any one of a coverage priority method and a prediction size priority method.

2. The medical insurance fraud detection method according to claim 1, characterized in that: The extracting time series features from the high-dimensional fusion feature sequence according to the claim time interval by using the long short-term memory network to generate a time series feature sequence includes: Taking the high-dimensional fusion feature sequence and the claim time interval as inputs of the long short-term memory network; Calculating a forget gate based on the high-dimensional fusion feature sequence and the claim time interval using a first preset formula; The information amount of the candidate memory unit is dynamically adjusted according to the claim time interval, and the output hidden state of each time step is calculated, and the output hidden states are spliced ​​to generate a temporal feature sequence.

3. The medical insurance fraud detection method according to claim 2, characterized in that: The first preset formula is: ， ； in, is the forget gate, 、 、 are randomly initialized weights, is the bias generated by random initialization, is the high-dimensional fusion feature sequence, for the claim time interval, 、 For sequence 、 The moment of claim.

4. The medical insurance fraud detection method according to claim 1, characterized in that: The dynamic attention mechanism is used to aggregate the time series feature sequences to generate a fixed-dimensional fraud representation vector, including: Mapping the aggregated target into a query vector through a fully connected layer, and constructing a key-value pair based on the temporal feature sequence; Calculating a score for each value based on the query vector and the key-value pair using a scaled dot product function to obtain an initial score, and normalizing the initial score to obtain a normalized score; The normalized score and the input value are weightedly summed to perform feature aggregation on the time series feature sequence to generate the fraud representation vector of fixed dimension.

5. The medical insurance fraud detection method according to any one of claims 1 to 4, characterized in that: The step of performing feature extraction on the fraud representation vector through a fully connected network to generate a high-density feature vector includes: Performing multi-layer nonlinear transformation on the fraud representation vector through the fully connected network to obtain a transformed feature vector; Performing feature compression on the changed feature vector to generate an initial high-density feature vector; If static features exist, feature extraction is performed on the static features, and the refined static features are concatenated with the dynamic features in the initial high-density feature vector to generate the high-density feature vector.

6. A medical insurance fraud detection device, characterized in that: include: A feature fusion module is used to obtain heterogeneous medical data, perform standardized encoding and feature fusion on the heterogeneous medical data, and generate a high-dimensional fusion feature sequence; A time series feature extraction module is used to extract time series features from the high-dimensional fusion feature sequence according to the claim time interval through a long short-term memory network to generate a time series feature sequence; A dynamic attention aggregation module, configured to perform feature aggregation on the temporal feature sequence using a dynamic attention mechanism to generate a fixed-dimensional fraud representation vector; a feature refinement module, configured to perform feature refinement processing on the fraud representation vector through a fully connected network to generate a high-density feature vector; a conformal calibration module, configured to generate a prediction result based on the high-density feature vector using a classifier, and conformally calibrate the prediction result to generate a fraud prediction set that meets a preset confidence level; Wherein, the feature fusion module includes: A data acquisition unit, configured to acquire the heterogeneous medical data; A first vector generating unit, configured to map the discrete categorical data in the heterogeneous medical data into a continuous vector of fixed dimension through an embedding layer to obtain a first vector; A second vector generating unit, configured to extract semantic embedding from text data in the heterogeneous medical data to generate a second vector; A third vector generating unit, configured to perform normalization processing on the continuous numerical features in the heterogeneous medical data to generate a third vector; a nonlinear fusion unit, configured to perform feature splicing on the first vector, the second vector, and the third vector, and to perform nonlinear fusion on the spliced ​​features to generate the high-dimensional fused feature sequence; The conformal calibration module includes: A sample division unit, configured to divide the label set corresponding to the high-density feature vector to generate a plurality of independent calibration samples; a sample classification unit, configured to classify each of the independent calibration samples based on the high-density feature vector using the classifier, and generate the prediction result corresponding to each of the independent calibration samples; A prediction screening unit is used to calculate an inconsistency score based on the prediction result according to a preset conformal prediction method, generate a quantile threshold, and perform prediction screening according to the quantile threshold and the preset confidence level to generate the fraud prediction set, wherein the preset conformal prediction method is any one of a coverage priority method and a prediction size priority method.

7. A computer device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the medical insurance fraud detection method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the medical insurance fraud detection method according to any one of claims 1 to 5.

Images