Medical insurance fraud cross identification method based on fusion of multiple machine learning algorithms

Abstract

The invention provides a medical insurance fraud cross identification method based on fusion of multiple machine learning algorithms, and belongs to the field of machine learning technology and insurance risk management cross technology in information technology. The technical problems that in medical insurance fraud recognition, the recognition precision is not high, characteristics are complex, behaviors are hidden, and distinguishing is difficult are solved. According to the technical scheme, the method comprises the following steps of S1, performing data preprocessing on an original data set D0; s2, finding out feature information with high correlation by using the heat map; s3, constructing a medical insurance fraud detection model based on multiple machine learning algorithms; s4, checking the model; and S5, extracting main medical features influencing fraud discrimination through an algorithm. The method has the advantages that the accuracy and efficiency of medical insurance fraud detection are improved, and a more effective risk management solution is provided for insurance companies.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of machine learning technology in information technology and insurance risk management, and in particular to a medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms. Background Art

[0002] After decades of unremitting development, China's social medical security system has gradually matured and become a key way to promote social civilization and improve people's well-being. However, with the continuous expansion of the scale of medical insurance funds and the increasing coverage of the population, medical insurance fraud has also become more and more serious. It not only pushes up medical expenses and causes a huge waste of medical resources, but also seriously threatens the security of medical insurance funds, infringes on the legitimate rights and interests of the people, and poses a severe challenge to the healthy and sustainable development of my country's social medical insurance.

[0003] Given the complexity of medical treatment and the fuzziness of the boundary between fraud and normal medical treatment processes, even legitimate insured persons may exhibit certain atypical behaviors, and fraudsters do their best to disguise themselves, making it increasingly difficult to distinguish fraud from normal medical treatment processes, which undoubtedly increases the difficulty of identifying medical insurance fraud. In addition, medical insurance fraud incidents often involve huge amounts of money and a wide range of people. Coupled with the complexity of medical procedures and the concealment of fraud, the supervision of medical fraud has become particularly difficult.

[0004] Traditional medical fraud identification algorithms mainly rely on artificially set models. These models are built on the basis of sorting out and summarizing existing fraud cases. The construction process requires close cooperation among experts in multiple professional fields such as medical care and data models, and the professional requirements are extremely high. However, the differences in clinical pathways among different doctors, as well as the deficiencies in the collection of pharmacopoeias and the lack of clinical pathway standards in the traditional medical insurance system, have greatly limited the effectiveness of traditional algorithms. To meet this challenge, machine learning algorithms have been introduced into the field of medical fraud identification. These algorithms can deeply mine relevant medical knowledge and treatment rules from massive medical big data, and build a discrimination model for medical insurance fraud in combination with fraud labels. The model built by the machine learning algorithm can monitor new medical insurance data in real time, detect fraudulent behavior in a timely manner, effectively protect the legitimate rights and interests of policyholders, and improve the efficiency of the use of medical insurance funds.

[0005] At present, the model algorithms used in the research of medical insurance fraud identification are indeed of various types, covering BP neural network algorithm, multinomial Bayes, outlier detection, decision tree and clustering and other methods. The use of a single model algorithm in medical insurance fraud identification generally has limitations, which are mainly reflected in the following aspects: First, the model may be too simplified and it is difficult to fully capture the complexity and diversity of fraudulent behavior. The time-series recurrent neural network is based on the comparison model to predict the difference between the patient's medical behavior and the actual medical behavior to determine whether there is fraud. This comparison model is difficult to capture complex fraudulent behavior; second, the model performance is highly dependent on the quality and quantity of training data, and is susceptible to data noise and imbalanced distribution. In the research on medical insurance fraud based on neural network models, the treatment method for data imbalance is too simple, which affects the effect of model learning; third, the model has limited generalization ability and is often only applicable to specific data sets or fraud patterns. It is difficult to adapt to the ever-changing fraud methods. Risk analysis and fraud judgment of insurance policies and claims information through data mining can only be based on specific data. In contrast, the multi-model fusion strategy shows significant advantages. By combining the advantages of different algorithms, multi-model fusion can more comprehensively capture the complex relationships in the data and improve the accuracy and robustness of fraud identification. At the same time, multi-model fusion can also effectively reduce the overfitting risk that a single model may bring and enhance the generalization ability of the model. By utilizing the different sensitivities and interpretability of different algorithms to data, multi-model fusion can identify richer fraud behavior characteristics and patterns, providing more comprehensive and in-depth insights for medical insurance fraud identification. Therefore, the multi-model fusion strategy has broad application prospects and important practical value in medical insurance fraud identification. Summary of the invention

