Medical insurance abnormal data detection method and training method of medical insurance abnormal detection model

Abstract

The invention provides a medical insurance abnormal data detection method and a medical insurance abnormal detection model training method, and the method comprises the steps: screening medical insurance total data according to a screening condition, and obtaining a first screening result containing a plurality of pieces of medical insurance data; for each piece of medical insurance data, extracting a target field and a field value corresponding to the target field from the medical insurance data according to a preset target field, and further synthesizing a target prompt word; and converting the target cue word into a vector form to obtain to-be-detected medical insurance data, inputting the to-be-detected medical insurance data into a pre-trained medical insurance anomaly detection model, and outputting a detection result corresponding to the to-be-detected medical insurance data. According to the mode, the medical insurance total data can be screened, the target field and the corresponding field value can be extracted, the target cue word is synthesized, the target cue word is converted into the vector form suitable for analysis of the medical insurance anomaly detection model, and the obtained to-be-detected medical insurance data is detected by adopting the medical insurance anomaly detection model, so that the manual workload can be reduced, and the detection efficiency is improved. And the detection accuracy and the detection efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for detecting abnormal medical insurance data and a method for training a medical insurance abnormality detection model. Background Art

[0002] With the expansion of medical insurance service coverage and the increasing complexity of medical services, the medical insurance system is facing more and more management and review challenges, especially in the detection and prevention of medical insurance fraud. Traditional methods of detecting fraud often rely on public reports, the formulation of detection rules and manual review. In addition, the medical insurance business process is complex and changeable, and manual operations cannot quickly respond to policy adjustments and the differentiated needs of medical insurance treatment scenarios. In summary, the existing methods that rely on manual detection have the problems of large manual workload, low detection accuracy and efficiency. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting abnormal medical insurance data and a method for training a medical insurance abnormality detection model to reduce manual workload and improve detection accuracy and efficiency.

[0004] The present invention provides a method for detecting abnormal medical insurance data, which includes: obtaining preset full medical insurance data; screening the full medical insurance data according to preset screening conditions to obtain a first screening result; wherein the first screening result includes multiple medical insurance data; for each medical insurance data, according to a preset target field, extracting a target field and a field value corresponding to the target field from the medical insurance data; synthesizing a target prompt word based on the target field and the field value corresponding to the target field; converting the target prompt word into a vector form to obtain medical insurance data to be detected; inputting the medical insurance data to be detected into a pre-trained medical insurance anomaly detection model, and outputting a detection result corresponding to the medical insurance data to be detected through the medical insurance anomaly detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

[0005] Furthermore, the medical insurance anomaly detection model includes: a general sub-model and multiple task sub-models; each task sub-model is respectively communicated with the general sub-model; each task sub-model corresponds to a different anomaly detection task; the medical insurance data to be detected is input into a pre-trained medical insurance anomaly detection model, and the step of outputting the detection result corresponding to the medical insurance data to be detected through the medical insurance anomaly detection model includes: inputting the medical insurance data to be detected into the general sub-model, and outputting the general representation vector corresponding to the medical insurance data to be detected through the general sub-model; receiving a model selection instruction for a target anomaly detection task, and selecting a target task sub-model corresponding to the target anomaly detection task from multiple task sub-models according to the model selection instruction; sending the general representation vector to the target task sub-model to output the detection result corresponding to the medical insurance data to be detected through the target task sub-model.

[0006] Furthermore, the method also includes: sending the test results to a preset terminal device through a preset API interface, so that the test results are displayed through the terminal device.

[0007] Furthermore, the method also includes: if the test result indicates that the medical insurance data to be tested is abnormal data, saving the medical insurance data to be tested and the test result corresponding to the medical insurance data to be tested.

[0008] Furthermore, there are multiple target fields; based on the target fields and the field values ​​corresponding to the target fields, the step of synthesizing the target prompt words includes: if the field values ​​corresponding to each target field are not missing, synthesizing the target prompt words based on each target field and the field values ​​corresponding to each target field; if the field value corresponding to the designated field is missing among multiple target fields, processing the field value corresponding to the designated field according to a preset missing value processing method to obtain a processed field value; synthesizing the target prompt words based on the designated field, the processed field value, and each field except the designated field and the field values ​​corresponding to each field among the multiple target fields.

[0009] The present invention provides a training method for a medical insurance anomaly detection model, the method comprising: obtaining sample medical insurance data; wherein the sample medical insurance data is pre-labeled with standard results; inputting the sample medical insurance data into an initial model to output initial results corresponding to the sample medical insurance data through the initial model; based on a preset loss function, calculating the loss value of the standard result and the initial result, updating the weight parameters of the initial model based on the loss value, repeating the step of obtaining the sample medical insurance data until the loss value converges, and obtaining the medical insurance anomaly detection model.

[0010] The present invention provides a medical insurance abnormal data detection device, which includes: a first acquisition module, used to acquire preset full medical insurance data; a screening module, used to screen the full medical insurance data according to preset screening conditions to obtain a first screening result; wherein the first screening result includes multiple medical insurance data; an extraction module, used to extract a target field and a field value corresponding to the target field from each medical insurance data according to a preset target field; a synthesis module, used to synthesize a target prompt word based on the target field and the field value corresponding to the target field; a conversion module, used to convert the target prompt word into a vector form to obtain medical insurance data to be detected; a first output module, used to input the medical insurance data to be detected into a pre-trained medical insurance abnormality detection model, and output a detection result corresponding to the medical insurance data to be detected through the medical insurance abnormality detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

[0011] The present invention provides a training device for a medical insurance anomaly detection model, which includes: a second acquisition module, used to acquire sample medical insurance data; wherein the sample medical insurance data is pre-marked with standard results; a second output module, used to input the sample medical insurance data into an initial model to output initial results corresponding to the sample medical insurance data through the initial model; a third acquisition module, used to calculate the loss value of the standard result and the initial result based on a preset loss function, update the weight parameters of the initial model based on the loss value, and repeatedly execute the step of acquiring the sample medical insurance data until the loss value converges to obtain the medical insurance anomaly detection model.

