Medical insurance abnormal data detection method and device based on unsupervised learning

Abstract

The invention provides a medical insurance abnormal data detection method and device based on unsupervised learning, and the method comprises the steps: inputting each piece of obtained to-be-detected medical insurance data into an abnormal data detection model, so as to output an output result corresponding to the to-be-detected medical insurance data; wherein the abnormal data detection model comprises a selector and a plurality of network modules based on variational auto-encoders; the module parameters of the network modules are different, and the network modules correspond to different abnormal data types; the selector is used for selecting a network module matched with the to-be-detected medical insurance data from a plurality of network modules; based on each output result, detecting medical insurance abnormal data from the multiple pieces of to-be-detected medical insurance data; according to the method, the network module matched with the to-be-detected medical insurance data can be automatically selected from the multiple network modules contained in the abnormal data detection model according to the to-be-detected medical insurance data, flexible identification of different abnormal data types is achieved, and the efficiency and accuracy of medical insurance abnormal data detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for detecting abnormal medical insurance data based on unsupervised learning. Background Art

[0002] With the expansion of medical insurance service coverage and the increasing complexity of medical services, the medical insurance system faces severe challenges in management and review, especially in the detection of anomalies in medical insurance data. There are two main traditional detection methods: one is to rely on manual review, which usually requires screening suspected abnormal data first, and then conducting manual interviews one by one to verify. The other is a supervised learning method, which trains models with labeled samples to identify potential abnormal patterns in the data. The large amount of medical insurance data and the scarcity of violation labels make supervised learning's reliance on a large amount of high-quality labeled data a major challenge. Medical insurance data are mostly unstructured medical records, drug lists, and inspection reports, and involve privacy protection and compliance requirements, making large-scale data labeling very difficult. In summary, the efficiency and accuracy of related technologies in detecting abnormal medical insurance data are poor. Summary of the invention

[0003] The purpose of the present invention is to provide a method and device for detecting abnormal medical insurance data based on unsupervised learning, so as to improve the efficiency and accuracy of detecting abnormal medical insurance data.

[0004] The present invention provides a medical insurance abnormal data detection method based on unsupervised learning, the method comprising: obtaining multiple medical insurance data to be detected; for each medical insurance data to be detected, inputting the medical insurance data to be detected into a pre-trained abnormal data detection model, and outputting an output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model comprises: a selector and multiple network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module matching the medical insurance data to be detected from multiple network modules; based on each output result, detecting the medical insurance abnormal data from the multiple medical insurance data to be detected.

[0005] Furthermore, each medical insurance data to be detected is in vector form; a preset weight matrix is ​​pre-stored in the selector; the input end of the selector is the input end of the abnormal data detection model, and the output end of the selector is connected to the input end of each network module; for each medical insurance data to be detected, the medical insurance data to be detected is input into a pre-trained abnormal data detection model, and the step of outputting the output result corresponding to the medical insurance data to be detected through the abnormal data detection model includes: for each medical insurance data to be detected, the medical insurance data to be detected is input into the selector, and the product of the medical insurance data to be detected and the preset weight matrix is ​​calculated through the selector to obtain multiple weighted results; wherein the multiple weighted results correspond to the multiple network modules one by one; the multiple weighted results are sorted in order from large to small to obtain weighted sorting results; the first preset number of weighted results in the weighted sorting results are determined as multiple target weighted results; based on the multiple target weighted results, the output result corresponding to the medical insurance data to be detected is determined.

[0006] Furthermore, based on multiple target weighted results, the step of determining the output result corresponding to the medical insurance data to be tested includes: normalizing the multiple target weighted results to obtain multiple probability results; activating each target network module corresponding to each probability result, and inputting the medical insurance data to be tested into each target network module, and outputting the corresponding first result through each target network module; multiplying the probability result corresponding to each target network module with the corresponding first result to obtain the first product result corresponding to each target network module; adding each first product result to obtain the output result corresponding to the medical insurance data to be tested.

[0007] Furthermore, based on each output result, the step of detecting abnormal medical insurance data from multiple medical insurance data to be detected includes: for each piece of medical insurance data to be detected, based on the output result corresponding to the medical insurance data to be detected and the medical insurance data to be detected, calculating the reconstruction error corresponding to the medical insurance data to be detected; according to the reconstruction errors corresponding to each piece of medical insurance data to be detected, detecting abnormal medical insurance data from multiple medical insurance data to be detected.

[0008] Further, according to the reconstruction error corresponding to each piece of medical insurance data to be detected, the step of detecting abnormal medical insurance data from the plurality of medical insurance data to be detected includes: sorting the reconstruction error corresponding to each piece of medical insurance data to be detected in descending order to obtain a sorting result; obtaining the data quantity and a preset abnormal proportion value of the medical insurance data to be detected; Calculate the product of the number of data and the abnormal proportion value to obtain the product result, and determine the product result as the target sequence number; determine the reconstruction error corresponding to the target sequence number in the sorting result as the abnormal threshold; determine the medical insurance data to be tested whose reconstruction error is greater than the abnormal threshold as medical insurance abnormal data.

[0009] Furthermore, the abnormal data detection model is trained in the following manner: obtaining sample medical insurance data; inputting the sample medical insurance data into the initial model to output the processing results corresponding to the sample medical insurance data through the initial model; wherein the initial model includes: an initial selector and multiple initial modules based on variational autoencoders; based on a preset loss function, calculating an auxiliary loss value, and a first loss value between the processing result and the sample medical insurance data; wherein the auxiliary loss value is used to: indicate the degree of difference in the importance of each initial module; training the initial model based on the auxiliary loss value and the first loss value to obtain a trained abnormal data detection model.

[0010] Furthermore, a sample weight matrix is ​​pre-stored in the initial selector; the auxiliary loss value is determined in the following manner: the sample medical insurance data is input into the initial selector, and the product of the sample medical insurance data and the sample weight matrix is ​​calculated by the initial selector to obtain multiple sample weighted results; based on the multiple sample weighted results, the standard deviation and the average are determined; based on the preset adjustment parameters, the standard deviation and the average, the auxiliary loss value is calculated.

