A comprehensive detection method and system for prescription tampering

Abstract

The present invention discloses a comprehensive detection method and system for prescription tampering, comprising: obtaining medical insurance diagnosis and treatment data and extracting characteristic information, and then calculating the prescription quantity and drug dosage changes based on the characteristic information; obtaining descriptive characteristics and trend characteristics by statistically analyzing the prescription quantity and drug dosage changes, and performing primary anomaly detection based on these characteristics; forming a data set with the descriptive characteristics, trend characteristics and introduced behavioral characteristics obtained by the primary anomaly detection, performing secondary anomaly detection on the data set using an isolation forest algorithm to determine and mark initial abnormal data points; performing data reconstruction based on the marked initial abnormal data points using a deep learning model, performing final anomaly detection by comparing the reconstruction error with an abnormal threshold to determine the final abnormal data points, and determining the prescription tampering based on the final abnormal data points, thereby improving the detection accuracy of the prescription tampering.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical health data mining, and in particular relates to a comprehensive detection method and system for prescription tampering behavior. Background Art

[0002] With the existence of prescription tampering in the medical insurance system, prescription tampering includes illegally changing the dosage, type or quantity of drugs, etc. Such behavior not only causes huge economic losses, but also may endanger the health and even life safety of patients. Therefore, it is of great significance to develop effective technology to detect and prevent such behavior.

[0003] Currently, the detection of prescription tampering behavior mainly relies on traditional statistical analysis methods and rule-checking systems. These methods perform well when dealing with simple and obvious prescription tampering behaviors, but often do not work well when faced with complex medical data and carefully designed prescription tampering behaviors. Statistical methods usually require predefined thresholds and rules, which makes them incapable of dealing with new types of tampering behavior. In addition, with the rapid increase in the amount of medical data, traditional methods are low in data processing speed and efficiency, and it is difficult to adapt to the rapid growth and update of medical insurance data.

[0004] Machine learning, especially deep learning technology, has shown great potential in the field of medical data analysis. Unsupervised learning models, such as Isolation Forest, Local Outlier Factor (LOF), One-Class SVM, and autoencoders, have been used to detect various abnormal behaviors, such as the local temperature change anomaly detection method based on deep learning of microwave heating temperature field distribution characteristics disclosed in the patent application with application publication number CN106682685A, which uses the isolation forest algorithm for anomaly detection, and the data flow anomaly detection method and system based on graph deep learning disclosed in the patent application with application publication number CN115080964A, which uses graph deep learning and isolation forest algorithm for anomaly detection. These models can learn from a large amount of unlabeled data and automatically identify abnormal patterns in the data. However, these models usually require a lot of tuning and expertise to achieve optimal performance, and a single model is often difficult to cope with diverse prescription tampering behaviors. Summary of the invention

[0005] In view of the above, the purpose of the present invention is to provide a comprehensive detection method and system for prescription tampering behavior, which is based on statistical analysis, unsupervised learning and deep learning to comprehensively realize accurate detection of prescription tampering behavior, and is suitable for various types of prescription tampering behaviors.

[0006] To achieve the above-mentioned purpose of the invention, an embodiment provides a comprehensive detection method for prescription tampering behavior, comprising the following steps:

[0007] Obtain medical insurance diagnosis and treatment data and extract characteristic information, then calculate the number of prescriptions and changes in drug dosage based on the characteristic information;

[0008] By statistically analyzing the changes in prescription quantity and drug dosage, descriptive characteristics and trend characteristics are obtained, and primary anomaly detection is performed based on these characteristics;

[0009] The descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection constitute a data set, and the isolation forest algorithm is used to perform secondary anomaly detection on the data set to determine and mark initial abnormal data points, wherein the segmentation points used in the isolation forest algorithm are determined based on at least two of the random segmentation strategy, the median segmentation strategy, and the feature importance segmentation strategy;

[0010] A deep learning model is used to reconstruct data based on the marked initial abnormal data points. The final abnormal data points are determined by comparing the reconstruction error with the abnormal threshold for final anomaly detection. The prescription tampering behavior is determined based on the final abnormal data points.

[0011] Preferably, after obtaining the medical insurance diagnosis and treatment data and extracting the characteristic information, the prescription quantity and drug dosage changes are calculated based on the characteristic information, including:

[0012] Obtain medical insurance diagnosis and treatment data and extract characteristic information, including drug purchase information, drug type, amount, prescribing doctor, drug purchase time, and medical insurance designated institution name and code;

[0013] The number of prescriptions and the changes in drug dosage in prescriptions for each patient within a fixed time window are calculated based on the feature information.