[0012] 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 abnormality data detection methods, or the above-mentioned medical insurance abnormality detection model training methods.

[0013] The present invention 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 any of the above-mentioned medical insurance anomaly data detection methods, or the above-mentioned medical insurance anomaly detection model training methods.

[0014] The medical insurance abnormal data detection method and the training method of the medical insurance abnormal detection model provided by the present invention obtain preset full medical insurance data; screen the full medical insurance data according to preset screening conditions to obtain a first screening result; wherein the first screening result includes multiple medical insurance data; for each medical insurance data, according to the preset target field, extract the target field and the field value corresponding to the target field from the medical insurance data; synthesize the target prompt word based on the target field and the field value corresponding to the target field; convert the target prompt word into a vector form to obtain the medical insurance data to be detected; input the medical insurance data to be detected into a pre-trained medical insurance abnormal detection model, and output the detection result corresponding to the medical insurance data to be detected through the medical insurance abnormal detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data. This method can screen the full amount of medical insurance data, extract the target field and the corresponding field value, synthesize the target prompt word, and convert the target prompt word into a vector form suitable for the analysis of the medical insurance abnormal detection model, and use the medical insurance abnormal detection model to detect the obtained medical insurance data to be detected, which can reduce the manual workload and improve the detection accuracy and detection efficiency. 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 insurance fraud detection process provided by an embodiment of the present invention;

[0017] Figure 2 A flowchart of a method for detecting abnormal medical insurance data provided by an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of the overall architecture of a medical insurance abnormal data detection system provided by an embodiment of the present invention;

[0019] Figure 4 A schematic diagram of the structure of a medical insurance anomaly detection model provided by an embodiment of the present invention;

[0020] Figure 5 A schematic diagram of the architecture of a general sub-model provided by an embodiment of the present invention;

[0021] Figure 6 A schematic diagram of the overall training process of a medical insurance anomaly detection model provided by an embodiment of the present invention;

[0022] Figure 7 A schematic diagram of the structure of a medical insurance abnormal data detection device provided by an embodiment of the present invention;

[0023] Figure 8 A schematic diagram of the structure of a training device for a medical insurance anomaly detection model provided by an embodiment of the present invention;

[0024] Fig. 9 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] 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.

[0026] At present, traditional methods of detecting insurance fraud often rely on public reports, the formulation of detection rules and manual review, which have problems such as large workload, low accuracy and low efficiency. In addition, the medical insurance business process is complex and changeable, and manual operations are difficult to quickly respond to policy adjustments and the differentiated needs of medical insurance treatment scenarios, making it difficult to detect insurance fraud in a timely manner. At the same time, with the continuous renovation and increased concealment of insurance fraud methods, traditional detection methods can no longer meet the current needs of medical insurance management.

[0027] The diversification and decentralization of medical insurance services also increase the complexity of management. In the current medical insurance management system, although various medical insurance departments have accumulated a large amount of medical treatment data, these data are often stored in different information systems and lack a unified data management and analysis platform. This leads to serious data silos, poor information sharing between departments, and difficulty in forming effective data linkage and comprehensive analysis. In addition, due to the complexity and diversity of medical insurance treatment data, traditional data analysis methods often have difficulty in accurately capturing patients' abnormal treatment characteristics, making it difficult to accurately discover medical insurance fraud data.

[0028] like Figure 1 The schematic diagram of a traditional insurance fraud detection process is shown in FIG. The general situation of the traditional insurance fraud detection process is as follows:

[0029] 1. Based on mass reporting:

[0030] Enthusiastic people mainly submit reporting information through hotline to report suspected medical insurance fraud. These reporting information will be directly sent to the manual review stage. Based on the content of the report and in combination with relevant policies and regulations, the reviewer will make a comprehensive judgment on the patient's medical records, medical expenses and other information to determine whether there is insurance fraud. After obtaining the review structure, the relevant personnel can be interviewed for confirmation.

[0031] (II) Active screening paths by regulatory agencies:

[0032] (1) The regulatory agency first conducts a comprehensive screening of the medical insurance data throughout the year, and uses SQL (Structured Query Language) and other technologies to screen out those who have used a large amount of medical insurance funds throughout the year as suspected fraud targets based on relevant constraints;

[0033] (2) For suspicious objects after data screening, the regulatory agency will hand them over to the law enforcement department for manual review. The manual reviewers will conduct a detailed review of these suspicious data, including comparing the patient's medical records, medical expenses and other information, and making a comprehensive judgment based on other relevant information;

[0034] (3) After receiving the manual review results, if it is still considered that there is a high possibility of insurance fraud, the relevant personnel will be interviewed to confirm and specific handling measures will be taken. If insurance fraud is confirmed, the relevant departments will handle it in accordance with the law to ensure the rational use of medical insurance funds and the legitimate rights and interests of patients.

[0035] As the scale of medical insurance funds expands, the manual review process for insurance fraud becomes increasingly complex. The approval process has the following main shortcomings:

[0036] 1. Low approval efficiency and inability to implement efficient collaboration;

[0037] 2. Approval is highly subjective, and the evaluation criteria may vary, resulting in missed or misjudgment;

[0038] 3. High human resource operating costs;

[0039] 4. Lack of efficient data processing and analysis capabilities limits the efficiency and accuracy of the audit;

[0040] 5. Weak ability to respond to business changes, making it difficult to capture insurance fraud behaviors that keep pace with the times and difficult to discover new insurance fraud behaviors;

[0041] Based on this, an embodiment of the present invention provides a method for detecting abnormal medical insurance data and a method for training a medical insurance abnormality detection model. This technology can be applied to applications that require detection of abnormal medical insurance data.

[0042] To facilitate understanding of this embodiment, firstly, a method for detecting abnormal medical insurance data disclosed in an embodiment of the present invention is introduced. Figure 2 As shown, the method comprises the following steps:

[0043] Step S202, obtaining preset full medical insurance data;

[0044] The above-mentioned full medical insurance data may include: hospital data, doctor data, patient data, unit data, prescription data, drug data, violation data, etc.; in actual implementation, the full medical insurance data of multiple patients can be obtained from a variety of different data sources, for example, the data source can be a medical institution, social security agency, etc.