[0011] Furthermore, the step of obtaining multiple medical insurance data to be tested includes: obtaining multiple original medical insurance data; pre-processing each original medical insurance data according to a preset processing method to obtain multiple medical insurance data to be tested; wherein each medical insurance data to be tested is in vector form; the preset processing method includes at least one of the following: data screening processing, data extraction processing, missing value processing, normalization processing and text encoding processing.

[0012] The present invention provides a medical insurance abnormal data detection device based on unsupervised learning, and the device includes: an acquisition module, used to acquire multiple medical insurance data to be detected; an output module, used to input the medical insurance data to be detected into a pre-trained abnormal data detection model for each medical insurance data to be detected, and output the output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and multiple network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from multiple network modules; and the detection module is used to detect the medical insurance abnormal data from the multiple medical insurance data to be detected based on each output result.

[0013] An electronic device provided by the present invention includes a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement any of the above-mentioned medical insurance abnormal data detection methods based on unsupervised learning.

[0014] The present invention provides a method, device and electronic device for detecting abnormal medical data based on unsupervised learning, the method comprising: obtaining multiple medical insurance data to be detected; for each of the medical insurance data to be detected, inputting the medical insurance data to be detected into a pre-trained abnormal data detection model, and outputting the output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model comprises: a selector and multiple network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from the multiple network modules; based on each of the output results, the medical insurance abnormal data is detected from the multiple medical insurance data to be detected. This method can automatically select a network module that matches the medical insurance data to be detected from the multiple network modules included in the abnormal data detection model according to the medical insurance data to be detected, realize flexible recognition of different abnormal data types, and improve the efficiency and accuracy of medical insurance abnormal data detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A schematic diagram of a traditional detection method provided by an embodiment of the present invention; Figure 2 A flowchart of a method for detecting abnormal medical insurance data based on unsupervised learning provided by an embodiment of the present invention; Figure 3 A schematic diagram of a processing process of an abnormal data detection model provided by an embodiment of the present invention; Figure 4 A schematic diagram of a medical insurance abnormal data detection system based on unsupervised learning provided by an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a medical insurance abnormal data detection device based on unsupervised learning provided by an embodiment of the present invention; Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] like Figure 1 The diagram of a traditional detection method shown in the figure shows that there are two main traditional detection methods for medical insurance data anomaly detection: one is to rely on manual review, which requires data screening of medical insurance data first, manual review, and then interview confirmation; the other is a supervised learning method, which requires data labeling first, supervised learning, and then using the supervised learning model to judge anomalies; however, for the method that relies on manual review, the entire process of this method is not only labor-intensive, inefficient, time-consuming and labor-intensive, but also prone to inaccurate detection, missed judgments and misjudgments, making it difficult to meet the needs of efficient management. The specific disadvantages are as follows: 1. Low approval efficiency: The manual review process is cumbersome and it is difficult to quickly process a large amount of data, resulting in a long review cycle; 2. Highly subjective: the review criteria may vary depending on the experience and judgment of the auditors, which may easily lead to missed judgments and misjudgments; 3. High labor cost: A large amount of human resources are required, resulting in high operating costs; 4. Limited data processing capabilities: Lack of efficient data processing and analysis tools affects the efficiency and accuracy of the audit.

[0019] As for supervised learning, as an emerging anomaly detection method, supervised learning methods rely on the learning mechanism of labeled data and show potential in identifying abnormal medical insurance behavior. Unlike the traditional method that relies on manual rules, supervised learning can train models through labeled samples and identify potential abnormal patterns in data, thus playing a role in insurance fraud prevention and control. However, supervised learning has the following disadvantages: 1. Dependence on labeled data: A large amount of high-quality labeled data is required for training, but violation labels in medical insurance data are scarce and difficult to obtain. This reality limits the practical application of supervised learning in insurance fraud detection and makes it difficult to meet the needs of efficiency and accuracy.

[0020] 2. High training cost: Data labeling and model training are time-consuming and labor-intensive, and require the participation of professionals; 3. Insufficient generalization ability: It may show poor recognition ability in the face of new and unlabeled abnormal patterns; 4. Privacy and compliance risks: When processing medical insurance data, privacy protection and compliance requirements are involved, which limit the use and analysis of data.

[0021] Based on this, an embodiment of the present invention provides a method and device for detecting abnormal medical insurance data based on unsupervised learning. This technology can be applied to applications that require detection of abnormal medical insurance data.

[0022] To facilitate understanding of this embodiment, firstly, a method for detecting abnormal medical insurance data based on unsupervised learning disclosed in an embodiment of the present invention is introduced. Figure 2 As shown, the method comprises the following steps: Step S202, obtaining multiple pieces of medical insurance data to be tested; Each of the above-mentioned medical insurance data to be tested usually contains dozens of obtainable information fields; in actual implementation, when it is necessary to detect whether there is abnormal medical insurance data, it is usually necessary to obtain multiple medical insurance data to be tested first; for example, multiple medical insurance data to be tested can be obtained through hospitals and other institutions or platforms.

[0023] Step S204: for each piece of medical insurance data to be detected, the medical insurance data to be detected is input into a pre-trained abnormal data detection model, and the output result corresponding to the medical insurance data to be detected is output through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and a plurality of network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from the plurality of network modules; The above-mentioned abnormal data types may include: fictitious non-existent medical expenses or medical services, forged medical documents and receipts, etc.; in actual implementation, the abnormal data detection model usually includes a selector and multiple network modules connected to it, and the network module can also be called an expert module. Each network module can have the same variational autoencoder structure, but the module parameters of each network module are different. By setting different module parameters for each network module, each network module can correspond to a different abnormal data type; in actual implementation, after obtaining multiple medical insurance data to be detected, for each medical insurance data to be detected, the medical insurance data to be detected is input into a pre-trained abnormal data detection model. Through the selector in the abnormal data detection model, a network module matching the medical insurance data to be detected can be selected, and then the output result corresponding to the medical insurance data to be detected can be output.