[0014] Preferably, descriptive characteristics and trend characteristics are obtained by statistically analyzing the changes in the number of prescriptions and drug dosages, including:

[0015] Descriptive characteristics obtained from statistical analysis of the number of prescriptions and changes in drug dosage include the average drug dosage, the standard deviation of drug dosage, the distribution of drug types based on the number of times and proportion of drugs appearing in prescriptions within a fixed time window, and the number of prescriptions and prescription frequency determined based on the number of prescriptions a patient receives within a fixed time window;

[0016] Trend features include changes in prescription quantity and drug dosage in statistical time windows, where the statistical time windows include 1 day, 3 days, 7 days, one month, and one quarter;

[0017] The introduced behavioral features include historical prescription patterns, the frequency of doctors prescribing medications, the frequency of drug substitutions, the diversity of doctors' prescriptions, and prescription drug combination patterns.

[0018] Preferably, a moving average is constructed and used to smooth the time series data of prescription quantities and the time series data of drug dosages to obtain the changes in prescription quantities and drug dosages over 1 day, 3 days, 7 days, and one month. Among them, the moving average , where k is the size of the moving window, represents the i-th data point, and t represents a certain time point;

[0019] The changes in prescription quantities and drug dosages for each quarter are calculated through the STL decomposition algorithm.

[0020] Preferably, the splitting points adopted in the Isolation Forest algorithm are determined based on at least two of the random splitting measurement, the median splitting strategy, and the feature importance splitting strategy, including:

[0021] Generate a random number r, where 0 ≤ r < 1. If 0 ≤ r < p1, select the random splitting strategy. If p1 ≤ r < p1 + p2, select the median splitting strategy. If p1 + p2 ≤ r < p1 + p2 + p3, select the feature importance splitting strategy. The generated random number r can select at least two strategies according to the above measurement selection method. Among them, p1 is the probability of the random splitting strategy, p2 is the probability of the median splitting strategy, and p3 is the probability of the feature importance splitting strategy;

[0022] Calculate the splitting points under the selected strategies respectively, and perform weighted combination on the splitting points under the selected strategies according to the weight of each strategy to obtain the final splitting point.

[0023] Preferably, the deep learning model adopts a variational autoencoder with an attention mechanism, which is optimized in parameters before being applied. The loss function adopted during parameter optimization includes the reconstruction loss and the KL divergence loss.

[0024] Preferably, the final abnormal data points are determined by comparing the reconstruction error with the anomaly threshold, including:

[0025] Prepare a test data set, input the test data set into the deep learning model, calculate the reconstruction error of each data point in the test data set, and set the anomaly threshold based on the distribution of the reconstruction error;

[0026] Compare the reconstruction error of the initial abnormal data points with the anomaly threshold, and determine the initial abnormal data points higher than the anomaly threshold as the final abnormal data points.

[0027] To achieve the above-mentioned purpose, the embodiment of the present invention provides a comprehensive detection system for prescription tampering, including:

[0028] A data processing module, which is used to obtain medical insurance diagnosis and treatment data and extract characteristic information, and then calculate the prescription quantity and drug dosage changes based on the characteristic information;

[0029] A primary detection module, which is used to obtain descriptive features and trend features by statistically analyzing the changes in the number of prescriptions and drug dosages, and perform primary anomaly detection based on these features;

[0030] A secondary detection module is used to form a data set with descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection, and perform secondary anomaly detection on the data set using an isolation forest algorithm to determine and mark initial abnormal data points, wherein the segmentation points used in the isolation forest algorithm are determined based on at least two of a random segmentation strategy, a median segmentation strategy, and a feature importance segmentation strategy;

[0031] The final detection module is used to reconstruct data based on the marked initial abnormal data points using a deep learning model, perform final anomaly detection by comparing the reconstruction error with the anomaly threshold to determine the final abnormal data points, and determine the prescription tampering behavior based on the final abnormal data points.

[0032] To achieve the above-mentioned purpose of the invention, an embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned comprehensive detection method for prescription tampering behavior.

[0033] In order to achieve the above-mentioned purpose of the invention, an embodiment further provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-mentioned comprehensive detection method for prescription tampering behavior.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] Descriptive features and trend features are determined through statistical analysis, and primary anomaly detection is performed based on these features. Behavioral features are simultaneously introduced into the descriptive features and trend features obtained from primary anomaly detection. These features are used as data sets to perform secondary anomaly detection through the isolation forest algorithm. During detection, the segmentation points are determined by at least two of the random segmentation strategy, the median segmentation strategy, and the feature importance segmentation strategy. This allows for more accurate and efficient screening of primary abnormal data points. Then, based on the initial abnormal data points, the final abnormal data points are ultimately judged through data reconstruction based on a deep learning model. This method combines statistical analysis, unsupervised learning such as the independent forest algorithm, and deep learning to perform anomaly detection, which can not only identify anomalies from different angles, but also verify and complement each other, thereby improving the overall detection accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 is a flow chart of a comprehensive detection method for prescription tampering provided by an embodiment;