[0045] Step S204, screening the entire amount of medical insurance data according to a preset screening condition to obtain a first screening result; wherein the first screening result includes a plurality of medical insurance data;

[0046] The above screening conditions can be set according to actual needs. For example, taking the screening condition of "over 20,000" as an example, the full amount of medical insurance data is screened according to the screening condition, and the multiple medical insurance data corresponding to the multiple patients obtained all have medical insurance reimbursement amounts of 20,000 or more. Among the multiple screened medical insurance data, there may be data involving insurance fraud, or there may be no data involving insurance fraud, which needs to be confirmed through subsequent testing methods.

[0047] Step S206, for each piece of medical insurance data, according to the preset target field, extract the target field and the field value corresponding to the target field from the medical insurance data;

[0048] The above-mentioned target fields can be set according to actual needs. These target fields are usually fields related to detecting whether the medical insurance data is abnormal. There are usually multiple target fields, for example, they can include: patient ID (Identity Document), diagnosis ID, settlement ID, gender, age, consultation time, consultation hospital, diagnosis information, total medical expenses, amount within the scope, drug diagnosis and treatment services, etc.; in actual implementation, the relevant target fields can be extracted for the multiple medical insurance data filtered out above, and the field value corresponding to each target field can be obtained.

[0049] Step S208, synthesizing a target prompt word based on the target field and the field value corresponding to the target field;

[0050] In actual implementation, for each piece of medical insurance data, after obtaining the above-mentioned target field and the field value corresponding to the target field, the target prompt word can be synthesized based on the obtained information. For example, a preset prompt word template can be obtained, and the field values ​​corresponding to each target field obtained can be inserted into the prompt word template to form a complete target prompt word. For example, the prompt word template can include age, gender, total number of visits, frequency of visits, etc.

[0051] Step S210, converting the target prompt word into a vector form to obtain the medical insurance data to be tested;

[0052] In actual implementation, after obtaining the above-mentioned target prompt words, the target prompt words can be converted into a vector form that can be used by the medical insurance anomaly detection model, that is, a standard data vector, to obtain the medical insurance data to be detected; in actual applications, it can be determined first whether the target prompt words belong to text data or numerical data, so as to adopt different forms of processing methods. For example, if the target prompt words belong to text data, a large language model is usually used to extract effective information and encode it into a word vector form; if the target prompt words belong to numerical data, a feedforward neural network can be used to encode them into a numerical vector form. In other words, the obtained medical insurance data to be detected enters the subsequent operation process in the form of vectors.

[0053] Step S212, input the medical insurance data to be detected into a pre-trained medical insurance anomaly detection model, and output the detection result corresponding to the medical insurance data to be detected through the medical insurance anomaly detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

[0054] In actual implementation, after obtaining the above-mentioned medical insurance data to be tested, the medical insurance data to be tested can be input into a pre-trained medical insurance anomaly detection model to detect the medical insurance data to be tested through the medical insurance anomaly detection model to confirm the test result corresponding to the medical insurance data to be tested; for example, the test result may be: the medical insurance data to be tested is abnormal data; or, the test result may be: the medical insurance data to be tested is not abnormal data.

[0055] The above-mentioned medical insurance abnormal data detection method obtains preset full medical insurance data; screens the full medical insurance data according to preset screening conditions to obtain a first screening result; wherein the first screening result includes multiple medical insurance data; for each medical insurance data, according to the preset target field, extracts the target field and the field value corresponding to the target field from the medical insurance data; synthesizes the target prompt word based on the target field and the field value corresponding to the target field; converts the target prompt word into a vector form to obtain the medical insurance data to be detected; inputs the medical insurance data to be detected into a pre-trained medical insurance abnormality detection model, and outputs the detection result corresponding to the medical insurance data to be detected through the medical insurance abnormality detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data. This method can screen the full medical insurance data, extract the target field and the corresponding field value, synthesize the target prompt word, and convert the target prompt word into a vector form suitable for the analysis of the medical insurance abnormality detection model, and use the medical insurance abnormality detection model to detect the obtained medical insurance data to be detected, which can reduce the manual workload and improve the detection accuracy and detection efficiency.

[0056] The embodiment of the present invention also provides another medical insurance abnormal data detection method, which is implemented on the basis of the above embodiment method. In this method, the medical insurance abnormality detection model includes: a general sub-model and multiple task sub-models; each task sub-model is respectively communicated with the general sub-model; each task sub-model corresponds to a different abnormality detection task, and each task sub-model is usually a neural network obtained by simple training; in actual implementation, each task sub-model can have the same structure or a different structure. When each task sub-model has the same structure, the model parameters of each task sub-model are usually different, so that each task sub-model corresponds to a different abnormality detection task; for example, the abnormality detection task can be an abnormality detection task corresponding to "full 20,000", an abnormality detection task corresponding to "full 100,000", etc.; the method includes the following steps:

[0057] Step 1: Obtain the preset full amount of medical insurance data;

[0058] Step 2: screening the entire amount of medical insurance data according to a preset screening condition to obtain a first screening result; wherein the first screening result includes a plurality of medical insurance data;

[0059] Step 3: for each piece of medical insurance data, according to the preset target field, extract the target field and the field value corresponding to the target field from the medical insurance data;

[0060] Step 4: if the field values ​​corresponding to each target field are not missing, synthesize the target prompt words based on each target field and the field values ​​corresponding to each target field;

[0061] In actual implementation, for each piece of medical insurance data, if the field values ​​corresponding to each target field extracted from the medical insurance data are not missing, it can be considered that each extracted target field and the corresponding field value are complete. In this case, the target prompt word can be directly synthesized based on each target field and the corresponding field value.

[0062] Step 5: If the field value corresponding to the specified field is missing in multiple target fields, the field value corresponding to the specified field is processed according to the preset missing value processing method to obtain the processed field value;

[0063] Step six, synthesizing a target prompt word based on the designated field, the processed field value, and each field other than the designated field and the field value corresponding to each field in the plurality of target fields.