[0024] Step S206: based on each output result, abnormal medical insurance data is detected from the multiple pieces of medical insurance data to be detected.

[0025] According to the output results corresponding to each medical insurance data to be tested, abnormal medical insurance data can be detected from multiple medical insurance data to be tested. For example, if the difference between a certain medical insurance data to be tested and the output result is greater than a certain threshold, then the medical insurance data to be tested can be considered as abnormal medical insurance data, etc.

[0026] The above-mentioned medical insurance abnormal data detection method based on unsupervised learning obtains multiple medical insurance data to be detected; for each medical insurance data to be detected, the medical insurance data to be detected is input into a pre-trained abnormal data detection model, and the output result corresponding to the medical insurance data to be detected is output through the abnormal data detection model; wherein, the abnormal data detection model includes: a selector and multiple network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from multiple network modules; based on each output result, the medical insurance abnormal data is detected from multiple medical insurance data to be detected. This method can automatically select a network module that matches the medical insurance data to be detected from the multiple network modules contained in the abnormal data detection model, realize flexible recognition of different abnormal data types, and improve the efficiency and accuracy of medical insurance abnormal data detection.

[0027] The embodiment of the present invention also provides another method for detecting abnormal medical insurance data based on unsupervised learning. The method is implemented on the basis of the method of the above embodiment. In the method, each piece of medical insurance data to be detected is in vector form; for example, if the medical insurance data to be detected is numerical data, the minimum-maximum normalization method can be used to scale the medical insurance data to be detected so that the numerical value is within the range of [-1,1] to obtain normalized data, and the normalized data can be represented as a vector; if the medical insurance data to be detected is unstructured text, Word2Vec (a technology for converting words into fixed-length vectors) and other methods can be used to encode the medical insurance data to be detected to convert it into a vector form; a preset weight matrix is ​​pre-stored in the selector; the input end of the selector is the input end of the abnormal data detection model, and the output end of the selector is connected to the input end of each network module; the method includes the following steps: Step 1: Obtain multiple pieces of medical insurance data to be tested; Step 2: for each piece of medical insurance data to be detected, input the medical insurance data to be detected into a selector, and calculate the product of the medical insurance data to be detected and a preset weight matrix through the selector to obtain multiple weighted results; wherein the multiple weighted results correspond to the multiple network modules one by one; In actual implementation, since each piece of medical insurance data to be tested is in vector form, for each piece of medical insurance data to be tested, after the medical insurance data to be tested is input into the abnormal data detection model, it can be first multiplied by the preset de-weighting matrix stored in the selector in the abnormal data detection model, and multiple weighted results are output through the selector, each weighted result corresponds to a network module, that is, the number of weighted results is the same as the number of network modules.

[0028] Step 3, sorting the multiple weighted results in descending order to obtain a weighted sorting result; Step 4: determining a preset number of weighted results in the weighted sorting results as a plurality of target weighted results; The above preset number can be set according to actual needs; in actual implementation, the values ​​of multiple weighted results output by the selector are usually different, and the multiple weighted structures can be sorted in order from large to small to obtain a weighted sorting result; in this embodiment, the Top-K strategy can be adopted, that is, the top K weighted results with the highest values ​​among multiple weighted results are selected as the target weighted results, where K is an artificially given hyperparameter.

[0029] Step 5: Based on the weighted results of multiple targets, determine the output result corresponding to the medical insurance data to be detected; wherein the abnormal data detection model includes: a selector and multiple network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from multiple network modules; Since each weighted result corresponds to a one-to-one network module, the corresponding network module can be selected according to the above target weighted result, and then the output result corresponding to the medical insurance data to be tested is determined.

[0030] The following is an explanation of the relevant principles of the network module of the variational autoencoder: Since a large amount of medical insurance data is unlabeled data, in order to detect insurance fraud without relying on abnormal labels, this solution chooses to use a network module based on a variational autoencoder to learn the latent variable representation of the data through an encoding-decoding structure and generate new samples that conform to the data distribution. In order to achieve this goal, it is necessary to maximize the marginal likelihood, that is, given a data sample The probability of the model generating a new sample when , the mathematical definition of marginal likelihood is:

[0031] in, represents the marginal likelihood, is a hidden variable, is the joint distribution of the generative model, It is a posteriori The approximate distribution of .

[0032] Since it is difficult to directly calculate the marginal likelihood, the variational lower bound (EvidenceLower Bound, ELBO) is used in the variational autoencoder to approximate it. Its form is: ; Among them, the first Represents the reconstruction error, which is used to measure the The accuracy of the generated samples, Indicates that in the intermediate hidden vector Based on The probability of is the Kullback-Leibler divergence, which is used to quantify the approximate posterior With prior The difference between . Optimizing this lower bound can be used as the goal of the variational autoencoder. Taking the variational lower bound as the loss function, the loss function of the network module based on the variational autoencoder can be expressed as: ; In the process of code implementation, this solution chooses to use Fully connected Neural Network (FCNN) for encoding and decoding, and Function as activation function and applied to and .

[0033] ; If we directly get the normal distribution Since random sampling is not differentiable, the model cannot be back-propagated and cannot be trained using the existing deep learning framework. Therefore, the reparameterization technique is used in the implementation to reparameterize the random sampling process into a deterministic mapping, namely: ; In this way, the sampling process becomes differentiable, thus achieving gradient-based end-to-end optimization.

[0034] When decoding, two fully connected neural networks are also used to obtain the distribution that the reconstructed samples satisfy The mean and variance of : ; Used with encoder The activation function is different and is used in the decoder The activation function makes all outputs positive and smoother. Finally, the reparameterization technique is used again to obtain the reconstructed output samples: ; For details about the relevant principles of the network module of the above-mentioned variational autoencoder, please refer to the relevant technology and will not be repeated here.