[0038] Figure 2 It is a schematic diagram of the structure of the comprehensive detection system for prescription tampering provided in the embodiment. DETAILED DESCRIPTION

[0039] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0040] The inventive concept of the present invention is: in order to cope with the increasingly severe prescription tampering behavior, the present invention aims to provide a system and method for detecting prescription tampering behavior, which is a more comprehensive, efficient and adaptive solution, and combines statistical analysis, unsupervised learning and deep learning technology. Among them, statistical analysis can quickly screen out abnormal candidates, unsupervised learning such as isolation forest algorithm further analyzes the potential abnormal patterns of these candidates, and deep learning model can mine deeper data associations through its powerful feature extraction ability. The system and method can effectively identify, have high efficiency, high accuracy and good adaptability, can process large-scale medical data, and can adapt to the continuous changes in prescription tampering behavior, effectively improving the anti-tampering ability of the medical insurance system.

[0041] like Figure 1 As shown, the embodiment provides a comprehensive detection method for prescription tampering behavior, comprising the following steps:

[0042] S1, obtains medical insurance diagnosis and treatment data and extracts feature information, and then calculates the prescription quantity and drug dosage changes based on the feature information.

[0043] In the embodiment, medical insurance diagnosis and treatment data, including medical insurance settlement data, prescription information data and patient information data, are extracted from the medical insurance database. Sensitive original data in the medical insurance diagnosis and treatment data are desensitized to remove all sensitive information. The desensitized data is transferred to the local database.

[0044] In an embodiment, the desensitized data can also be preprocessed, including: merging the data, associating multiple data tables according to primary keys and foreign keys to form a master table; removing or correcting errors and inconsistencies in the data set to ensure the accuracy and completeness of the data; cleaning the data, filling in the data and cleaning up missing and erroneous data records as appropriate; using statistical methods to identify and process abnormal data that does not conform to the normal distribution; converting and unifying the data to ensure consistency in the data format.

[0045] In the embodiment, feature information is extracted from the pre-processed data, specifically referring to the medical insurance expert's advice, to extract feature information such as drug purchase information, drug type, amount, prescribing doctor, drug purchase time, medical insurance designated institution name and code, etc. The number of prescriptions for each patient within a fixed time window is also calculated based on the feature information to detect frequent prescription activities, and the change in drug dosage in each patient's prescription within a fixed time is also calculated.

[0046] S2, by statistically analyzing the changes in prescription quantity and drug dosage, descriptive characteristics and trend characteristics are obtained, and primary anomaly detection is performed based on these characteristics.

[0047] In the embodiment, descriptive statistical analysis and trend analysis are used to find patterns and anomalies in the data. Descriptive statistical analysis specifically includes:

[0048] (a) Calculate the average drug dose by calculating the average dose of each drug in all prescriptions based on the drug dosage of each drug in a single prescription;

[0049] (b) Count the distribution of drug types, count the number of times each drug appears in the prescription within a fixed time window, and calculate the proportion of each drug in the number of prescriptions to characterize the distribution of drug types;

[0050] (c) Calculate the number of prescriptions and the frequency of prescriptions. For each patient and each doctor, calculate the number of prescriptions within a fixed time window, and determine the number of prescriptions and the frequency of prescriptions based on the prescription data, where the prescription frequency is equal to the ratio of the number of prescriptions within the fixed time window to the total observation time;

[0051] (d) Calculate the variance and standard deviation of drug dosages to assess dosage variability;

[0052] The above-mentioned average drug dosage, standard deviation of drug dosage, distribution of drug types, and number of prescriptions and prescription frequency are collectively referred to as descriptive characteristics.

[0053] In the embodiment, trend analysis is used to identify patterns in data that change over time, focusing on the time series changes in drug dosage or prescription quantity, which are collectively referred to as trend features. Specifically, after setting statistical time windows of different intervals, the curves and fluctuations of drug dosage and prescription quantity over time are analyzed to analyze and identify periodic fluctuations, trends and mutation points. Among them, the statistical time windows include 1 day, 3 days, 7 days, one month, and one quarter.

[0054] Construct a moving average and use it to smooth the time series data of prescription quantity and drug dosage to see the trend clearly, that is, to obtain the changes in prescription quantity and drug dosage for 1 day, 3 days, 7 days, and one month. , where k is the size of the moving window, represents the i-th data point, and t represents a certain time point;

[0055] For the changes in the number of prescriptions and drug dosages in each quarter, seasonal decomposition is performed based on time. If some prescription information in the data is related to the season, STL decomposition, a highly flexible time series analysis tool, is used to decompose the seasonality, trend, and residual components in complex time series data. Its main function is to accurately extract and quantify the long-term trends and cyclical fluctuations in time series data, so that analysts can clearly see the basic behavior and time-dependent structure of the data. Through iterative local weighted regression (Loess), STL decomposition can not only adapt to various seasonal patterns, but also flexibly adjust to handle the nonlinear characteristics of seasonal changes, thus providing a solid foundation for trend prediction and anomaly detection.