[0064] The number of the above-mentioned designated fields may be one or more; the above-mentioned missing value processing method may include deleting missing values, filling missing values, etc.; in actual implementation, for each medical insurance data, one or more designated fields may be missing in the multiple target fields extracted from the medical insurance data. Therefore, it is usually necessary to detect each target field to identify missing values, and then delete the designated field or fill the field value of the designated field according to the attribute of the field value corresponding to the designated field; for example, if it is difficult to determine the field value that needs to be filled in the designated field based on the existing data, the designated field can be deleted; if the field value that needs to be filled in the designated field can be determined based on the existing data, the designated field can be filled; for example, the average value of the field value corresponding to the designated field in other medical insurance data can be filled in the designated field of the medical insurance data, etc.; after the field value of the designated field is processed, the target prompt word can be synthesized based on each updated target field and the corresponding field value. In actual application, if there are many missing field values ​​in the multiple target fields extracted from the medical insurance data, it is generally believed that the medical insurance data has little reference significance, and the medical insurance data can be directly deleted.

[0065] Step 7: Convert the target prompt word into a vector form to obtain the medical insurance data to be tested;

[0066] Step 8, inputting the medical insurance data to be tested into the general sub-model, and outputting the general representation vector corresponding to the medical insurance data to be tested through the general sub-model;

[0067] The above-mentioned universal characterization vector can be understood as a multi-dimensional vector used to characterize the medical insurance data to be tested. The above-mentioned universal sub-model can be a Transformer model, etc.; in actual implementation, the medical insurance data to be tested can be input into the universal sub-model, and the medical insurance data to be tested can be processed by the universal sub-model to output the corresponding universal characterization vector, which contains all the characterization information of the medical insurance data to be tested.

[0068] Step nine, receiving a model selection instruction for a target anomaly detection task, and selecting a target task sub-model corresponding to the target anomaly detection task from a plurality of task sub-models according to the model selection instruction;

[0069] In actual implementation, the user can issue corresponding model selection instructions according to the target anomaly detection task that actually needs to be performed. Since each task sub-model corresponds to a different anomaly detection task, after receiving the model selection instruction issued by the user, the target task sub-model corresponding to the target anomaly detection task can be selected from multiple task sub-models.

[0070] Step ten, sending the universal representation vector to the target task sub-model to output the test results corresponding to the medical insurance data to be tested through the target task sub-model.

[0071] The above-mentioned universal characterization vector is sent to the selected target task sub-model, and the universal characterization vector is processed by the target task sub-model to output the corresponding detection result.

[0072] Step eleven: send the test results to a preset terminal device through a preset API interface, so that the test results are displayed through the terminal device.

[0073] In actual implementation, the original medical insurance supervision system can call the medical insurance anomaly detection model through a preset API interface to send the above-mentioned detection results to a preset terminal device, and display the detection results through the page of the terminal device to assist the reviewer in completing the final anomaly judgment.

[0074] Step twelve: if the test result indicates that the medical insurance data to be tested is abnormal data, save the medical insurance data to be tested and the test result corresponding to the medical insurance data to be tested.

[0075] For each piece of medical insurance data to be tested, if the test result indicates that the medical insurance data to be tested is abnormal data, the medical insurance data to be tested and its corresponding test result can be written into the database. In addition, the original medical insurance data corresponding to the medical insurance data to be tested can also be saved in the database to provide retention of abnormal data and abnormal detection results.

[0076] For easier understanding, see Figure 3 The overall architecture diagram of a medical insurance abnormal data detection system shown in the figure includes: a data collection layer, a data processing layer, a general model layer and a data interaction layer. The data collection layer is the input port of the original data of the project, and provides the original data (corresponding to the above-mentioned full amount of medical insurance data) for the model data analysis. The data processing layer reads the original data from the data collection layer, and can use SQL to filter the obtained original data to obtain the first screening result of multiple medical insurance data containing multiple patients. On the basis of each medical insurance data obtained, multiple types of other data can be attached at the same time for splicing, including but not limited to personnel information (such as age, etc.), organizational information (such as medical insurance type, treatment department, etc.), prescription information, diagnosis information, etc., to enrich the data coverage and provide more hidden information. The data processing layer can perform text data standardization, missing value processing, data cleaning and other processing on each medical insurance data obtained by screening, and finally obtain the standard data vector that can be used by the general anomaly detection large model (corresponding to the above-mentioned medical insurance anomaly detection model), that is, obtain the medical insurance data to be detected. A general anomaly detection model is set in the general model layer. The medical insurance data to be detected can be input into the general anomaly detection model. A general representation vector is output through the general sub-model in the general anomaly detection model. In the downstream task, the corresponding target task sub-model can be selected according to different needs, and the general representation vector is input into the target task sub-model to output the detection result corresponding to the medical insurance data to be detected. The general anomaly detection model usually does not need to be adjusted frequently after deployment. For example, it can be fine-tuned and updated as needed every year. The data interaction layer provides API interface and database archiving operations. The original medical insurance supervision system can call the general anomaly detection model through the API interface, and display the detection results output by the general anomaly detection model on the front-end page to assist the auditors in completing the final anomaly judgment. The database archiving module can write the detection results output by the general anomaly detection model and the original data of the abnormal items into the database to provide retention of abnormal information and results.

[0077] As another alternative, different methods can be used to extract medical treatment characteristics based on medical insurance data. In addition to the vertical large model used in this solution (corresponding to the above-mentioned medical insurance anomaly detection model), graph neural networks can also be used. By treating patients as different nodes, connecting them by using the correlation between payment fees, hospitals visited, and drug details, and then using known insurance fraud cases as label data, using graph network models for training and learning, we can capture potential relationships in the data and discover new insurance fraud patterns; this solution is an alternative to using vertical large models to extract medical insurance treatment characteristics.

[0078] The above-mentioned medical insurance abnormal data detection method provides a vertical large model for discovering medical insurance fraud data based on medical insurance treatment characteristics, that is, the above-mentioned medical insurance anomaly detection model, which conducts in-depth mining and analysis of patient treatment data to achieve comprehensive monitoring and anomaly detection of medical insurance data, thereby effectively identifying potential medical insurance fraud behaviors, improving manual review efficiency, improving the intelligence level of medical insurance fraud detection, and improving the accuracy and efficiency of medical insurance fraud detection, providing more accurate and efficient technical support for medical insurance management, and improving the quality of government services.

