Auto-encoder-based medical insurance fraud detection method, apparatus and device, and medium

Abstract

The invention provides a medical insurance fraud detection method and device based on an auto-encoder, equipment and a medium, and is applied to a medical insurance system, and the method comprises the steps: obtaining a medical insurance claim data set, and carrying out the structural processing of initial data in the medical insurance claim data set, so as to obtain a dimension-reduced data set after dimension reduction; dividing dimension reduction data in the dimension reduction data set into a training set according to a proportion; inputting the dimensionality reduction data in the training set into an auto-encoder so that the auto-encoder is trained to obtain an anomaly detection model; and if to-be-processed medical insurance claim data is received, detecting the to-be-processed medical insurance claim data through the anomaly detection model to obtain a corresponding anomaly score, and performing risk marking on the to-be-processed medical insurance claim data according to the anomaly score and a preset risk threshold. According to the method, the initial data is subjected to structured processing to obtain the dimension-reduced data after dimension reduction, so that redundant features of the data are reduced, the generalization ability of the model is enhanced, and the over-fitting of the model is avoided, thereby improving the accuracy of the model for medical insurance fraud detection.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a medical insurance fraud detection method, device, equipment and medium based on an autoencoder. Background Art

[0002] With the rapid development of medical informatization, medical data has shown the characteristics of large data volume, multiple data types, and wide data sources, which provides sufficient data foundation and technical support for the detection of medical insurance fraud.

[0003] As a typical AI unsupervised learning model, the autoencoder offers a new technical perspective for healthcare fraud detection. It identifies anomalous patterns through data reconstruction errors and is particularly well-suited for processing high-dimensional, unstructured medical data. However, medical data has significant domain characteristics, such as high dimensionality, complex categorical fields, and extreme data imbalance (fraud samples account for <0.1%). These characteristics can cause the autoencoder to overfit to frequently occurring normal medical visit data, ignoring the long-tail distribution of normal medical visit data, and causing marginal normal samples to be misclassified as fraudulent. Summary of the Invention

[0004] Embodiments of the present invention provide a medical insurance fraud detection method, apparatus, device, and medium based on an autoencoder, which aim to solve the problem of class imbalance in samples, thereby improving classification accuracy.

[0005] In a first aspect, an embodiment of the present invention provides a medical insurance fraud detection method based on an autoencoder, which includes: obtaining a medical insurance claim data set, performing structured processing on the initial data in the medical insurance claim data set to obtain a reduced dimension data set after dimensionality reduction; dividing the reduced dimension data in the reduced dimension data set into a training set in proportion; inputting the reduced dimension data in the training set into the autoencoder to train it to obtain an anomaly detection model; if unprocessed medical insurance claim data is received, the unprocessed medical insurance claim data is detected by the anomaly detection model to obtain a corresponding anomaly score, and the unprocessed medical insurance claim data is risk-marked according to the anomaly score and a preset risk threshold.

[0006] In a second aspect, an embodiment of the present invention further provides a medical insurance fraud detection device based on an autoencoder, which includes a unit for executing the above method.

[0007] In a third aspect, an embodiment of the present invention further provides a computer device, which includes a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the above method when executing the computer program.

[0008] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the above method can be implemented.

[0009] The present application provides a medical insurance fraud detection method, device, equipment and medium based on an autoencoder, which is applied to a medical insurance system. The medical insurance fraud detection method includes: obtaining a medical insurance claim data set, performing structured processing on the initial data in the medical insurance claim data set to obtain a reduced dimension data set after dimensionality reduction; dividing the reduced dimension data in the reduced dimension data set into a training set in proportion; inputting the reduced dimension data in the training set into the autoencoder to train it to obtain an anomaly detection model; if the medical insurance claim data to be processed is received, the medical insurance claim data to be processed is detected by the anomaly detection model to obtain a corresponding anomaly score, and the medical insurance claim data to be processed is risk-marked according to the anomaly score and a preset risk threshold. The present application obtains reduced dimension data after dimensionality reduction by performing structured processing on the initial data of the medical insurance claim data set, and divides the reduced dimension data into a training set in proportion so that the autoencoder can be trained. Since the training data has no redundant features, the computational complexity of the anomaly detection model is reduced, the training speed of the model is accelerated, and the processing efficiency of the model is improved; at the same time, the model can be simplified to enhance the generalization ability of the model, avoid overfitting of the model, and ultimately improve the accuracy of the model in detecting medical insurance fraud. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 A schematic flow chart of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention;

[0012] Figure 2 A schematic diagram of a sub-process of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention;

[0013] Figure 3 A schematic diagram of a sub-process of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention;

[0014] Figure 4 A schematic diagram of a sub-process of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention;

[0015] Figure 5A schematic diagram of a sub-process of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention;

[0016] Figure 6 A schematic block diagram of a medical insurance fraud detection device based on an autoencoder provided in an embodiment of the present invention;

[0017] Figure 7 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0022] See also Figure 1 , Figure 1A schematic flow chart of a medical insurance fraud detection method based on an autoencoder provided in an embodiment of the present invention. In this application, the medical insurance fraud detection method based on an autoencoder is applied to the medical insurance system, and in particular, the medical insurance claim data is detected to determine whether there is fraud in the medical insurance claim data reported by the user, such as identifying whether there are fraudulent behaviors such as inflated items, repeated reimbursements, and over-prescription of medicines; such as detecting the abnormal pattern of "a hypertensive patient is prescribed 10 boxes of antihypertensive drugs at a time", and linking with the medical insurance settlement system to achieve real-time fraud interception, automatically triggering manual review of high-risk reimbursements, thereby achieving accurate detection of medical insurance fraud, avoiding fraud risks, thereby reducing the loss of medical insurance funds, and improving the utilization rate of medical insurance funds.