[0056] The above features determined by trend analysis and descriptive statistics are used to initially detect abnormal patterns and abnormal points in prescription data. The features obtained by these analyses provide preliminary screening and early warning functions for prescription tampering detection, and provide important features and benchmark data for subsequent unsupervised learning and deep learning models, thereby improving the accuracy and effectiveness of the model. In the embodiment, an example is given to illustrate the primary abnormality detection based on the above features:

[0057] Average drug dosage: After statistical analysis of a certain drug, the average dosage is 100 mg, the standard deviation is 15 mg, and prescriptions exceeding 135 mg are marked as abnormal;

[0058] Drug type distribution: Analysis found that the number of prescriptions for a particular drug suddenly increased in a specific period of time. For example, the average number of prescriptions for a drug was 50 per month, but it increased to 100 in a certain month, showing an unusual pattern;

[0059] Prescription frequency: Calculate the prescription frequency of each doctor. If a doctor writes 50 prescriptions per month on average and the frequency suddenly increases to 100 in a certain month, it may indicate illegal prescription activities.

[0060] Time series trend: Through the moving average, it is found that the drug dosage of a patient has increased significantly in a short period of time, for example, from an average of 50 mg per prescription to 100 mg, indicating possible abuse; however, through trend analysis combined with seasonal correlation, it is found that there is a seasonal correlation between drug dosage, so abuse can be ruled out;

[0061] These features and benchmark data provide the basis for subsequent unsupervised learning and deep learning models, which can help the models more accurately identify abnormal behaviors and potential prescription tampering, while also improving the models' sensitivity to abnormal patterns.

[0062] S3, the descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection constitute a data set, and the isolation forest algorithm is used to perform secondary anomaly detection on the data set to determine the initial abnormal data points and mark them.

[0063] In the embodiment, descriptive statistics and trend analysis are used to perform primary anomaly detection to obtain descriptive features, such as average dose, standard deviation, drug type distribution, prescription frequency, and dosage changes. Trend features within the time window are introduced, such as changes in the number of prescriptions and drug dosages in a season for 1 day, 3 days, 7 days, 30 days, to capture abnormal changes in the short term. Patient and doctor behavior characteristics are also introduced, including historical prescription patterns, the frequency of doctors prescribing drugs, the frequency of drug substitutions, the diversity of doctors' prescriptions, and prescription drug combination patterns. These features constitute the data set of the present invention. During the data set construction process, all features are standardized so that the data are on the same scale in order to improve the convergence and performance of the model.

[0064] For the detection of prescription tampering behavior in the medical insurance scenario, the data has the characteristics of high latitude and large data set. Therefore, the embodiment selects the Isolation Forest algorithm as the unsupervised learning model because it is particularly suitable for processing high-dimensional data and large data sets, and can effectively identify abnormal behavior. Its basic principle is to isolate observations by randomly selecting features and segmentation points. Usually, abnormal points are easier to be isolated.

[0065] When using the Isolation Forest Algorithm as an unsupervised learning model, the model needs to be optimized, and the segmentation strategy is one of the key steps. Data points are isolated by randomly selecting segmentation points in the feature space. Integrating different segmentation strategies can enhance the robustness and detection capabilities of the model.

[0066] 1. Random segmentation strategy: The random segmentation strategy is the basis of the isolation forest algorithm, which isolates data points by randomly selecting features and segmentation points. This method is simple and efficient and is applicable to various data distributions.

[0067] Random Features: ,in represents the features selected in the i-th iteration, and d is the total number of features;

[0068] Randomly choose split points: ,in Indicates the selected split point, Characteristics The minimum and maximum values ​​of .

[0069] 2. Segmentation strategy based on statistical characteristics: This strategy selects segmentation points based on the statistical characteristics of the features, such as segmentation based on the mean, median or quartile, which can more accurately reflect the distribution characteristics of the data.

[0070] In the current medical insurance scenario, medical insurance diagnosis and treatment data contain a large amount of patient information, prescription information and drug information. These data may have the following characteristics: Heterogeneity and asymmetry: The data may have large heterogeneity and asymmetry. For example, there are significant differences in drug usage habits and prescription patterns among different patients. There are outliers: Due to certain prescription tampering behaviors, there may be extreme values ​​or outliers in the data.