[0079] This method embeds the characteristics of medical insurance fraud data into a general large language model through fine-tuning training of the general large language model, solving the problem that the general large language model cannot efficiently cope with specific fields, and realizing a vertical large model for discovering medical insurance fraud data using medical insurance treatment characteristics.

[0080] This solution uses a large model approach to centrally screen medical insurance transaction data, classifies the medical insurance data to be tested through the medical insurance anomaly detection model, obtains data anomaly conclusions, and realizes automated detection without the need for large amounts of manpower investment.

[0081] This solution uses the powerful fitting ability of the medical insurance anomaly detection model to extract the multi-dimensional data indicator features in the medical insurance treatment data. The medical insurance anomaly detection model transforms the input features layer by layer, maps the original feature space to different hidden layers, and then realizes the fusion and enhancement of features. Through the fusion of multi-dimensional data features, it can realize the detection and discovery of anomalies in insurance fraud data.

[0082] This solution fully utilizes the large model's ability to transform and map data of multiple dimensions as a basis, and fine-tunes and trains a general data base, which can achieve richer dimensional extraction for medical insurance data and has good scalability.

[0083] Through one-stop deployment, the solution can run independently in a restricted network environment and meet the high confidentiality requirements of medical insurance data.

[0084] This solution provides a vertical big model (corresponding to the above-mentioned medical insurance anomaly detection model) for discovering medical insurance fraud data based on medical insurance treatment characteristics. This model can be applied to medical insurance business anomaly detection, providing accuracy and detection speed far higher than traditional means. It can support natural language input data, and after relevant data processing, the data features are extracted through the vertical big model, and anomaly detection is performed. Compared with traditional algorithms or manual verification methods, it can significantly improve detection speed and accuracy.

[0085] Through the data collection layer and the data processing layer, the vertical big model can collect the original full amount of medical insurance data from multiple data sources and process it into standard data vectors suitable for vertical big model analysis.

[0086] The vertical large model uses the general model layer to extract features of the medical insurance data to be tested. The model is highly versatile in the medical insurance business field and provides strong scalability.

[0087] The vertical model integrates three granular features of user level, order level and detail level through multi-granular feature fusion, which can more comprehensively describe the behavior patterns of medical insurance users, thereby improving the accuracy of anomaly detection. Among them, the user level is user-related information (such as name, age, etc.); the order level includes: transaction amount, hospital department corresponding to the medical insurance transaction, etc.; the detail level can be understood as the order details in each transaction.

[0088] The solution also designs an easy-to-integrate API interface to facilitate interaction and data exchange with other systems, improving the scalability and maintainability of the system.

[0089] This solution can also save the general representation vector to the database, so that large vertical models can be quickly migrated and fine-tuned on different downstream tasks. This improves the efficiency of "multiple tasks, one model", and the model can share knowledge between different tasks, achieving higher robustness and generalization.

[0090] The embodiment of the present invention also provides a method for training a medical insurance anomaly detection model, the method comprising the following steps:

[0091] Step A, obtaining sample medical insurance data; wherein the sample medical insurance data is pre-marked with standard results;

[0092] 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.

[0093] Here we take the medical insurance "over 20,000" behavior as an example, and the data processing methods for other insurance fraud behaviors are similar. The data is analyzed through correlation analysis and other methods, and finally all the medical information of the patients arranged in chronological order within a time window is selected (each medical information includes: visit time, hospital, diagnosis information, total medical expenses, amount within the scope, and drug diagnosis and treatment services, a total of six fields). As the input of the "over 20,000" medical insurance anomaly detection model. At the same time, in the transaction details data for the whole year of 2022, according to the positive and negative sample ratio of one to five, the positive and negative sample training sets: a total of 119,017, and the verification set: a total of 13,527, are sampled for model training and verification of the model during training.

[0094] The acquired original sample medical insurance data can be processed in the manner described in the foregoing embodiments, such as data screening, data extraction, missing value processing, prompt word synthesis, etc.; the synthesized prompt words are converted into vector form to obtain the above-mentioned sample medical insurance data, wherein the prompt word format incorporates the prompt word engineering method, which can make full use of the text analysis ability obtained by the initial model in pre-training, thereby improving the final performance of the "full 20,000" medical insurance anomaly detection task.

[0095] Step B, inputting the sample medical insurance data into the initial model, so as to output the initial results corresponding to the sample medical insurance data through the initial model;

[0096] Step C, based on the preset loss function, calculate the loss value of the standard result and the initial result, update the weight parameters of the initial model based on the loss value, repeat the steps of obtaining sample medical insurance data until the loss value converges, and obtain the medical insurance anomaly detection model.

[0097] In the medical insurance law enforcement and supervision system, the traditional screening of insurance fraud behavior will first use the database query method to obtain all suspicious insured persons who meet the abnormal conditions, and then manually review the transactions of suspicious persons, and finally refuse to pay for the transaction details. Using artificial intelligence algorithms, insurance fraud screening can be modeled as a classification task. In this classification task, non-suspicious persons are negative samples, while suspicious persons are positive samples. The goal of the model is to construct various data such as the insured person's personal information and transaction information into effective features that the model can process, and finally realize anomaly detection through the medical insurance anomaly detection model, so as to achieve the goal of artificial intelligence empowering anomaly detection. The medical insurance anomaly detection model aims to quickly screen out illegal transactions and output suspicious scores through the relevant data indicators of suspicious insured persons, thereby reducing the pressure of manual review and improving the efficiency of illegal screening.

[0098] like Figure 4 The structure diagram of a medical insurance anomaly detection model is shown in Figure 1. The large model base uses the most competitive Qwen2-7B model of the same model scale as the large model base of the "full 20,000" anomaly detection model. Qwen2 is an upgraded large language model series launched by the Qianwen large model team, covering multiple sizes from 0.5B to 72B parameters, supporting 27 languages ​​and significantly improving code and mathematical processing capabilities. The Qwen2 series of models have been open sourced on Hugging Face and ModelScope, using the Apache 2.0 license, promoting new developments in AI (Artificial Intelligence) technology in multilingual, long text processing and security. The use of this large model base can combine low resource overhead, high training and reasoning efficiency, and excellent model performance, and is the best choice among the model base selection solutions currently explored.