[0023] The present application provides a medical insurance fraud detection method, apparatus, device and medium based on an autoencoder, which are applied to a medical insurance system. The medical insurance fraud detection method based on an autoencoder includes: obtaining a medical insurance claim data set, performing structured processing on the initial data in the medical insurance claim data set to obtain a reduced dimension data set after dimensionality reduction; dividing the reduced dimension data in the reduced dimension data set into a training set in proportion; inputting the reduced dimension data in the training set into an autoencoder to train it to obtain an anomaly detection model; if unprocessed medical insurance claim data is received, the unprocessed medical insurance claim data is detected by the anomaly detection model to obtain a corresponding anomaly score, and the unprocessed medical insurance claim data is risk-marked according to the anomaly score and a preset risk threshold.

[0024] This application obtains reduced-dimensionality data by performing structured processing on the initial data of the medical insurance claims dataset, and divides the reduced-dimensionality data into a training set in proportion so that the autoencoder can be trained. Since the training data has no redundant features, the computational complexity of the anomaly detection model is reduced, the training speed of the model is accelerated, and the processing efficiency of the model is improved; at the same time, the model can be simplified to enhance the generalization ability of the model, avoid overfitting of the model, and ultimately improve the accuracy of the model in detecting medical insurance fraud.

[0025] Figure 1 1 is a flow chart of a method for detecting medical insurance fraud based on an autoencoder according to an embodiment of the present invention. As shown in the figure, the method includes the following steps S110-S150.

[0026] S110, obtaining a medical insurance claim dataset, and performing structured processing on initial data in the medical insurance claim dataset to obtain a reduced dimension dataset after dimensionality reduction;

[0027] In this embodiment, the medical insurance claim data set is the medical insurance claim data of users over the years, including normal data and fraud data, wherein normal data refers to normal medical insurance claim data, and fraud data refers to medical insurance claim data of fraudulent claims; one medical insurance claim data corresponds to a user's medical insurance claim application, and the medical insurance claim data includes features and labels, and the features refer to the user's hospitalization data, outpatient treatment data, insured person information data, medical treatment amount and other feature data; the labels include normal types and fraud types; wherein the initial data is the medical insurance claim data, and the medical insurance claim data set includes multiple initial data. Since the initial data is high-dimensional, unstructured data, and its features are relatively complex, the autoencoder may overfit when learning the initial data. This is because the autoencoder may rely too much on unique identifiers during the learning process, and thus fail to effectively capture the common patterns and potential structures of the data, thereby affecting the generalization ability of the model. For example, in the case of high-cardinality coding features (such as doctor ID information, claim ID information, and diagnosis code information), the autoencoder will pay more attention to these features, resulting in overfitting problems, which will cause the normal type of initial data on the edge to be misjudged as fraud type, thereby making the model inaccurate in detecting medical insurance fraud.

[0028] In this embodiment, after obtaining the medical insurance claim data set, all the initial data in the medical insurance claim data set are structured, so that the initial data becomes structured data. The structured processing also includes dimensionality reduction processing to reduce the dimensionality of the initial data, thereby removing redundant features of the initial data, and then the initial data after dimensionality reduction constitute the dimensionality reduction data set, so that the dimensionality reduction data in the dimensionality reduction data set can be applied in the model, reducing the complexity of the model calculation, speeding up the model training, and improving the processing efficiency of the model; at the same time, it can also simplify the model to enhance the generalization ability of the model, avoid overfitting of the model, and ultimately improve the accuracy of the model in detecting medical insurance fraud.

[0029] S120, dividing the dimension-reduced data in the dimension-reduced data set into a training set in proportion;

[0030] Specifically, the dimensionality reduction data set includes multiple dimensionality reduction data, that is, data after structured processing; the training set is used for model training to improve the learning ability of the model, and the trained model can detect the medical insurance claim data to be tested and determine its type. Therefore, the data in the training set plays an important role in the operation of the model.

[0031] In this embodiment, the reduced dimensionality data in the reduced dimensionality data set is divided into the training set in proportion so that the reduced dimensionality data in the training set can be trained and learned by the model. For complex models, more reduced dimensionality data is required for effective learning, and therefore, more reduced dimensionality data needs to be divided into the training set.