[0006] The purpose of the present invention is to provide a medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms, which uses multiple machine learning algorithms to comprehensively identify a large amount of medical data and monitor users who may commit medical fraud. It is particularly suitable for massive data with more features. The invention aims to improve the accuracy and efficiency of medical insurance fraud detection through innovative machine learning algorithms and technologies, and provide insurance companies with more effective risk management solutions.

[0007] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms, comprising the following steps:

[0008] S1. Perform data preprocessing on the original data set D0. Information overlap will affect the results of fraud factor identification and lead to inaccurate judgment results. First, calculate the correlation coefficient between features, and then draw a heat map of the Pearson correlation coefficient. The heat map shows the correlation between multiple variables, and the depth and change of color are used to indicate the degree of correlation between different variables.

[0009] S2. Use the heat map to find the feature information with high correlation, and eliminate the feature information with significant correlation to reduce the correlation between features. Use the above algorithm to draw the heat map, find the feature columns with correlation coefficient greater than 0.9 and eliminate them, complete the preprocessing of the data set, and the data set after elimination is D1.

[0010] S3. Construction of a medical insurance fraud detection model based on multiple machine learning algorithms. The exported user information is of an unlabeled type and cannot be directly read and identified. The medical information of fraudulent users is an outlier in the data set. Machine learning is used to find outliers or outliers to find relevant data. The accuracy of the judgment results is ensured by constructing a medical insurance fraud model with multiple machine learning algorithms. Eight machine learning algorithms (angle-based outlier detection algorithm, local anomaly factor detection algorithm, high-dimensional data anomaly detection algorithm, histogram-based anomaly detection algorithm, isolated forest anomaly detection method, K nearest neighbor algorithm, local outlier factor detection method, principal component analysis method) are used to detect outliers in the data set D1. The proportion of outliers is set, and each algorithm outputs the corresponding result. If there is possible fraud, the output is 1, and if there is no fraud, the output is 0.

[0011] S3.1, perform angle-based outlier detection on the dataset D1, using the ABOD outlier degree algorithm,

[0012]

[0013] Formula (1) is used to calculate the angles between points. When the angles are acute or obtuse, machine learning determines that the point is an outlier. The outlier output is 1, and the normal point output is 0.

[0014] S3.2, perform outlier detection of local anomaly factors (CBLOF anomaly detection method) on the data set D1, cluster the data, select the Euclidean distance, select a suitable K value for clustering, substitute the position of the point, calculate the distance to the large cluster, and the point that is too far away is an outlier. The outlier point is recorded as 1 and the normal point is recorded as 0.

[0015] S3.3, perform high-dimensional data outlier detection (Feature Bagging) on ​​the data set D1. This detection method is an integrated learning method. It takes features as the detection object and constructs a decision tree model to select features and variance. The combination function is selected to combine the scores after detection. Based on this, the outliers are judged, and the outliers are recorded as 1 and the normal points are recorded as 0.

[0016] S3.4. Perform anomaly detection based on histogram (HBOS) on the dataset D1. The anomaly score of each sample is calculated. The higher the score, the greater the possibility of an anomaly. The following formula is used to calculate the anomaly score of the sample:

[0017] P(p)=P1(p)*P2(p)*P3(P)*…*P a (p) (2)

[0018]

[0019] Specify an anomaly score. Samples with a score higher than the anomaly score are considered to have an abnormal output result of 1, and samples with a score lower than the anomaly score are considered to have an abnormal output result of 0.