[0071] Based on the characteristics of the above data, the median-based segmentation strategy is the best. The main reasons include: 1) Robustness to outliers: The median is not affected by outliers and can better reflect the actual center position of the data, which is very important for capturing normal patterns and abnormal behaviors in the data; 2) Applicable to asymmetric data: The median is applicable to asymmetric data distribution and can more accurately reflect the actual situation of medical insurance diagnosis and treatment data; 3) Computational efficiency: Although the computational complexity of the median is slightly higher, it is still efficient in modern computing environments.

[0072] Therefore, in the segmentation strategy based on statistical characteristics, the embodiment of the present invention adopts a median-based segmentation strategy that is robust to outliers and suitable for asymmetric data characteristics, and its segmentation point is The calculation formula is: ,in It is a feature of the median.

[0073] 3. Segmentation strategy based on feature importance: This strategy first calculates the importance of the feature (for example, variance, information gain, etc.), and then selects the segmentation point based on the importance, focusing more on the segmentation of important features.

[0074] In the prescription tampering behavior detection in the medical insurance scenario, since this task scenario is an unsupervised learning task and there is no clear target variable to calculate the information gain, the present invention selects variance as a measure of feature importance.

[0075] Calculate feature variance ;

[0076] Calculating feature importance Using feature variance as an importance measure, features with larger variance show greater variation in the data and are therefore given higher probability weights when weighted selection of split points; this method can effectively identify features that have a greater impact on model predictions;

[0077] Weighted selection of split points: ,in Is a selection feature The probability of is determined by weighting the probability according to the feature importance to determine the split point.

[0078] 3. Segmentation strategy based on feature importance: This strategy first calculates the importance of the feature (for example, variance, information gain, etc.), and then selects the segmentation point based on the importance, focusing more on the segmentation of important features.

[0079] In the prescription tampering behavior detection in the medical insurance scenario, since this task scenario is an unsupervised learning task and there is no clear target variable to calculate the information gain, the present invention selects variance as a measure of feature importance.

[0080] Calculate feature variance ;

[0081] Calculating feature importance Using feature variance as an importance measure, features with larger variance show greater variation in the data and are therefore given higher probability weights when weighted selection of split points; this method can effectively identify features that have a greater impact on model predictions;

[0082] Weighted selection of splitting point: , where is the probability of selecting a feature The splitting point is determined by weighting the probability according to the feature importance.

[0083] First, set the probability p1 and weight w1 of the random splitting strategy, set the probability p2 and weight w2 of the median-based splitting strategy, and set the probability p3 and weight w3 of the feature-importance-based splitting strategy. Specifically, after testing and tuning the parameters, set p1 = 0.45, w1 = 0.32, p2 = 0.35, w2 = 0.28, p3 = 0.20, w3 = 0.40;

[0084] Then, the random selection strategy includes: generating a random number r, where 0 ≤ r < 1. If 0 ≤ r < p1, select the random splitting strategy; if p1 ≤ r < p1 + p2, select the median-based splitting strategy; if p1 + p2 ≤ r < p1 + p2 + p3, select the feature-importance-based splitting strategy. The generated random number r can select at least two strategies according to the above measurement and selection method;

[0085] Next, calculate the splitting point, including: calculating the splitting points under the selected strategies respectively. Assume that three strategies are selected. Calculate the splitting points θ_random, θ_digits, and θ_importance based on the random splitting strategy, the median-based splitting strategy, and the feature-importance-based splitting strategy respectively;

[0086] Finally, weighted combination of splitting points, including: weighted combination of the splitting points under the selected strategies according to the weight of each strategy to obtain the final splitting point θ_final. Assume that three strategies are selected, then θ_final = w1 * θ_random + w2 * θ_digits + w3 * θ_importance.

[0087] Through this method, combining the combination strategy and the weighting strategy, the final splitting point θ_final takes into account both the randomness of different strategies and reflects the importance of features, thereby enhancing the anomaly detection ability of the Isolation Forest algorithm in complex scenarios.

[0088] In the embodiment, the data points in the data set are divided into two parts according to the final segmentation point θ_final, and isolated trees are continuously constructed to form an isolated forest, and the abnormal score of each data point is calculated, and the data points are sorted by score, and the normal points and abnormal points are distinguished by setting a threshold value, and the data points with specific abnormal scores higher than the threshold value are marked as abnormal data points. The detected abnormal data points are analyzed, their data values ​​and distributions are checked, and potential prescription tampering behaviors are identified and marked. For example, in the present invention, abnormal dosage changes and drug type replacements are found through an isolated forest to determine the prescriptions with tampering behaviors. These marks are used to screen abnormal data points for deep learning model training, enhance the model's recognition performance of marked prescription tampering behaviors, and further screen and clean the noise and abnormal points in the data through the isolated forest algorithm, so that subsequent deep learning models can be trained on cleaner data.