[0032] More specifically, the reduced-dimensionality data in the reduced-dimensionality dataset is further divided into a validation set and a test set in proportion. The validation set is used for hyperparameter tuning and model selection. By monitoring the reconstruction error (such as MSE) of the validation set, the model with the smallest error is selected as the optimal parameter, and hyperparameters such as the learning rate, batch size, and latent space dimension are adjusted. The test set is used to evaluate generalization ability. In this embodiment, the reduced-dimensionality data in the reduced-dimensionality dataset is divided into the training set, the validation set, and the test set in a ratio of 8:1:1.

[0033] S130, inputting the dimension-reduced data in the training set into an autoencoder to train it to obtain an anomaly detection model;

[0034] Specifically, an autoencoder is an unsupervised learning neural network model that learns the low-dimensional potential representation of data by imitating the compression and reconstruction process of data. Its core goal is to make the output replicate the input as much as possible, while forcing the network to extract key features of the data through the "bottleneck layer"; the training set is used for training and learning the autoencoder.

[0035] In this embodiment, all the dimensionality reduction data in the training set are input into the autoencoder, so that the autoencoder is trained and learned according to the dimensionality reduction data, and finally an anomaly detection model is obtained, and then the anomaly detection model is enabled to perform classification detection on the medical insurance claim data to be processed. Since the data input to the autoencoder is dimensionality reduction data, the computational complexity of the autoencoder is reduced, the training speed of the autoencoder is accelerated, and the training efficiency is improved; at the same time, the redundant features of the input data are removed, the anomaly detection model is simplified, the generalization ability of the anomaly detection model is enhanced, and finally the classification accuracy of the anomaly detection model is improved.

[0036] More specifically, the autoencoder is trained on the training set based on the Pytorch deep learning framework, allowing the autoencoder to learn the potential representation of medical insurance data. This means that the high-dimensional input data is mapped to a low-dimensional latent space through the neural network encoder. In this low-dimensional space, each point represents the key information of the input data and its inherent structure, and can summarize the main features of the data. This representation method helps the model learn and understand data more efficiently, thereby facilitating subsequent generation or classification tasks.

[0037] In one embodiment, the autoencoder includes a first autoencoder and a second autoencoder, and step S130 includes step S131.

[0038] S131. Input the dimension-reduced data in the training set into the first autoencoder and the second autoencoder for training.

[0039] Specifically, the autoencoder includes a first autoencoder and a second autoencoder, and adopts a collaborative autoencoder module. Compared with the single autoencoder structure, a robust representation of the potential result can be obtained. In this embodiment, dual autoencoders are used in collaboration (first autoencoder A1 and second autoencoder A2). The first autoencoder A1 and the second autoencoder A2 adopt a symmetric encoder-decoder structure to input the reduced dimensionality data in the training set into the first autoencoder A1 and the second autoencoder A2. The first autoencoder A1 and the second autoencoder A2 respectively map the reduced dimensionality data to a continuous and dense latent space through the encoder part, which means that most areas in this space contain reasonable representations of the data, and almost every position can generate valid outputs. In the dense latent space, when any two similar points are mapped to the output data, similar data samples will be generated, thereby ensuring the continuity and smoothness of the model; then the latent representation is reconstructed through the decoder part, thereby avoiding omissions and improving training accuracy. Therefore, the anomaly detection model learns the normal pattern based on the first autoencoder and the second autoencoder. The normal pattern refers to the medical behavior pattern that appears frequently in the data and complies with business rules, reflecting the compliant and reasonable diagnosis and treatment and fee settlement process, that is, normal type of data.

[0040] In one embodiment, step S110 may include step S111 .

[0041] S111 , using principal component analysis and principal component extraction to perform structured processing on the initial data to obtain a dimension-reduced data set.

[0042] Specifically, Principal Component Analysis (PCA) is a classic unsupervised linear dimensionality reduction method. It maps the original high-dimensional data into a new set of low-dimensional coordinates (the principal components) through an orthogonal transformation, such that the axes of the new coordinate system correspond to the directions of maximum data variance. Its core goal is to compress the data dimensions with minimal information loss, while removing redundant features and revealing the data's inherent structure.

[0043] When the initial data in the medical insurance claim dataset is structured to obtain a reduced dimension dataset, principal component analysis and principal component extraction are used to structure the initial data to obtain a reduced dimension dataset. The principal component analysis and principal component extraction methods can effectively remove redundancy, retain key information, reduce computational complexity, and improve efficiency. The specific steps are:

[0044] First, the initial data is subjected to data standardization to standardize the initial data and eliminate the influence of different dimensions. The formula is: Z is the result after standardization, X is the feature, μ is the feature mean, that is, the mean of the current feature is calculated for each feature, and σ is the standard deviation of the feature.

[0045] The second step is to calculate the covariance matrix to reflect the linear relationship between different features: Where n is the number of samples, Z is the standardized data matrix, and Z T is the transpose of the normalized data matrix, and C is the covariance matrix.

[0046] The next step is to select the principal components, which are selected based on the size of the eigenvalues ​​of the covariance matrix C. Usually, the first k eigenvectors with larger eigenvalues ​​are selected.