[0035] Specifically, step five can be implemented by following steps 50 to 53: Step 50, normalizing the multiple target weighted results to obtain multiple probability results; In actual implementation, the values ​​of these K target weighted results can be kept unchanged, and the values ​​of the other weighted results can be set to negative infinity. Then, after the softmax operation, the probability of the network module whose weighted result value is set to negative infinity being selected will drop to 0. Thus, the network module selection distribution can be obtained The formula is: ; ; in, Indicates the medical insurance data to be tested. Represents the preset weight matrix inside the selector, Represents the weighted result of K targets; Represents the K probability results after normalizing the K target weighted results.

[0036] The probability distribution of selecting different target network modules is obtained through softmax operation . The softmax is an activation function that can normalize a numerical vector into a probability distribution vector, and the sum of all probabilities is 1. Its formula is as follows: ; in, represents the i-th element in the input vector, which corresponds to the i-th target weighted result in this embodiment; n is the length of the vector, which corresponds to the number K of target weighted results in this embodiment; For example, the three target weighted results are normalized and the probability results corresponding to the three target weighted results are: 0.7, 0.2 and 0.1. These three probability results correspond to the probability of selecting each target network module.

[0037] Step 51, activating each target network module corresponding to each probability result, and inputting the to-be-detected medical insurance data into each target network module, and outputting the first result corresponding to each target network module respectively; Activate each target network module corresponding to each probability result, and according to the distribution The selected K target network modules are , then the medical insurance data to be tested is assigned to the selected K target network modules for processing, and each target network module outputs the corresponding first result respectively.

[0038] Step 52, multiplying the probability result corresponding to each target network module by the corresponding first result to obtain the first product result corresponding to each target network module; Step 53, adding each first multiplication result to obtain the output result corresponding to the medical insurance data to be tested.

[0039] For each target network module, the probability result corresponding to the target network module is multiplied by the first result to obtain the first product result corresponding to the target network module, and the first product results of each target network module are added together to obtain the output result corresponding to the medical insurance data to be detected. It can be expressed by the following formula: ; in, Represents the probability result corresponding to the i-th target network module selected by the selector, It indicates the first result of the output corresponding to the selected i-th target network module; Indicates the output result corresponding to the medical insurance data to be tested; Indicates the number of the target network module.

[0040] For easier understanding, see Figure 3 The figure shows a processing diagram of an abnormal data detection model that uses a mixture of experts (MoE) strategy. The mixture of experts strategy is a method of dynamically selecting network substructures. It regards different subnetworks in the model as different "experts" and allocates computing resources to different "experts" through comprehensive scheduling, so that the model can learn richer information. The mixture of experts strategy can significantly improve computing efficiency and model performance when processing large-scale data and tasks.

[0041] The hybrid expert strategy usually consists of two parts: a selector and an expert module (corresponding to the above network module). The selector selects the most matching expert module for a given input, and the expert module is used to allocate and execute actual data processing operations. Figure 3As shown in the figure, its working process is as follows: first, use the selector to generate the probability distribution of each VAE layer (Variational Autoencoder) being selected, that is, the probability of each expert module being selected, and select the appropriate VAE layer for data processing. Finally, multiply the output distribution of the selector by the output of the selected VAE layer, and add the results to get the final output.

[0042] Step 6: for each piece of medical insurance data to be detected, based on the output result corresponding to the medical insurance data to be detected and the medical insurance data to be detected, calculate the reconstruction error corresponding to the medical insurance data to be detected; After calculating the output result corresponding to each piece of medical insurance data to be tested, the reconstruction error corresponding to the medical insurance data to be tested can be calculated according to the following formula: ; in, represents the reconstruction error corresponding to the i-th medical insurance data to be tested; Indicates the output result corresponding to the i-th medical insurance data to be tested; represents the i-th medical insurance data to be detected that is input into the abnormal data detection model; N is the number of medical insurance data to be detected.

[0043] Step 7: Detect abnormal medical insurance data from multiple pieces of medical insurance data to be detected based on the reconstruction error corresponding to each piece of medical insurance data to be detected. The step seven can be implemented by following the steps 70 to 74: Step 70, sorting the reconstruction errors corresponding to each piece of medical insurance data to be detected in descending order to obtain a sorting result; The reconstruction errors corresponding to each piece of medical insurance data to be tested are sorted in descending order to obtain the sorting results as follows: ; Step 71, obtaining the data quantity of the medical insurance data to be detected and a preset abnormal ratio value; Step 72, calculating the product of the data quantity and the abnormal ratio value to obtain the product result, and determining the product result as the target sequence number; The above data quantity can be represented by N; the above abnormal proportion value can be represented by It can be set according to actual needs. In actual implementation, after obtaining the data quantity and the abnormal proportion value, the product of the two can be calculated and the product result can be used as the target sequence number. For example, N is 100, is 5%, then It is 5, which is the target sequence number.

[0044] Step 73, determining the reconstruction error corresponding to the target sequence number in the sorting result as the abnormal threshold; Step 74: determine the medical insurance data to be detected whose reconstruction error is greater than the abnormal threshold as abnormal medical insurance data.

[0045] In actual implementation, the reconstruction error corresponding to the target sequence number in the sorting result can be used as the abnormal threshold. When the reconstruction error is greater than the abnormal threshold, the corresponding medical insurance data to be detected is determined as abnormal data. Specifically, since the sorting result is obtained by sorting the reconstruction error from large to small, the medical insurance data to be detected corresponding to the reconstruction error before the target sequence number will be determined as medical insurance abnormal data. Specifically, the abnormal threshold T can be obtained according to the following expression: ; in, represents the reconstruction error corresponding to the target number i; represents the abnormal proportion value; N represents the number of data.

[0046] For easier understanding, see Figure 4 The figure shows a schematic diagram of a medical insurance abnormal data detection system based on unsupervised learning, which shows the overall application framework of this solution, mainly including a selector and multiple expert modules based on variational autoencoders (corresponding to the above-mentioned network modules). First, the original medical insurance data is preprocessed to obtain the medical insurance data to be detected encoded in the form of a vector, and the medical insurance data to be detected is input into the selector. The selector calculates the weighted results of each expert module based on the characteristics of the vector corresponding to the medical insurance data to be detected, and activates the two expert modules with the highest values ​​in the weighted results (such as the two gray expert modules shown in the figure). These activated expert modules will receive the input medical insurance data to be detected, while other unactivated expert modules will not participate in the processing.