[0099] like Figure 5 The schematic diagram of the architecture of a general sub-model shown in FIG. 1 shows the model process from inputting the target prompt word to outputting the general representation vector. The functions of each part in the figure can refer to the relevant technology. The following is a brief description of the main parts:

[0100] 1. Tokenization operation: This operation quantizes the input target prompt words and decomposes the tables, text and other data in the medical insurance data into tokens that the model can understand. The tokens are also in the form of vectors. Among them, Text refers to the target prompt words synthesized based on the original full medical insurance data. The full medical insurance data can contain structured tables, medical advice and other text data. Then, the target prompt words are converted into a token sequence through Tokenizer, that is, the target prompt words are converted into vectors.

[0101] 2. The backbone of the model

[0102] a) Embedding operation. The model maps the discrete word sequence generated in the previous step into a continuous embedding vector, which is the hidden layer (hidden_states). The hidden layer contains the step-by-step representation of the input text in the multi-layer model, captures the contextual semantics of the word, and retains the similarity relationship between words for further model layer processing;

[0103] b) Core layer. This layer is composed of multiple decoders stacked together. In the decoder, the input hidden state is first normalized by RMS Norm, and then the correlation between hidden states is captured by the self-attention mechanism and residual connection. Finally, a multilayer perceptron (MLP) is used for nonlinear transformation, and the hidden state obtained by the self-attention layer is added to generate an updated hidden state for the next layer.

[0104] c) Normalization layer (RMSNorm). This is a layer normalization method. The calculation formula is as follows, where x is the input data to be normalized, γ is a learnable scaling parameter, and RMS(x) is the root mean square value of the input. RMSNorm is used after each layer to help the model converge better by standardizing the vector length of the hidden state. Unlike the common LayerNorm, RMSNorm does not subtract the mean, has higher calculation efficiency, and is a lighter and more stable normalization method.

[0105]

[0106] Among them, x i represents the i-th data to be normalized, and n represents the number of data to be normalized.

[0107] 3. Output part

[0108] After the final hidden state passes through the linear layer, the general sub-model can output a general representation vector, and through the subsequent target task sub-model (not shown in the figure), the final prediction result can be output. The target task sub-model usually also includes a linear layer Linear, and the linear layer Linear used is a binary classification layer, which is used to classify the predicted abnormal results. The loss function (Loss) is used to measure the difference between the predicted result output by the model and the true label, helping the model to optimize parameters. Finally, after training and optimization, the model will output the final prediction result. In classification problems, the loss function usually uses cross entropy (Cross-Entropy), and the formula can be shown as follows, where y represents the true label (corresponding to the above standard result), Represents the prediction result of the model (corresponding to the above initial result), which measures the difference between the probability distribution of the model output and the true label, and minimizes the loss by maximizing the log-likelihood function.

[0109]

[0110] To reduce the consumption of computing resources, the model can be trained and fine-tuned based on LoRA (Low-Rank Adaptation) technology. LoRA is an efficient fine-tuning technology designed for fine-tuning of LLMs (Large Language Models) to reduce the consumption of training parameters and computing resources. By introducing a low-rank matrix into the model weight matrix, LoRA can quickly adapt and learn new information for specific tasks while keeping the original model parameters unchanged. This approach not only improves the efficiency and flexibility of the fine-tuning process, but also saves a lot of computing resources, while maintaining performance comparable to full parameter fine-tuning in a variety of tasks and fields.

[0111] The model training process can be formalized as the following formula:

[0112] W'=W 0 +ΔW

[0113] Among them, W 0 is the initialization parameter of the model pre-training, and ΔW is the parameter that needs to be updated during model training. If full fine-tuning is performed, the updated parameter size(ΔW)=size(W 0 ), that is, all initialization parameters need to be fine-tuned. This is costly when computing resources are limited. For LoRA technology, only ΔW needs to be fine-tuned. Assuming the original W 0 The dimension is Then the update of model training can be expressed as:

[0114]

[0115] Among them, rank r<<min(d, k); A and B are two low-rank matrices introduced. In the process of using LoRA for training, W 0 Fixed, only low-rank matrices A and B are involved in training. Since the dimensions of low-rank matrices A and B are much smaller than the original parameter dimensions, using LoRA fine-tuning technology can greatly reduce the number of parameters that need to be trained, reduce the demand for computing resources and speed up training.

[0116] The implementation and training of the model can be based on the PyTorch framework. PyTorch is an open source machine learning library in the Python language. It is based on PyTorch and is widely used in applications such as natural language processing. The underlying training will use the unified computing device architecture CUDA (Compute Unified Device Architecture). The CUDA model uses the CPU (Central Processing Unit) as the host to process high-level program logic, and the GPU (Graphic Processing Unit) as the device to process a large amount of parallel data calculations. In order to achieve LoRA fine-tuning, the model uses the PEFT (Performance Enhancement and Fine-Tuning) library during training. The library can be integrated with various large model frameworks and can easily and efficiently fine-tune the parameters of large models on low-resource hardware.

[0117] The above method is verified below. For a total of 30,000 training data, after debugging various learning rates, batch sizes, and training epochs, the final learning rate is set to 0.1, the batch size is 256, and the training epoch is 20 as the training parameters of the model. Under this parameter configuration, the model took 8 hours to complete the training using the computing power of 1 A800, and achieved an accuracy of 0.811 and a recall of 0.824 on the validation set. AUC (Area Under the Curve) is an important evaluation indicator in machine learning, used to measure the performance of binary classification models. It is based on the ROC curve (Receiver Operating Characteristic Curve, a tool for evaluating the performance of classification models), that is, the curve relationship between the true positive rate (TPR) and the false positive rate (FPR). The AUC value ranges from 0 to 1, and the larger the value, the better the classification performance of the model. A key advantage of AUC is that it provides a comprehensive evaluation of the performance of a model at all possible thresholds, which makes it unaffected by the choice of a specific threshold. In addition, AUC has good robustness for class imbalanced data sets, and for the task of discovering medical insurance fraud data, the proportion of positive samples is much lower than that of negative samples, so the AUC indicator is more reliable for measuring the performance of the model. The AUC of the model on the validation set reached 0.911, which proves that the model has excellent classification performance for positive and negative samples.