[0047] Finally, the data is transformed and the initial data is projected onto the selected principal component to obtain the data after dimensionality reduction. The formula is: X reduced =ZV k , where V k Represents the matrix composed of the first k eigenvectors selected, X reduced is the data after dimensionality reduction, i.e., the dimensionality reduction data. The dimensionality reduction data set consists of all the dimensionality reduction data.

[0048] In one embodiment, if Figure 2 As shown, the autoencoder also includes an abnormal autoencoder, and step S131 includes steps S1311-S1314.

[0049] S1311, inputting the dimension-reduced data in the training set into the first autoencoder and the second autoencoder to obtain decoded and reconstructed data;

[0050] S1312, performing reconstruction error calculation on the decoded reconstructed data and the dimensionality reduced data to obtain reconstruction error data;

[0051] S1313, sorting the reconstructed error data, and selecting data whose reconstructed error data is greater than a preset error as high error data according to the abnormality ratio;

[0052] S1314: Input the high-error data into the abnormal autoencoder to enable it to learn.

[0053] Specifically, in order to further improve the detection accuracy of the model, the autoencoder also includes an abnormal autoencoder C, which adopts an autoencoder that is isomorphic to the first autoencoder and the second autoencoder. The abnormal autoencoder focuses on reconstructing the high-error samples discarded by the first autoencoder and the second autoencoder, thereby reconstructing the abnormal samples so that the abnormal detection model learns abnormal patterns. The abnormal pattern refers to a pattern that appears in the data at a low frequency and deviates from the normal rules, which may indicate fraud, abuse or system errors. In this embodiment, it refers to fraud-type data. Therefore, through the joint training mechanism of the first autoencoder, the second autoencoder and the abnormal autoencoder, the abnormal detection model can simultaneously learn the feature expressions of normal patterns and abnormal patterns, so that the abnormal detection model has a more accurate abnormality capture capability, thereby further improving the detection accuracy of the abnormal detection model.

[0054] In this embodiment, first, the reduced-dimensionality data in the training set is input into the first autoencoder A1 and the second autoencoder A2, and decoded and reconstructed data is generated according to the first autoencoder and the second autoencoder. Specifically, the reduced-dimensionality data of the training set is input into the first autoencoder and the second autoencoder, and the encoders of the first autoencoder and the second autoencoder map the reduced-dimensional data to a continuous and dense latent space. In the dense latent space, when any two similar points are mapped to the output data, similar data samples are generated. The potential representation is then reconstructed by the decoder. The formula is:

[0055] h=Encoder(x,θ e )

[0056]

[0057] Where Encoder is the encoder, Decoder is the decoder, x is the dimensionality reduction data in the training set, and h is the encoding result, i.e., the potential representation. To decode and reconstruct data, θ e ,θ d They are the encoder and decoder, that is, the weights and biases of neurons in a multi-layer deep neural network, which are learned through the neural network back propagation algorithm.

[0058] Then, reconstruction error data is obtained by calculating the reconstruction error between the decoded reconstructed data and the dimensionality-reduced data, that is, calculating the difference between the original input data (i.e., the dimensionality-reduced data in the training set) and the model reconstruction output (i.e., the decoded reconstructed data). The reconstruction error can be defined in many ways, the most common of which is the mean square error (MSE), which is calculated as follows:

[0059]

[0060] Among them, x i is the dimension-reduced data in the training set, The first autoencoder and the second autoencoder generate decoded reconstructed data, where n is the number of samples, i.e., the number of the dimensionality reduction data, thereby obtaining the reconstructed error data, and optimizing the network structure with the goal of minimizing the reconstructed error data.

[0061] The reconstructed error data is then sorted and, based on the anomaly ratio, data with reconstructed error data greater than a preset error is selected as high-error data. The anomaly ratio is a manually set parameter that can be adjusted based on actual needs; the preset error is a manually set threshold that can be set based on needs. The reconstructed error data is then sorted in ascending order and data with reconstructed error data greater than the preset error is selected as high-error data. High-error data has a high error and is likely to be fraudulent. For example, if the anomaly ratio is set to 9%, the top 9% of reconstructed error data will be selected as high-error data.

[0062] Finally, the high-error data is input into the abnormal autoencoder to enable it to learn. Specifically, the high-error data is input into the abnormal autoencoder, and the encoder of the abnormal autoencoder maps the high-error data to a continuous and dense latent space. In the dense latent space, when any two similar points are mapped to the output data, similar data samples will be generated. The latent representation is then reconstructed through the decoder, and finally the reconstruction of the high-error data is output to obtain the reconstruction error of the high-error data, so that the anomaly detection model learns the abnormal pattern.

[0063] Therefore, in this embodiment, through the joint training mechanism of the first autoencoder, the second autoencoder and the abnormal autoencoder, the abnormality detection model can simultaneously learn the feature expressions of normal mode and abnormal mode, so that the abnormality detection model has more accurate abnormality capture capability, thereby further improving the detection accuracy of the abnormality detection model.