[0047] Each expert module has the same variational autoencoder structure. The input vector-based medical insurance data to be tested is first compressed by the encoder to generate the mean and variance of the normal distribution, and then randomly sampled on this distribution to obtain the intermediate layer hidden vector. In order to retain part of the original input information, the expert module directly adds the original input to the intermediate layer hidden vector through conditional connection. Then, the hidden vector is reconstructed by the decoder, and the reconstructed vector is output.

[0048] After obtaining the reconstructed output of each expert module, the selector assigns a corresponding weight to the output vector of each expert module, performs weighted product on the output of each expert module, and obtains the first product results corresponding to each. Finally, multiple first product results are added together to obtain the final output of the abnormal data detection model. This architecture enables the abnormal data detection model to adaptively select expert modules according to data features, improving the flexibility and accuracy of processing complex data.

[0049] When using the hybrid expert strategy, in addition to the variational autoencoder network (VAE) used in this solution, the internal expert module can also use other network types suitable for unsupervised learning, such as the classic autoencoder (AE) and conditional variational autoencoder (CVAE). Both of the above methods and VAE use an encoding-decoding structure to generate new samples that are as similar to the original input as possible by learning the latent variable representation and distribution of the data, thereby realizing the detection of insurance fraud without relying on abnormal labels. This solution can be used as an alternative to using hybrid expert variational autoencoders for unsupervised anomaly detection of medical insurance data.

[0050] The above-mentioned medical insurance abnormal data detection method based on unsupervised learning can automatically select a matching network module from multiple network modules included in the abnormal data detection model according to the medical insurance data to be detected, thereby realizing flexible recognition of different abnormal data types and improving the efficiency and accuracy of medical insurance abnormal data detection.

[0051] The embodiment of the present invention also provides another method for detecting abnormal medical insurance data based on unsupervised learning. The method is implemented on the basis of the method in the above embodiment, and the method includes the following steps: Step 11, obtaining multiple pieces of original medical insurance data; Step 12: pre-process each piece of original medical insurance data according to a preset processing method to obtain multiple pieces of medical insurance data to be tested; wherein each piece of medical insurance data to be tested is in vector form; the preset processing method includes at least one of the following: data screening processing, data extraction processing, missing value processing, normalization processing and text encoding processing.

[0052] The above-mentioned data screening process can be to screen multiple pieces of original medical insurance data in the database according to pre-set screening conditions to select the excess data that meets the screening conditions; for example, the screening condition can be "the annual medical insurance amount exceeds 20,000", which can screen out the medical insurance data to be tested with an annual medical insurance amount exceeding 20,000. The above-mentioned data extraction process can be to extract relevant field data from the screening results after the data screening process, including the patient ID (Identity Document), diagnosis ID, settlement ID, gender, age, visit time, hospital, diagnosis information, total medical expenses, amount within the scope, drug diagnosis and treatment services, etc.; among them, the data in the screening results may contain multiple fields, some of which are irrelevant to detecting whether the medical insurance data is abnormal, and the field data of concern can be screened out from multiple fields through data extraction processing. The above-mentioned missing value processing can be used to process the missing values ​​in the relevant field data obtained after data extraction processing, including identifying missing values, deleting missing values ​​and filling missing values; it can also query and supplement from the personnel information data table according to the missing data attributes, delete the entries of missing transaction data, etc.; for example, if a field in a certain medical insurance data to be tested lacks the corresponding field value, the average value can be calculated according to the existing values ​​corresponding to the field in other medical insurance data to be tested to determine the missing value to be filled; if there are many missing values ​​in a certain medical insurance data to be tested, it is generally considered that the medical insurance data to be tested is invalid and can be deleted; the above-mentioned missing data attributes can be category data, numerical data, etc., where the numerical data can be a specific value; category data is used to distinguish different categories, for example, 0 and 1 represent different genders, etc. For different missing data attributes, a suitable way to fill missing values ​​can be selected. The above-mentioned normalization processing can be understood as using the minimum-maximum normalization method for numerical data to scale the data so that the value is within the range of [-1,1]. The above-mentioned text encoding processing can be understood as encoding unstructured text using Word2Vec and other methods for unstructured text.

[0053] In actual implementation, after obtaining multiple pieces of original medical insurance data, when it is necessary to use an abnormal data detection model for anomaly detection, if the data volume of the multiple pieces of original medical insurance data is relatively large, the original medical insurance data can be randomly sampled according to a certain proportion, for example, 20% of the multiple pieces of original medical insurance data can be selected, and then processed according to the above preset processing method to obtain multiple pieces of medical insurance data to be tested.

[0054] Step 13: for each piece of medical insurance data to be detected, input the medical insurance data to be detected into a pre-trained abnormal data detection model, and output the output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and multiple network modules based on variational autoencoders; each network module has different module parameters, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from multiple network modules; The abnormal data detection model is trained by following steps A to D: Step A, obtaining sample medical insurance data; In order to train a model that can effectively identify insurance fraud, it is first necessary to collect training data for this task to improve the classification performance of the model. The amount of raw data for the task of discovering medical insurance fraud is extremely large. According to statistics, in actual production scenarios, more than 100 million pieces of data to be classified will be generated each year. For example, in 2023, about 150 million transaction details that need to be classified will be generated throughout the year, and each piece of data contains dozens of accessible information fields. Due to resource constraints, it is impossible for all of this data to be used for model training. Therefore, data cleaning, sampling, and analysis are particularly important.

[0055] In actual implementation, multiple indicators that affect the test results can be pre-set, and the data can be analyzed through correlation analysis methods such as PCA (Principal Component Analysis). Through data screening, data extraction, missing value processing, normalization, text encoding, etc., all the medical information of the patients arranged in chronological order within a time window is finally selected as the input of the initial model. For example, in the transaction details data for the whole year of 2022, according to the positive and negative sample ratio of one to five, a total of 119,017 positive and negative sample training sets and 13,527 validation sets are sampled for model training and verification of the final model.