[0089] S4, using the deep learning model to reconstruct data based on the marked initial abnormal data points, performing final anomaly detection by comparing the reconstruction error with the abnormal threshold to determine the final abnormal data points, and determining the prescription tampering behavior based on the final abnormal data points.

[0090] In the embodiment, the deep learning model adopts a variational autoencoder that introduces an attention mechanism. After trend analysis, descriptive statistics and unsupervised learning, the data is screened and marked many times, and the data is preprocessed and input into the variational autoencoder for training. The attention mechanism is added to the model structure to allow the variational autoencoder to automatically learn and distribute the weights of feature importance. In addition, for some text features in the data, the features are converted into low-dimensional vectors by adding an embedding layer to capture the intrinsic relationship. The variational autoencoder is parameter optimized before being applied, and the loss function used in parameter optimization includes reconstruction loss and KL divergence loss, and its reconstruction loss is the difference between the reconstructed data output by the variational autoencoder and the input data, and the KL divergence loss is the difference between the reconstructed data output by the variational autoencoder and the input data.

[0091] After the variational autoencoder is trained, a test data set is prepared and input into the deep learning model. The reconstruction error of each data point in the test data set is calculated, and an abnormal threshold is set according to the distribution of the reconstruction error. The abnormal threshold can be set to the mean of the reconstruction error distribution.

[0092] When using the trained variational autoencoder for anomaly detection, the primary abnormal data points in the data to be detected are obtained through steps S1-S3, and the trained variational autoencoder is used to reconstruct the data. After calculating the reconstruction error, the reconstruction error is compared with the abnormal threshold, and the initial abnormal data points above the abnormal threshold are determined as the final abnormal data points. The prescription tampering behavior can be obtained based on these final abnormal data points.

[0093] Using variational autoencoders as the core model, the variational autoencoder for medical insurance data can better capture abnormal patterns in the data and improve the ability to detect prescription tampering. At the same time, combining trend analysis and descriptive statistics for data analysis and screening, as well as data optimization and preliminary detection results of isolation forests, makes the entire process more efficient and accurate, and can fully identify abnormal behaviors in medical insurance data.

[0094] like Figure 2 As shown, the embodiment also provides a comprehensive detection system 20 for prescription tampering behavior, including: a data processing module 21, a primary detection module 22, a secondary detection module 23, and a final detection module 24, wherein the data processing module 21 is used to obtain medical insurance diagnosis and treatment data and extract feature information, and then calculate the prescription quantity and drug dosage changes based on the feature information; the primary detection module 22 is used to obtain descriptive features and trend features by statistically analyzing the prescription quantity and drug dosage changes, and perform primary anomaly detection based on these features; the secondary detection module 23 is used to form a data set with the descriptive features, trend features and introduced behavioral features obtained by the primary anomaly detection, and perform secondary anomaly detection on the data set using the isolation forest algorithm to determine and mark initial abnormal data points, wherein the segmentation points used in the isolation forest algorithm are determined based on at least two of a random segmentation strategy, a median segmentation strategy, and a feature importance segmentation strategy; the final detection module 24 is used to reconstruct data based on the marked initial abnormal data points using a deep learning model, perform final anomaly detection by comparing the reconstruction error with the abnormal threshold to determine the final abnormal data points, and determine the prescription tampering behavior based on the final abnormal data points.

[0095] It should be noted that when constructing the comprehensive detection system for prescription tampering provided in the above-mentioned embodiment, the division of the above-mentioned functional modules should be used as an example, and the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the functions described above. In addition, the comprehensive detection system for prescription tampering provided in the above-mentioned embodiment and the embodiment of the method for constructing the comprehensive detection of prescription tampering belong to the same concept. The specific implementation process is detailed in the embodiment of the comprehensive detection method for prescription tampering, which will not be repeated here.

[0096] The above-mentioned comprehensive detection method and system for prescription tampering can solve the problems of low efficiency, low accuracy and poor adaptability encountered by existing technologies in prescription tampering detection. Specifically, statistical analysis will be used for preliminary data screening and abnormal indicator generation, unsupervised learning will be used to deeply explore the potential relationships and patterns in the data, and complex data structures will be analyzed and predicted through deep learning models. This comprehensive solution can not only improve the accuracy and efficiency of prescription tampering detection, but also effectively adapt to the continuous evolution of medical insurance fraud.

[0097] The above-mentioned comprehensive detection method and system for prescription tampering behavior can address the shortcomings of existing technologies in processing and analyzing complex and large-scale data sets. Specific statistical analysis methods play an important role in the initial identification and screening of potential fraudulent activities, providing a means to quickly identify abnormal trends and patterns. However, statistical methods alone are not enough to deal with more sophisticated and covert fraud strategies, which is where unsupervised learning technology enters the core stage. By introducing unsupervised learning, especially methods such as the isolation forest algorithm, this system is able to identify abnormal data points without explicit labels. These technologies can reveal subtle abnormal behaviors and fraud patterns by analyzing the inherent structure and relationships in the data.