[0064] More specifically, in order to improve computational efficiency during neural network training, a parallel computing method is used to divide the training set into several small training batches. In each training batch B, the model is first forward propagated, and the reconstruction error data of the samples are sorted in ascending order. The samples before β|B| are retained for backpropagation training, and the remaining (1-β)|B| high error data are temporarily stored in the buffer D. drop , where β is the sample selection rate and B is the data batch size, which are hyperparameters tuned during the validation phase. Through a strategy that dynamically decays the sample selection rate for high reconstruction error as training progresses and an active learning feedback mechanism based on auditors' annotation feedback on data, the system can continuously iterate as fraud patterns evolve, enabling more flexible adaptive optimization of the anomaly detection model.

[0065] In one embodiment, if Figure 3 As shown, step S130 also includes steps S1315-S1319.

[0066] S1315, dividing the training set into several small training batches, each training batch has a size of B;

[0067] S1316: Calculate the basic reconstruction loss of the first autoencoder and the second autoencoder, using the following formula: Wherein, x is the dimension-reduced data in the training set input to the first autoencoder A1 and the second autoencoder A2; and The reconstructed error data of the first autoencoder A1 and the second autoencoder A2 respectively;

[0068] S1317: Calculate the abnormal reconstruction loss of the abnormal autoencoder, using the following formula: Among them, D drop Refers to high error data, Refers to the reconstructed error data of the abnormal autoencoder C;

[0069] S1318. Calculate the total objective function. The calculation formula is: L total =L recon (A1,A2)+λL recon (C); where λ is a hyperparameter;

[0070] S1319: Input the total objective function into the anomaly detection model for optimization.

[0071] Specifically, the anomaly detection model involves a multi-task loss function. To reduce the loss, the first autoencoder, the second autoencoder, and the anomaly autoencoder are jointly optimized. First, the training set is divided into several small training batches, each of size B. Then, the basic reconstruction loss of the first autoencoder and the second autoencoder is calculated using the following formula: Wherein, x is the dimension-reduced data in the training set input to the first autoencoder A1 and the second autoencoder A2; and are the reconstruction error data of the first autoencoder A1 and the second autoencoder A2, respectively, to obtain the reconstruction loss of the first autoencoder and the second autoencoder; and then calculate the abnormal reconstruction loss of the abnormal autoencoder, and the calculation formula is: Among them, D drop Refers to high error data, Refers to the reconstruction error data of the abnormal autoencoder C, thereby obtaining the reconstruction loss of the abnormal autoencoder; finally, the total objective function is calculated, and the calculation formula is: L total =L recon (A1,A2)+λL recon (C), where λ is a hyperparameter that can be set according to actual conditions to control the degree of attention the anomaly detection model pays to the reconstruction of abnormal data, where the abnormal data refers to fraudulent data; and the overall objective function is then input into the anomaly detection model to optimize the anomaly detection model.

[0072] Therefore, in this embodiment, during the training phase, by minimizing the overall objective function, the anomaly detection model learns the ability to reconstruct normal data, thereby distinguishing abnormal data patterns through the size of the reconstruction error. Normal data generally refers to samples that conform to the expected pattern and do not exhibit abnormal behavior, such as medical expenses within the normal range or compliant claims requests. Abnormal data refers to samples that deviate significantly from the normal pattern, which may include, but are not limited to, extremely high medical expenses, repeated claims, and inconsistent patient medical records. By using the size of the reconstruction error, the anomaly detection model can determine which samples are abnormal data that do not conform to the normal pattern. When the reconstruction error exceeds a set threshold (set manually), these samples are considered abnormal. By optimizing the anomaly detection model, the detection accuracy of the anomaly detection model can be further improved.

[0073] S140: If pending medical insurance claim data is received, the pending medical insurance claim data is detected using the anomaly detection model to obtain a corresponding anomaly score, and risk-marking the pending medical insurance claim data based on the anomaly score and a preset risk threshold;

[0074] Specifically, in the field of applying the anomaly detection model to medical insurance claim data detection, the anomaly detection model classifies and marks the detected data; specifically, if pending medical insurance claim data is received, the pending medical insurance claim data refers to the newly input medical insurance claim data application of the user; the pending medical insurance claim data is detected by the anomaly detection model to obtain the anomaly score corresponding to the pending medical insurance claim data, and the pending medical insurance claim data is risk-marked according to the anomaly score and the preset risk threshold, wherein the preset risk threshold is a manually set value that can be obtained based on experience; the risk mark refers to an indication of the high or low fraud risk of the pending medical insurance claim data, and then humans can further judge whether the pending medical insurance claim data is normal data or fraud data (abnormal data) based on the value of the risk mark.

[0075] The anomaly detection model is used to score the anomaly of the pending medical insurance claim data. The lower the anomaly score, the more normally distributed (i.e., normal data) the pending medical insurance claim data is. The higher the anomaly score, the more abnormal the pending medical insurance claim data is (i.e., the risk of abnormal data is high). This allows for accurate judgment of the type of the pending medical insurance claim data to improve the accuracy of medical insurance fraud detection.