[0056] Step B, inputting the sample medical insurance data into the initial model, so as to output the processing results corresponding to the sample medical insurance data through the initial model; wherein the initial model includes: an initial selector and a plurality of initial modules based on variational autoencoders; The above-mentioned variational autoencoder is a type of deep learning method based on generative models, which aims to learn the potential distribution of data so as to generate new samples similar to the training data.

[0057] Step C, based on a preset loss function, calculating an auxiliary loss value and a first loss value between the processing result and the sample medical insurance data; wherein the auxiliary loss value is used to: indicate the degree of difference in importance of each initial module; The above preset loss function can be expressed as follows: ; in, is a reconstruction loss function used to calculate the first loss value; is the auxiliary loss function used to calculate the auxiliary loss value.

[0058] For unsupervised anomaly detection tasks, the reconstruction loss function can be used As one of the training tasks, it aims to reduce the difference between the reconstructed samples and the original sample medical insurance data, so that the model can better capture the characteristics of the original sample medical insurance data. Specifically, it can be expressed as follows: ; in, Indicates the above processing results, Represents sample medical insurance data.

[0059] The sample weight matrix is ​​pre-stored in the initial selector; the auxiliary loss value is determined by following steps a to c: Step a, inputting the sample medical insurance data into the initial selector, and calculating the product of the sample medical insurance data and the sample weight matrix through the initial selector to obtain multiple sample weighted results; Step b, determining the standard deviation and the average value based on the weighted results of multiple samples; Step c: based on the preset adjustment parameters, standard deviation and average value, the auxiliary loss value is calculated.

[0060] In actual training, in order to avoid a certain initial module being repeatedly activated while other initial modules are not fully trained, this scheme introduces an auxiliary loss function to force each initial module to have the same importance during the training process. The auxiliary loss function calculation formula is: ; ; in, It can be understood as the weighted result of the samples of each initial module; the first component of the auxiliary loss function The weighted results of the samples of each initial module in the entire batch are summed up, which can be regarded as the importance score of each initial module, that is, the probability of each initial module being selected without considering the input. These importance scores can be used to calculate the coefficient of variation (CV), which indicates the degree of difference between the importance scores of each initial module, defined as the standard deviation of the weighted results of multiple samples With the average The ratio is used to measure the discreteness of the weighted results of this group of samples. The calculation formula is as follows: ; If all initial modules have similar scores, the CV value is low, which is the ideal situation. Through the CV value, the auxiliary loss value can be updated during training to minimize the CV value, so that each initial module has the same importance. The auxiliary loss function The first parameter is a hyperparameter given by an artificial person. It is a measure of the importance of all initial modules having equal weights. For example, it can take any value between 0 and 1.

[0061] Step D: training the initial model based on the auxiliary loss value and the first loss value to obtain a trained abnormal data detection model.

[0062] The initial model is trained based on the auxiliary loss value and the first loss value until the auxiliary loss value and the first loss value converge to obtain a trained abnormal data detection model.

[0063] Step fourteen: based on each output result, detect abnormal medical insurance data from multiple pieces of medical insurance data to be detected.

[0064] The above-mentioned medical insurance anomaly data detection method based on unsupervised learning is an unsupervised medical insurance anomaly detection method based on hybrid expert variational autoencoder. This method improves the efficiency and accuracy of medical insurance anomaly data detection, while reducing dependence on labeled data. This method combines variational autoencoders with hybrid expert strategies. Without relying on abnormal labels, it can achieve comprehensive monitoring and anomaly detection of massive medical insurance treatment data, thereby providing more accurate and efficient technical support for medical insurance management and helping to improve the quality of government services.

[0065] The above-mentioned medical insurance abnormal data detection method based on unsupervised learning, through a hybrid expert mechanism, automatically selects the appropriate network module according to the characteristics of the input data, and realizes flexible identification of different types of abnormalities; it can capture multiple patterns and abnormal categories in the data, which helps to detect subtle or different types of abnormal behaviors. This scheme gives each network module different weighted results, and the output result after weighted summation has strong robustness, can effectively identify abnormalities under different conditions, reduce misjudgment, and improve the reliability of detection results. This scheme uses the reconstruction process of the variational autoencoder, and the model can identify potential anomalies through reconstruction errors. When the input data cannot be effectively reconstructed, the model tends to identify the input as abnormal and has a high detection sensitivity.

[0066] This solution adopts an unsupervised training method, which does not require manual data annotation. It directly adaptively learns abnormal features in medical insurance data through the model, reducing labor cost investment, reducing the burden of relying on labeled data, and achieving efficient anomaly detection. The model parameters used in this solution architecture are small and the reasoning speed is fast, so it can efficiently process massive medical insurance transaction data. With fast reasoning capabilities, the model can complete data analysis and anomaly detection in a short time, which is suitable for real-time or near real-time medical insurance data monitoring scenarios, improving detection efficiency. Through modular design, this solution can be independently deployed in an isolated network environment, meeting the high confidentiality requirements of medical insurance data and ensuring safe and reliable applications in sensitive data environments.

[0067] The above-mentioned medical insurance abnormal data detection method based on unsupervised learning can be applied to medical insurance business anomaly detection work. It does not require manual data labeling and provides accuracy and detection speed far higher than traditional means. This method uses the variational autoencoder structure for feature extraction and reconstruction, and makes abnormal judgments on data with large reconstruction errors, thereby realizing automatic anomaly detection and reducing the need for manual intervention. Compared with manual verification methods, it can significantly improve detection speed and accuracy.

[0068] This solution applies a hybrid expert strategy. Multiple different network modules can capture different features in the data and establish a flexible expert library through unsupervised training. It has richer feature extraction dimensions for medical insurance data and has good scalability for efficiently processing multiple types of data.

[0069] This solution can extract multi-dimensional features from the medical insurance data to be tested, embed the features of the input data into different network modules through layer-by-layer conversion of latent vectors, and then realize the fusion and enhancement of multi-dimensional features, so as to accurately identify abnormal patterns and abnormal categories in the data.