[0118] like Figure 6The diagram shows the overall training process of a medical insurance anomaly detection model. This solution uses a combination of mathematical statistics, data mining, and artificial intelligence methods to implement a large vertical model for discovering medical insurance fraud data through the "four steps" of data processing, model training, deployment and application, and anomaly expansion. This model systematically analyzes abnormal medical insurance data, mines the potential patterns of abnormal data, and promptly collects, processes, and warns suspicious data, providing a certain basis for the supervision of the medical insurance law enforcement system, reducing the burden of manual review work in the medical insurance law enforcement system, and improving the intelligence of medical insurance business. Figure 6 The design architecture of the overall training process of the model is described, which consists of four technical stages:

[0119] 1. Data processing, that is, the analysis and processing of medical insurance business big data, can use statistical and artificial intelligence analysis methods to accurately locate the key information required for medical insurance business anomaly detection tasks from massive medical insurance business data, and use standardized methods to pre-process the data to obtain sample medical insurance data to ensure data quality and availability.

[0120] 2. Model training, specifically including the construction and training of the medical insurance base model, will build the backbone architecture of the medical insurance element representation model and conduct training. The model as a whole will use the pre-processed sample medical insurance data as input, and train the model in a pre-training and fine-tuning manner to achieve higher detection quality and faster detection speed in the medical insurance business anomaly detection scenario.

[0121] 3. Deployment and application: Deploy the model on the back end of the medical insurance law enforcement system and connect it to the front end user interface. The model will receive abnormal data screened by the law enforcement system and return the corresponding abnormal diagnosis classification and confidence level to provide a basis for manual review, thereby improving the intelligence of the medical insurance system.

[0122] 4. Anomaly Exploration: The last stage will focus on discovering new anomaly scenarios and exploring more application models based on the mature medical insurance anomaly detection model. This will enable a more comprehensive reconstruction and upgrade of medical insurance business, so as to better utilize the massive medical insurance business data and achieve the goal of AI empowering medical insurance business.

[0123] The above-mentioned medical insurance anomaly detection model training method trains the initial model based on sample medical insurance data, so that the obtained medical insurance anomaly detection model can be more suitable for medical insurance detection application scenarios. Detecting the medical insurance data to be detected based on the trained medical insurance anomaly detection model can reduce manual workload and improve detection accuracy and efficiency.

[0124] The embodiment of the present invention provides a device for detecting abnormal medical insurance data. Figure 7As shown, the device includes: a first acquisition module 70, used to obtain a preset full amount of medical insurance data; a screening module 71, used to screen the full amount of medical insurance data according to a preset screening condition to obtain a first screening result; wherein the first screening result includes multiple medical insurance data; an extraction module 72, used to extract a target field and a field value corresponding to the target field from each medical insurance data according to a preset target field; a synthesis module 73, used to synthesize a target prompt word based on the target field and the field value corresponding to the target field; a conversion module 74, used to convert the target prompt word into a vector form to obtain the medical insurance data to be detected; a first output module 75, used to input the medical insurance data to be detected into a pre-trained medical insurance anomaly detection model, and output the detection result corresponding to the medical insurance data to be detected through the medical insurance anomaly detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

[0125] The above-mentioned medical insurance anomaly data detection device can screen the entire amount of medical insurance data, extract the target fields and the corresponding field values, synthesize the target prompt words, and convert the target prompt words into a vector form suitable for medical insurance anomaly detection model analysis. The obtained medical insurance data to be detected is detected using the medical insurance anomaly detection model, which can reduce manual workload and improve detection accuracy and efficiency.

[0126] Furthermore, the medical insurance anomaly detection model includes: a general sub-model and multiple task sub-models; each task sub-model is respectively communicated with the general sub-model; each task sub-model corresponds to a different anomaly detection task; the first output module is also used to: input the medical insurance data to be detected into the general sub-model, and output the general representation vector corresponding to the medical insurance data to be detected through the general sub-model; receive the model selection instruction for the target anomaly detection task, and select the target task sub-model corresponding to the target anomaly detection task from multiple task sub-models according to the model selection instruction; send the general representation vector to the target task sub-model, so as to output the detection result corresponding to the medical insurance data to be detected through the target task sub-model.

[0127] Furthermore, the device is also used to: send the test results to a preset terminal device through a preset API interface, so that the test results can be displayed through the terminal device.

[0128] Furthermore, the device is also used to: if the test result indicates that the medical insurance data to be tested is abnormal data, save the medical insurance data to be tested and the test result corresponding to the medical insurance data to be tested.

[0129] Furthermore, there are multiple target fields; the synthesis module is also used to: if the field values ​​corresponding to each target field are not missing, synthesize the target prompt word based on each target field and the field value corresponding to each target field; if the field value corresponding to the specified field is missing among multiple target fields, process the field value corresponding to the specified field according to a preset missing value processing method to obtain the processed field value; synthesize the target prompt word based on the specified field, the processed field value, and each field except the specified field among multiple target fields and the field value corresponding to each other field.

[0130] The medical insurance abnormal data detection device provided in the embodiment of the present invention has the same implementation principle and technical effects as those in the aforementioned medical insurance abnormal data detection method embodiment. For the sake of brief description, for matters not mentioned in the embodiment of the medical insurance abnormal data detection device, reference may be made to the corresponding contents in the aforementioned medical insurance abnormal data detection method embodiment.

[0131] The embodiment of the present invention provides a training device for a medical insurance anomaly detection model, such as Figure 8 As shown, the device comprises:

[0132] The second acquisition module 80 is used to acquire sample medical insurance data; wherein the sample medical insurance data is pre-marked with standard results;

[0133] The second output module 81 is used to input the sample medical insurance data into the initial model to output the initial results corresponding to the sample medical insurance data through the initial model;

[0134] The third acquisition module 82 is used to calculate the loss value of the standard result and the initial result based on a preset loss function, update the weight parameters of the initial model based on the loss value, and repeat the steps of acquiring sample medical insurance data until the loss value converges to obtain a medical insurance anomaly detection model.