[0076] In one embodiment, if Figure 4 As shown, step S140 includes steps S141-S144.

[0077] S141, obtaining the to-be-processed medical insurance claim data and inputting it into the anomaly detection model to obtain a reconstruction error;

[0078] S142. Perform weighted summation on the reconstruction errors to generate the anomaly score;

[0079] S143: Determine the difference between the abnormality score and the preset risk threshold to mark the risk type of the medical insurance claim data to be processed;

[0080] S144. Mark the to-be-processed medical insurance claim data as normal data or abnormal data according to the risk type and in combination with the business rule library.

[0081] Specifically, the anomaly score is the magnitude of the anomaly of the to-be-processed medical insurance claim data. The lower the anomaly score, the more normally distributed the to-be-processed medical insurance claim data is (i.e., normal data). The higher the anomaly score, the more abnormal the to-be-processed medical insurance claim data is (i.e., the risk of abnormal data is high).

[0082] In this embodiment, the anomaly detection model is used to perform reconstruction error processing on the medical insurance claim data to be processed to obtain the reconstruction error of the medical insurance claim data to be processed, that is, the trained first autoencoder, the second autoencoder and the anomaly autoencoder will output the reconstruction error respectively when inputting the medical insurance claim data to be processed. The reconstruction errors in this embodiment are calculated based on the MSE method; then the reconstruction errors are weighted and summed to generate the abnormality score S final , the specific formula is: Among them, ε is a minimum value, such as 1e-8, to prevent overflow caused by division by zero; the anomaly score is then judged against the size of the preset risk threshold to mark the risk type of the to-be-processed medical insurance claim data, and the higher the risk type, the more abnormal the data; finally, according to the risk type and the business rule library, the to-be-processed medical insurance claim data is marked as normal data or abnormal data. The business rule library includes, for example, the medical insurance refusal rules where the applied medical service or treatment does not match the patient's diagnosis or condition or lacks necessary medical documents, etc. Therefore, after determining the risk type of the to-be-processed medical insurance claim data, it is necessary to further determine whether it is normal data or abnormal data in combination with the provisions of the business rule library. If it is normal data, the normal compensation process will be carried out. If it is abnormal data, no compensation will be made, thereby improving the judgment of the compensation of medical insurance claim data and avoiding the risk of fraud.

[0083] In one embodiment, if Figure 5 As shown, the preset risk threshold includes a first risk threshold, a second risk threshold and a third risk threshold, the first risk threshold is greater than the second risk threshold, the second risk threshold is greater than the third risk threshold, and step S143 includes steps S1431-S1433.

[0084] S1431. If the anomaly score is greater than the first risk threshold, mark it as high risk;

[0085] S1432: If the anomaly score is less than the first risk threshold and greater than the second risk threshold, mark it as medium risk;

[0086] S1433: If the abnormality score is less than the second risk threshold and greater than the third risk threshold, mark it as low risk.

[0087] Specifically, in this embodiment, three levels of risk thresholds are set, and the preset risk thresholds include a first risk threshold, a second risk threshold and a third risk threshold. The first risk threshold is greater than the second risk threshold, the second risk threshold is greater than the third risk threshold, and the first risk threshold, the second risk threshold and the third risk threshold can be set according to actual needs.

[0088] If the anomaly score is greater than the first risk threshold, it is marked as high risk; if the anomaly score is less than the first risk threshold and greater than the second risk threshold, it is marked as medium risk; if the anomaly score is less than the second risk threshold and greater than the third risk threshold, it is marked as low risk.

[0089] More specifically, in this embodiment, the first risk threshold is set to 0.7, the second risk threshold is set to 0.5, and the third risk threshold is set to 0, wherein 0-0.5 is low risk, 0.5-0.7 is medium risk, and greater than 0.7 is high risk.

[0090] Therefore, the system can combine the three-level risk grading warning strategy to implement graded responses to pending medical insurance claim data suspected of fraud, so as to improve the accuracy of medical insurance claim data detection.

[0091] For example, if the reconstruction error of the pending medical insurance claim data is in a high-risk range and meets a certain refusal rule, the case is marked as high-risk; the auditor manually reviews the system's warning cases and labels each case based on their business experience and rules. If the case meets the standards of normal operation, has no abnormal features, and complies with business rules, it is marked as normal data; if the case shows abnormal features and there is evidence pointing to fraud, it is marked as abnormal data.

[0092] This application obtains reduced-dimensionality data after dimensionality reduction by performing structured processing on the initial data of the medical insurance claims data set, and divides the reduced-dimensionality data into a training set in proportion so that the autoencoder can be trained. Since the training data has no redundant features, the computational complexity of the anomaly detection model is reduced, and the training speed of the model is accelerated, thereby improving the processing efficiency of the model. At the same time, the model can be simplified to enhance the generalization ability of the model, avoid overfitting of the model, and ultimately improve the accuracy of the model in detecting medical insurance fraud. At the same time, through the multi-model collaborative training and dynamic anomaly modeling of the first autoencoder, the second autoencoder and the anomaly autoencoder, the comprehensive improvement of the supervision efficiency of the medical insurance fund is achieved. The system combines a three-level risk grading warning strategy to achieve a graded response to suspected fraudulent behavior to improve the accuracy of medical insurance fraud detection.