[0070] This solution uses a small parameter model with fast inference speed, so it can complete anomaly detection of a large amount of medical insurance transaction data in a short time. It can be applied to real-time or near real-time monitoring scenarios, and can quickly process data on a large scale, further improving detection efficiency and response speed.

[0071] The embodiment of the present invention provides a medical insurance abnormal data detection device based on unsupervised learning, such as Figure 5 As shown, the device comprises: An acquisition module 50 is used to acquire multiple pieces of medical insurance data to be tested; The output module 51 is used for inputting each piece of medical insurance data to be detected into a pre-trained abnormal data detection model, and outputting the output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and a plurality of network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from the plurality of network modules; The detection module 52 is used to detect abnormal medical insurance data from multiple pieces of medical insurance data to be detected based on each output result.

[0072] The above-mentioned medical insurance abnormal data detection device based on unsupervised learning can automatically select a matching network module from multiple network modules included in the abnormal data detection model according to the medical insurance data to be detected, thereby realizing flexible recognition of different abnormal data types and improving the efficiency and accuracy of medical insurance abnormal data detection.

[0073] Furthermore, each piece of medical insurance data to be detected is in vector form; a preset weight matrix is ​​pre-stored in the selector; the input end of the selector is the input end of the abnormal data detection model, and the output end of the selector is connected to the input end of each network module; the output module is also used for: for each piece of medical insurance data to be detected, the medical insurance data to be detected is input into the selector, and the product of the medical insurance data to be detected and the preset weight matrix is ​​calculated by the selector to obtain multiple weighted results; wherein the multiple weighted results correspond to multiple network modules one by one; the multiple weighted results are sorted in order from large to small to obtain weighted sorting results; the first preset number of weighted results in the weighted sorting results are determined as multiple target weighted results; based on the multiple target weighted results, the output result corresponding to the medical insurance data to be detected is determined.

[0074] Furthermore, the output module is also used to: normalize multiple target weighted results to obtain multiple probability results; activate each target network module corresponding to each probability result, and input the medical insurance data to be tested into each target network module, and output the corresponding first results through each target network module; multiply the probability result corresponding to each target network module by the corresponding first result to obtain the first product result corresponding to each target network module; add each first product result to obtain the output result corresponding to the medical insurance data to be tested.

[0075] Furthermore, the detection module is also used to: for each piece of medical insurance data to be detected, calculate the reconstruction error corresponding to the medical insurance data to be detected based on the output result corresponding to the medical insurance data to be detected and the medical insurance data to be detected; and detect abnormal medical insurance data from multiple pieces of medical insurance data to be detected according to the reconstruction errors corresponding to each piece of medical insurance data to be detected.

[0076] Furthermore, the detection module is also used to: sort the reconstruction errors corresponding to each medical insurance data to be detected in order from large to small to obtain a sorting result; obtain the data quantity and the preset abnormal proportion value of the medical insurance data to be detected; calculate the product of the data quantity and the abnormal proportion value to obtain the product result, and determine the product result as the target serial number; determine the reconstruction error corresponding to the target serial number in the sorting result as the abnormal threshold; determine the medical insurance data to be detected whose reconstruction error is greater than the abnormal threshold as medical insurance abnormal data.

[0077] Furthermore, the device also includes an abnormal data detection model training module, which is used to: obtain sample medical insurance data; input the sample medical insurance data into the initial model to output the processing results corresponding to the sample medical insurance data through the initial model; wherein the initial model includes: an initial selector and multiple initial modules based on variational autoencoders; based on a preset loss function, calculate an auxiliary loss value, and a first loss value between the processing result and the sample medical insurance data; wherein the auxiliary loss value is used to: indicate the degree of difference in the importance of each initial module; train the initial model based on the auxiliary loss value and the first loss value to obtain a trained abnormal data detection model.

[0078] Furthermore, a sample weight matrix is ​​pre-stored in the initial selector; the auxiliary loss value is determined by: inputting the sample medical insurance data into the initial selector, calculating the product of the sample medical insurance data and the sample weight matrix through the initial selector, and obtaining multiple sample weighted results; based on the multiple sample weighted results, determining the standard deviation and the average value; Based on the preset adjustment parameters, standard deviation and mean, the auxiliary loss value is calculated.

[0079] Furthermore, the acquisition module is also used to: acquire multiple pieces of original medical insurance data; According to a preset processing method, each piece of original medical insurance data is preprocessed to obtain multiple pieces of medical insurance data to be tested; wherein each piece of medical insurance data to be tested is in vector form; the preset processing method includes at least one of the following: data screening processing, data extraction processing, missing value processing, normalization processing and text encoding processing.

[0080] The medical insurance abnormal data detection device based on unsupervised learning provided in the embodiment of the present invention has the same implementation principle and technical effects as the aforementioned medical insurance abnormal data detection method embodiment based on unsupervised learning. For the sake of brief description, for the parts not mentioned in the embodiment of the medical insurance abnormal data detection device based on unsupervised learning, reference may be made to the corresponding contents in the aforementioned medical insurance abnormal data detection method embodiment based on unsupervised learning.

[0081] The embodiment of the present invention further provides an electronic device, see Figure 6 As shown, the electronic device includes a processor 130 and a memory 131, wherein the memory 131 stores machine executable instructions that can be executed by the processor 130, and the processor 130 executes the machine executable instructions to implement the above-mentioned medical insurance abnormal data detection method based on unsupervised learning.

[0082] Further, Figure 6 The electronic device shown further includes a bus 132 and a communication interface 133 , and the processor 130 , the communication interface 133 and the memory 131 are connected via the bus 132 .

[0083] The memory 131 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 133 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used. The bus 132 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0084] The processor 130 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 130. The above processor 130 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly embodied as a hardware decoding processor for execution, or a combination of hardware and software modules in the decoding processor for execution. The software module may be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 131, and the processor 130 reads the information in the memory 131 and completes the steps of the method of the above embodiment in combination with its hardware.