[0135] The training device of the above-mentioned medical insurance anomaly detection model trains the initial model based on sample medical insurance data, so that the obtained medical insurance anomaly detection model can be more suitable for medical insurance detection application scenarios. The medical insurance data to be detected is detected based on the trained medical insurance anomaly detection model, which can reduce manual workload and improve detection accuracy and efficiency.

[0136] The embodiment of the present invention further provides an electronic device, see Fig. 9 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 abnormality data detection method and medical insurance abnormality detection model training method.

[0137] Further, Fig. 9The 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 .

[0138] 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, Fig. 9 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.

[0139] 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.

[0140] 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 abnormality data detection method and medical insurance abnormality detection model training method. The specific implementation can be found in the method embodiment, which will not be repeated here.

[0141] The computer program products of the medical insurance abnormality data detection method and the medical insurance abnormality detection model training method provided in the embodiments of the present invention include a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be found in the method embodiments, which will not be repeated here.

[0142] 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 for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment 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.

[0143] 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, characterized in that: The method comprises: Obtain the preset full amount of medical insurance data; The full amount of medical insurance data is screened according to a preset screening condition to obtain a first screening result; wherein the first screening result includes a plurality of medical insurance data; For each piece of the medical insurance data, according to the preset target field, extract the target field and the field value corresponding to the target field from the medical insurance data; synthesizing a target prompt word based on the target field and a field value corresponding to the target field; Convert the target prompt word into a vector form to obtain the medical insurance data to be tested; The medical insurance data to be detected is input into a pre-trained medical insurance anomaly detection model, and the medical insurance anomaly detection model outputs the detection result corresponding to the medical insurance data to be detected; wherein, the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

2. The method according to claim 1, characterized in that The medical insurance anomaly detection model includes: a general sub-model and multiple task sub-models; each of the task sub-models is respectively connected to the general sub-model in communication; each of the task sub-models corresponds to a different anomaly detection task; The steps of inputting the to-be-detected medical insurance data into a pre-trained medical insurance anomaly detection model, and outputting the detection result corresponding to the to-be-detected medical insurance data through the medical insurance anomaly detection model include: Inputting the medical insurance data to be detected into the general sub-model, and outputting a general representation vector corresponding to the medical insurance data to be detected through the general sub-model; Receiving a model selection instruction for a target anomaly detection task, and selecting a target task sub-model corresponding to the target anomaly detection task from a plurality of task sub-models according to the model selection instruction; The general characterization vector is sent to the target task sub-model to output the detection result corresponding to the medical insurance data to be detected through the target task sub-model.

3. The method according to claim 1, characterized in that The method further comprises: The detection result is sent to a preset terminal device through a preset API interface, so that the detection result is displayed through the terminal device.

4. The method according to claim 1, characterized in that: The method further comprises: If the detection result indicates that the medical insurance data to be detected is abnormal data, the medical insurance data to be detected and the detection result corresponding to the medical insurance data to be detected are saved.

5. The method according to claim 1, characterized in that There are multiple target fields; based on the target fields and field values ​​corresponding to the target fields, the step of synthesizing the target prompt words includes: If the field value corresponding to each of the target fields is not missing, synthesizing the target prompt word based on each of the target fields and the field value corresponding to each of the target fields; If the field value corresponding to the designated field is missing among the plurality of target fields, the field value corresponding to the designated field is processed according to a preset missing value processing method to obtain a processed field value; A target prompt word is synthesized based on the designated field, the processed field value, and the field values ​​corresponding to each other field and each other field in the plurality of target fields except the designated field.

6. A training method for a medical insurance anomaly detection model, characterized in that: The method comprises: Obtaining sample medical insurance data; wherein the sample medical insurance data is pre-marked with standard results; Inputting the sample medical insurance data into the initial model to output the initial results corresponding to the sample medical insurance data through the initial model; Based on a preset loss function, the loss values ​​of the standard result and the initial result are calculated, the weight parameters of the initial model are updated based on the loss values, and the step of obtaining sample medical insurance data is repeated until the loss value converges to obtain a medical insurance anomaly detection model.

7. A medical insurance abnormal data detection device, characterized in that: The device comprises: The first acquisition module is used to obtain the preset full amount of medical insurance data; A screening module, used to screen the full amount of medical insurance data according to a preset screening condition to obtain a first screening result; wherein the first screening result includes a plurality of medical insurance data; An extraction module, for extracting the target field and the field value corresponding to the target field from each piece of the medical insurance data according to a preset target field; A synthesis module, used for synthesizing a target prompt word based on the target field and a field value corresponding to the target field; A conversion module, used to convert the target prompt word into a vector form to obtain the medical insurance data to be detected; The first output module is used to input the medical insurance data to be detected into a pre-trained medical insurance anomaly detection model, and output the detection result corresponding to the medical insurance data to be detected through the medical insurance anomaly detection model; wherein the detection result is used to indicate whether the medical insurance data to be detected is abnormal data.

8. A training device for a medical insurance anomaly detection model, characterized in that: The device comprises: A second acquisition module is used to acquire sample medical insurance data; wherein the sample medical insurance data is pre-marked with standard results; A second output module, used to input the sample medical insurance data into the initial model, so as to output the initial result corresponding to the sample medical insurance data through the initial model; The third acquisition module is used to calculate the loss value of the standard result and the initial result based on a preset loss function, update the weight parameter of the initial model based on the loss value, and repeat the step of acquiring sample medical insurance data until the loss value converges to obtain a medical insurance anomaly detection model.

9. An electronic device, characterized in that: It includes a processor and a memory, 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 abnormality data detection method described in any one of claims 1 to 5, or the medical insurance abnormality detection model training method described in claim 6.

10. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the medical insurance abnormality data detection method described in any one of claims 1-5, or the medical insurance abnormality detection model training method described in claim 6.

Images