[0093] Figure 6 FIG is a schematic block diagram of a medical insurance fraud detection device 300 based on an autoencoder provided in an embodiment of the present invention. Figure 6As shown, corresponding to the above-mentioned medical insurance fraud detection method based on autoencoder, the present invention also provides a medical insurance fraud detection device 300 based on autoencoder. The medical insurance fraud detection device based on autoencoder includes a unit for executing the above-mentioned medical insurance fraud detection method based on autoencoder, and the device can be configured in a computer device. Specifically, please refer to Figure 6 The medical insurance fraud detection device 300 based on the autoencoder includes a processing unit 301, a training unit 302, and a detection unit 303.

[0094] Processing unit 301 obtains a medical insurance claim dataset, performs structured processing on initial data in the medical insurance claim dataset to obtain a reduced dimension dataset; divides the reduced dimension data in the reduced dimension dataset into a training set in proportion; and performs structured processing on the initial data using principal component analysis and principal component extraction to obtain a reduced dimension dataset;

[0095] The training unit 302 inputs the reduced-dimensionality data in the training set into the autoencoder to train it to obtain an anomaly detection model; inputs the reduced-dimensionality data in the training set into the first autoencoder and the second autoencoder to train them; inputs the reduced-dimensionality data in the training set into the first autoencoder and the second autoencoder to obtain decoded reconstructed data; calculates the reconstruction error of the decoded reconstructed data and the reduced-dimensionality data to obtain reconstruction error data; sorts the reconstruction error data, and selects data whose reconstruction error data is greater than a preset error as high-error data; inputs the high-error data into the anomaly autoencoder to learn it; divides the training set into several small training batches, each of which is of size B; calculates the basic reconstruction loss of the first autoencoder and the second autoencoder, and the calculation formula is: Wherein, x is the dimension-reduced data in the training set input to the first autoencoder A1 and the second autoencoder A2; and are the reconstruction error data of the first autoencoder A1 and the second autoencoder A2 respectively; the abnormal reconstruction loss of the abnormal autoencoder is calculated, and the calculation formula is: Among them, D drop Refers to high error data, Refers to the reconstruction error data of the abnormal self-encoder C; calculate the total objective function, the calculation formula is: L total =L recon (A1,A2)+λL recon (C), where λ is a hyperparameter; the overall objective function is input into the anomaly detection model for optimization;

[0096] The detection unit 303, if receiving the medical insurance claim data to be processed, detects the medical insurance claim data to be processed through the anomaly detection model to obtain the corresponding anomaly score, and marks the risk of the medical insurance claim data to be processed according to the anomaly score and the preset risk threshold; obtains the medical insurance claim data to be processed, and inputs it into the anomaly detection model to obtain the reconstruction error; performs weighted summation on the reconstruction error to generate the anomaly score; determines the size of the anomaly score and the preset risk threshold to mark the risk type of the medical insurance claim data to be processed; marks the medical insurance claim data to be processed as normal data or abnormal data according to the risk type and in combination with the business rule library; if the anomaly score is greater than the first risk threshold, it is marked as high risk; if the anomaly score is less than the first risk threshold and greater than the second risk threshold, it is marked as medium risk; if the anomaly score is less than the second risk threshold and greater than the third risk threshold, it is marked as low risk.

[0097] It should be noted that technical personnel in the relevant field can clearly understand that the specific implementation process of the above-mentioned medical insurance fraud detection device based on autoencoder and each unit can refer to the corresponding description in the aforementioned method embodiment. For the convenience and conciseness of the description, it will not be repeated here.

[0098] The above-mentioned medical insurance fraud detection device 300 based on autoencoder can be implemented in the form of a computer program. The computer program can be used in Figure 7 Runs on the computer equipment shown.

[0099] See also Figure 7 , Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 500 can be a terminal or a server. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, personal digital assistant, wearable device, or other electronic device with communication capabilities. The server can be a standalone server or a server cluster consisting of multiple servers.

[0100] See Figure 7 The computer device 500 includes a processor 502 , a memory, and a network interface 505 connected via a system bus 501 , wherein the memory may include a non-volatile storage medium 503 and an internal memory 504 .

[0101] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions, which, when executed, can cause the processor 502 to execute a medical insurance fraud detection method based on an autoencoder.

[0102] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500.

[0103] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a medical insurance fraud detection method based on an autoencoder.

[0104] The network interface 505 is used to communicate with other devices through the network. Figure 7 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 500 to which the solution of the present application is applied. The specific computer device 500 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0105] The processor 502 is configured to run a computer program 5032 stored in the memory to implement the steps of the above-mentioned medical insurance fraud detection method based on an autoencoder.

[0106] It should be understood that in the embodiment of the present application, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0107] Those skilled in the art will appreciate that all or part of the steps in the method of the above-described embodiment can be implemented by instructing the relevant hardware through a computer program. The computer program includes program instructions, which can be stored in a storage medium that is computer-readable. The program instructions are executed by at least one processor in the computer system to implement the steps in the method of the above-described embodiment.