[0085] An embodiment of the present invention also provides a machine-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the above-mentioned medical insurance abnormal data detection method based on unsupervised learning. The specific implementation can be found in the method embodiment, which will not be repeated here.

[0086] The computer program product of the medical insurance abnormal data detection method, device and electronic device based on unsupervised learning provided in the embodiments of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can be found in the method embodiment, which will not be repeated here.

[0087] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting abnormal medical insurance data based on unsupervised learning, characterized in that: The method comprises: Obtain multiple pieces of medical insurance data to be tested; For each piece of the medical insurance data to be detected, the medical insurance data to be detected is input into a pre-trained abnormal data detection model, and the output result corresponding to the medical insurance data to be detected is output through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and a plurality of network modules based on variational autoencoders; the module parameters of each network module are different, and each network module corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from the plurality of network modules; Based on each of the output results, abnormal medical insurance data is detected from the plurality of medical insurance data to be detected.

2. The method according to claim 1, characterized in that: Each of the medical insurance data to be detected is in vector form; a preset weight matrix is ​​pre-stored in the selector; the input end of the selector is the input end of the abnormal data detection model, and the output end of the selector is connected to the input end of each of the network modules; The step of inputting each piece of the medical insurance data to be detected into a pre-trained abnormal data detection model, and outputting an output result corresponding to the medical insurance data to be detected through the abnormal data detection model comprises: For each piece of the medical insurance data to be detected, the medical insurance data to be detected is input into the selector, and the product of the medical insurance data to be detected and the preset weight matrix is ​​calculated by the selector to obtain a plurality of weighted results; wherein the plurality of weighted results correspond to the plurality of network modules one by one; Sorting the plurality of weighted results in descending order to obtain a weighted sorting result; Determining a preset number of weighted results among the weighted sorting results as a plurality of target weighted results; Based on the weighted results of the multiple targets, the output result corresponding to the medical insurance data to be tested is determined.

3. The method according to claim 2, characterized in that The step of determining the output result corresponding to the medical insurance data to be detected based on the multiple target weighted results includes: Normalizing the multiple target weighted results to obtain multiple probability results; Activate each target network module corresponding to each of the probability results, input the medical insurance data to be detected into each of the target network modules, and output the first results corresponding to each of the target network modules; Multiplying the probability result corresponding to each of the target network modules by the corresponding first result to obtain a first product result corresponding to each of the target network modules; Add each of the first multiplication results to obtain the output result corresponding to the medical insurance data to be tested.

4. The method according to claim 1, characterized in that The step of detecting abnormal medical insurance data from the plurality of medical insurance data to be detected based on each of the output results comprises: For each piece of the medical insurance data to be detected, based on the output result corresponding to the medical insurance data to be detected and the medical insurance data to be detected, calculate the reconstruction error corresponding to the medical insurance data to be detected; According to the reconstruction error corresponding to each piece of the medical insurance data to be detected, abnormal medical insurance data is detected from the multiple pieces of the medical insurance data to be detected.

5. The method according to claim 4, characterized in that The step of detecting abnormal medical insurance data from a plurality of medical insurance data to be detected according to the reconstruction error corresponding to each medical insurance data to be detected comprises: Sorting the reconstruction errors corresponding to each piece of the medical insurance data to be detected in descending order to obtain a sorting result; Obtaining the data quantity and preset abnormal ratio value of the medical insurance data to be detected; Calculate the product of the data quantity and the abnormal proportion value to obtain a product result, and determine the product result as the target sequence number; Determine the reconstruction error corresponding to the target sequence number in the sorting result as an abnormal threshold; The medical insurance data to be detected whose reconstruction error is greater than the abnormal threshold is determined as abnormal medical insurance data.

6. The method according to claim 1, characterized in that The abnormal data detection model is trained in the following way: Obtain sample medical insurance data; Input the sample medical insurance data into the initial model to output the processing result corresponding to the sample medical insurance data through the initial model; wherein the initial model includes: an initial selector and a plurality of initial modules based on variational autoencoders; Based on a preset loss function, an auxiliary loss value and a first loss value between the processing result and the sample medical insurance data are calculated; wherein the auxiliary loss value is used to: indicate the degree of difference in importance of each of the initial modules; The initial model is trained based on the auxiliary loss value and the first loss value to obtain the trained abnormal data detection model.

7. The method according to claim 6, characterized in that The sample weight matrix is ​​pre-stored in the initial selector; the auxiliary loss value is determined by the following method: Inputting the sample medical insurance data into the initial selector, and calculating the product of the sample medical insurance data and the sample weight matrix through the initial selector to obtain multiple sample weighted results; Determining a standard deviation and an average based on a plurality of said sample weighted results; The auxiliary loss value is calculated based on the preset adjustment parameter, the standard deviation and the average value.

8. The method according to claim 1, characterized in that: The step of obtaining multiple pieces of medical insurance data to be tested includes: Obtain multiple pieces of original medical insurance data; According to a preset processing method, each piece of the original medical insurance data is preprocessed to obtain multiple pieces of medical insurance data to be tested; wherein each piece of the medical insurance data to be tested is in vector form; the preset processing method includes at least one of the following: data screening processing, data extraction processing, missing value processing, normalization processing and text encoding processing.

9. A medical insurance abnormal data detection device based on unsupervised learning, characterized in that: The device comprises: An acquisition module is used to obtain multiple pieces of medical insurance data to be tested; An output module is used for inputting each piece of the medical insurance data to be detected into a pre-trained abnormal data detection model, and outputting an output result corresponding to the medical insurance data to be detected through the abnormal data detection model; wherein the abnormal data detection model includes: a selector and a plurality of network modules based on variational autoencoders; each of the network modules has different module parameters, and each of the network modules corresponds to a different abnormal data type; the selector is used to select a network module that matches the medical insurance data to be detected from the plurality of network modules; The detection module is used to detect abnormal medical insurance data from the multiple medical insurance data to be detected based on each of the output results.

10. An electronic device, characterized in that: It includes a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the medical insurance abnormal data detection method based on unsupervised learning as described in any one of claims 1-8.

Images