[0098] The above-mentioned comprehensive detection method and system for prescription tampering behavior, the introduction of deep learning technology, especially the ability to process high-dimensional data and model complex relationships, enables the present invention to understand and predict medical insurance fraud at a deeper level. Deep learning models, such as convolutional neural networks (CNN) and recurrent neural networks (RNN), are particularly suitable for extracting useful information from large amounts of unstructured or semi-structured data. These models can automatically learn complex features from raw data, which are often difficult for traditional algorithms to directly process.

[0099] In summary, integrating statistical analysis, unsupervised learning, and deep learning into a comprehensive detection technology solution not only improves the accuracy and efficiency of detection, but also significantly enhances the adaptability of the system, enabling it to effectively respond to the evolving prescription tampering behavior.

[0100] Based on the same inventive concept, an embodiment further provides a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the method is used to implement the above-mentioned comprehensive detection method for prescription tampering behavior, which specifically includes the following steps:

[0101] S1, obtains medical insurance diagnosis and treatment data and extracts characteristic information, and then calculates the prescription quantity and drug dosage changes based on the characteristic information;

[0102] S2, obtain descriptive characteristics and trend characteristics by statistically analyzing the changes in prescription quantity and drug dosage, and perform primary anomaly detection based on these characteristics;

[0103] S3, using the descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection to form a data set, and using the isolation forest algorithm to perform secondary anomaly detection on the data set to determine the initial abnormal data points and mark them;

[0104] S4, using the deep learning model to reconstruct data based on the marked initial abnormal data points, performing final anomaly detection by comparing the reconstruction error with the abnormal threshold to determine the final abnormal data points, and determining the prescription tampering behavior based on the final abnormal data points.

[0105] The computing device provided in the embodiment, in addition to the processor and memory, also includes hardware required for other services such as internal bus, network interface, memory, etc. at the hardware level. The memory is a non-volatile memory, and the processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the comprehensive detection method of prescription tampering described in S1-S4 above. Of course, in addition to the software implementation, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0106] Based on the same inventive concept, an embodiment further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the above-mentioned comprehensive detection method for prescription tampering behavior is implemented, which specifically includes the following steps:

[0107] S1, obtains medical insurance diagnosis and treatment data and extracts characteristic information, and then calculates the prescription quantity and drug dosage changes based on the characteristic information;

[0108] S2, obtain descriptive characteristics and trend characteristics by statistically analyzing the changes in prescription quantity and drug dosage, and perform primary anomaly detection based on these characteristics;

[0109] S3, using the descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection to form a data set, and using the isolation forest algorithm to perform secondary anomaly detection on the data set to determine the initial abnormal data points and mark them;

[0110] S4, using the deep learning model to reconstruct data based on the marked initial abnormal data points, performing final anomaly detection by comparing the reconstruction error with the abnormal threshold to determine the final abnormal data points, and determining the prescription tampering behavior based on the final abnormal data points.

[0111] In the embodiment, computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data.

[0112] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A comprehensive detection method for prescription tampering, characterized in that: The following steps are involved: Obtain medical insurance diagnosis and treatment data and extract characteristic information, then calculate the number of prescriptions and changes in drug dosage based on the characteristic information; Descriptive features and trend features are obtained by statistically analyzing the changes in the number of prescriptions and drug dosages, and primary anomaly detection is performed based on these features. Descriptive features include average drug dosage, standard deviation of drug dosage, distribution of drug types based on the number of times and proportion of drugs appearing in prescriptions within a fixed time window, and the number of prescriptions and prescription frequency determined based on the number of prescriptions for patients within a fixed time window; trend features include changes in the number of prescriptions and drug dosages within the statistical time window; The descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection constitute a data set. The isolation forest algorithm is used to perform secondary anomaly detection on the data set to determine and mark the initial abnormal data points. The introduced behavioral features include historical prescription patterns, the frequency of doctors prescribing drugs, the frequency of drug substitutions, the diversity of doctors' prescriptions, and the prescription drug combination pattern. The segmentation points used in the isolation forest algorithm are determined based on at least two of the random segmentation strategy, the median segmentation strategy, and the feature importance segmentation strategy, including: Generate a random number r, and 0≤ r < 1. If 0 ≤r < p1, select the random split strategy. If p1 ≤ r < p1+ p2, select the median-based split strategy. If p1 + p2 ≤ r < p1 + p2 + p3, select the feature importance split strategy. The generated random number r can select at least two strategies according to the above measurement selection method, where p1 is the probability of the random split strategy, p2 is the probability of the median-based split strategy, and p3 is the probability of the feature importance split strategy. Calculate the split points under the selected strategies respectively, and weightedly merge the split points under the selected strategies according to the weight of each strategy to obtain the final split point. A deep learning model is used to reconstruct data based on the marked initial abnormal data points, and the final abnormal data points are determined by comparing the reconstruction error with the abnormal threshold for final anomaly detection. The prescription tampering behavior is determined based on the final abnormal data points. The deep learning model adopts a variational autoencoder with an attention mechanism, which is parameter optimized before being applied. The loss functions used in parameter optimization include reconstruction loss and KL divergence loss.