[0108] Therefore, the present invention also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the steps of the aforementioned autoencoder-based medical insurance fraud detection method.

[0109] The storage medium may be any computer-readable storage medium that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk.

[0110] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0111] In the several embodiments provided herein, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the various units is merely a logical functional division, and actual implementation may employ other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be omitted or not implemented.

[0112] The steps in the methods of the embodiments of the present invention may be adjusted in order, combined, or deleted as needed. The units in the devices of the embodiments of the present invention may be combined, divided, or deleted as needed. Furthermore, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit.

[0113] If this integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, terminal, or network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present invention.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A medical insurance fraud detection method based on an autoencoder, characterized in that: Applied to the medical insurance system, the method includes: Obtain a medical insurance claim dataset, and perform structural processing on the initial data in the medical insurance claim dataset to obtain a reduced dimension dataset; Dividing the dimension-reduced data in the dimension-reduced data set into a training set in proportion; Inputting the dimension-reduced data in the training set into the autoencoder to train it to obtain an anomaly detection model; If unprocessed medical insurance claim data is received, the unprocessed medical insurance claim data is detected using the anomaly detection model to obtain a corresponding anomaly score, and the unprocessed medical insurance claim data is risk-marked based on the anomaly score and a preset risk threshold.

2. The method according to claim 1, characterized in that The autoencoder includes a first autoencoder and a second autoencoder, and inputting the dimension-reduced data in the training set into the autoencoder to train the autoencoder to obtain an anomaly detection model includes: The dimension-reduced data in the training set is input into the first autoencoder and the second autoencoder for training.

3. The method according to claim 2, characterized in that The structured processing of the initial data in the medical insurance claim dataset to obtain a reduced dimension dataset includes: The principal component analysis method and principal component extraction method are used to perform structured processing on the initial data to obtain a dimension-reduced data set.

4. The method according to claim 2, characterized in that The autoencoder further includes an anomaly autoencoder, and inputting the dimension-reduced data in the training set into the first autoencoder and the second autoencoder to train them to obtain an anomaly detection model includes: Inputting the dimension-reduced data in the training set into the first autoencoder and the second autoencoder to obtain decoded and reconstructed data; Performing reconstruction error calculation on the decoded reconstructed data and the dimensionality reduced data to obtain reconstruction error data; Sorting the reconstructed error data, and selecting data whose reconstructed error data is greater than a preset error as high-error data according to an abnormality ratio; The high-error data is input into the abnormal autoencoder to enable it to learn.

5. The method according to claim 4, characterized in that The step of inputting the dimension-reduced data in the training set into the autoencoder to train the autoencoder to obtain an anomaly detection model further includes: Divide the training set into several small training batches, each of which has a size of B; Calculate the basic reconstruction loss of the first autoencoder and the second autoencoder using the following formula: Wherein, x is the dimension-reduced data in the training set input to the first autoencoder A1 and the second autoencoder A2; and The reconstructed error data of the first autoencoder A1 and the second autoencoder A2 respectively; The abnormal reconstruction loss of the abnormal autoencoder is calculated as follows: Among them, D drop Refers to high error data, Refers to the reconstructed error data of the abnormal autoencoder C; Calculate the total objective function, the calculation formula is: L total =L recon (A1,A2)+λL recon (C), Among them, λ is a hyperparameter; The overall objective function is input into the anomaly detection model for optimization.

6. The method according to claim 1, characterized in that If the unprocessed medical insurance claim data is received, the unprocessed medical insurance claim data is detected by the anomaly detection model to obtain a corresponding anomaly score, and the unprocessed medical insurance claim data is risk-marked according to the anomaly score and a preset risk threshold, including: Obtaining the to-be-processed medical insurance claim data and inputting the data into the anomaly detection model to obtain a reconstruction error; Performing weighted summation on the reconstruction errors to generate the anomaly score; Determining the difference between the abnormality score and the preset risk threshold to mark the risk type of the medical insurance claim data to be processed; According to the risk type and in combination with the business rule library, the to-be-processed medical insurance claim data is marked as normal data or abnormal data.

7. The method according to claim 6, characterized in that The preset risk thresholds include a first risk threshold, a second risk threshold, and a third risk threshold, the first risk threshold is greater than the second risk threshold, and the second risk threshold is greater than the third risk threshold. The determining of the difference between the abnormality score and the preset risk thresholds to mark the risk type of the to-be-processed medical insurance claim data includes: If the anomaly score is greater than the first risk threshold, it is marked as high risk; If the anomaly score is less than the first risk threshold and greater than the second risk threshold, it is marked as medium risk; If the anomaly score is less than the second risk threshold and greater than the third risk threshold, it is marked as low risk.

8. A medical insurance fraud detection device based on an autoencoder, characterized in that: The method comprises a unit for executing the method according to any one of claims 1 to 7.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the method according to any one of claims 1 to 7 can be implemented.

Images