2. The comprehensive detection method for prescription tampering according to claim 1, characterized in that: After obtaining medical insurance diagnosis and treatment data and extracting feature information, the number of prescriptions and changes in drug dosage are calculated based on the feature information, including: Obtain medical insurance diagnosis and treatment data and extract characteristic information, including drug purchase information, drug type, amount, prescribing doctor, drug purchase time, and medical insurance designated institution name and code; The number of prescriptions and the changes in drug dosage in prescriptions for each patient within a fixed time window are calculated based on the feature information.

3. The comprehensive detection method for prescription tampering according to claim 1, characterized in that: The statistical time windows include 1 day, 3 days, 7 days, one month, and one quarter.

4. The comprehensive detection method for prescription tampering according to claim 3, characterized in that: Construct a moving average and use it to smooth the time series data of prescription quantity and drug dosage to obtain the changes in prescription quantity and drug dosage for 1 day, 3 days, 7 days, and one month. , where k is the size of the moving window, represents the i-th data point, and t represents a certain time point; The quarterly changes in prescription quantity and drug dosage were calculated using the STL decomposition algorithm.

5. The comprehensive detection method for prescription tampering according to claim 1, characterized in that: The final abnormal data point is determined by comparing the reconstruction error with the abnormal threshold, including: Prepare a test data set and input the test data set into the deep learning model, calculate the reconstruction error of each data point in the test data set, and set the anomaly threshold based on the distribution of the reconstruction error; The reconstruction error of the initial abnormal data point is compared with the abnormal threshold, and the initial abnormal data point that is higher than the abnormal threshold is determined as the final abnormal data point.

6. A comprehensive detection device for prescription tampering, characterized in that: include: A data processing module, which is used to obtain medical insurance diagnosis and treatment data and extract characteristic information, and then calculate the prescription quantity and drug dosage changes based on the characteristic information; A primary detection module, which is used to obtain descriptive features and trend features by statistically analyzing the changes in the number of prescriptions and drug dosages, and perform primary anomaly detection based on these features, wherein the descriptive features include the average drug dosage, the standard deviation of drug dosage, the distribution of drug types based on the number of times and the proportion of times that drugs appear in prescriptions within a fixed time window, and the number of prescriptions and prescription frequency determined based on the number of prescriptions for patients within a fixed time window; trend features include changes in the number of prescriptions and drug dosages within the statistical time window; The secondary detection module is used to form a data set with descriptive features, trend features and introduced behavioral features obtained by primary anomaly detection, and use the isolation forest algorithm to perform secondary anomaly detection on the data set to determine and mark initial abnormal data points, wherein the introduced behavioral features include historical prescription patterns, frequency of doctors prescribing drugs, frequency of drug substitutions, diversity of doctors' prescriptions, and prescription drug combination patterns. The segmentation points used in the isolation forest algorithm are determined based on at least two of the random segmentation strategy, the median segmentation strategy, and the feature importance segmentation strategy, including: Generate a random number r, and 0≤ r < 1. If 0 ≤r < p1, select the random split strategy. If p1 ≤ r < p1+ p2, select the median-based split strategy. If p1 + p2 ≤ r < p1 + p2 + p3, select the feature importance split strategy. The generated random number r can select at least two strategies according to the above measurement selection method, where p1 is the probability of the random split strategy, p2 is the probability of the median-based split strategy, and p3 is the probability of the feature importance split strategy. Calculate the split points under the selected strategies respectively, and weightedly merge the split points under the selected strategies according to the weight of each strategy to obtain the final split point. The final detection module is used to reconstruct data based on the marked initial abnormal data points using a deep learning model, determine the final abnormal data points by performing final anomaly detection by comparing the reconstruction error with the abnormal threshold, and determine the prescription tampering behavior based on the final abnormal data points. The deep learning model uses a variational autoencoder that introduces an attention mechanism, which undergoes parameter optimization before being applied. The loss functions used in parameter optimization include reconstruction loss and KL divergence loss.

7. A computing device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the one or more processors execute the executable code, they are used to implement the comprehensive detection method for prescription tampering behavior described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the comprehensive detection method for prescription tampering described in any one of claims 1-5 is implemented.

Images