[0020] S3.5. Isolation Forest outlier detection is performed on the data set D1. This detection is a parameter-free and unsupervised learning algorithm. It divides the feature space and continuously generates multiple trees and methods for anomaly detection. The cutting point is randomly generated from the given data, and a random plane is generated based on the cutting point to cut the data space. One cutting will produce two subspaces. The ones smaller than the cutting point under the selected dimension are placed on the left branch, and those larger than the cutting point are placed on the right branch. The cutting and classification operations are repeated on the left and right branches respectively, and new leaf nodes are continuously constructed until only one data is left on the leaf node and cutting can no longer be continued. The principle is as follows: Figure 2 As shown. The segmented features are used as the basis for detecting outliers, and outliers are marked as 1 and normal points are marked as 0.

[0021] S3.6, perform K-nearest neighbor outlier detection (KNN) on the dataset D1. It classifies new samples based on the distance measurement between samples. The category of the new sample is determined by the most common category among its nearest k neighboring points. When there is a sample data set, each data of the sample has a classification label. When new data without labels is input, each feature of the new data is compared with the features of the data in the sample set, and finally the classification label closest to the feature in the sample set is extracted. The algorithm principle is as follows Figure 3 As shown, based on this, abnormal points are judged, and abnormal points are recorded as 1 and normal points are recorded as 0.

[0022] S3.7. Perform local outlier factor detection (LOF algorithm) on the data set D1 to identify outliers through density calculation. The LOF algorithm determines the point by calculating the density deviation between the data point and its neighboring points. When the density of a data point is much lower than that of its neighbors, the disk determines that the point is an outlier. Generally speaking, if LOF>1, it is considered an outlier. The distance between the data point and its neighbors is farther than expected. The outlier point is recorded as 1 and the normal point is recorded as 0.

[0023] S3.8, perform principal component analysis on the data set D1, and express the matrix x,

[0024]

[0025] Calculate the eigenvectors and eigenvalues ​​through the matrix:

[0026]

[0027] λ1≥λ2≥…≥λ p ≥0(R is a semi-positive definite matrix, and

[0028] The outliers are identified based on the contribution value of each feature, with abnormal users being marked as 1 and normal users being marked as 0.

[0029] S3.9. When six or more algorithms all output 1 and meet the two additional judgment conditions that the pooled payment amount is large and accounts for a large proportion of the total treatment cost, the algorithm will ultimately determine that there is fraudulent behavior and label the data set as fraudulent or non-fraudulent based on this.

[0030] S4. Model verification. The data set containing all user information is randomly divided into a training set and a test set. The decision tree model and random forest model machine learning algorithm are used as prediction models. After reading the training set, the test set is used for detection and the results are output. The random forest algorithm sorts the important features that affect the fraud structure, and can obtain the important features of medical fraud, which can help better detect the medical behavior of fraudsters. After the test set was tested, the accuracy rate exceeded 90%, which verified the accuracy of the discriminant model.

[0031] S5. The algorithm can also extract the main medical features that affect the identification of fraud and sort them according to the importance of the features, extracting important features to help us further analyze and confirm medical insurance fraud.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention provides a method for comprehensively identifying medical fraud users using multiple machine learning algorithms. Compared with the traditional medical fraud algorithm, which is a model set manually after summarizing the existing fraud models, this solution is not restricted by manpower and computing power, and a large amount of data can be fully utilized to analyze and mine the medical data of the insured, extract behavioral characteristics that may affect medical fraud, and use this as a basis for judgment. The recognition model based on the machine learning algorithm can efficiently process huge and complex medical information, make full use of each medical feature, and deeply mine the user's potential medical behavior and potential fraud clues that may be hidden. By analyzing the data through multiple machine learning algorithms, effective information related to medical fraud can be automatically extracted from the massive data for analysis, which can widely involve information in multiple dimensions such as medical amount, date of consultation, use of drugs, patient information, and consistency of consultation. First, the information is classified to find out the abnormal values, and then the rules and commonalities are found according to the historical fraud data, and the main information and main features for identifying fraud are automatically constructed. It can automatically update the recognition ability of data according to the dynamic changes of data, and even in the face of new fraudulent behaviors, it can maintain high sensitivity and accuracy.

[0034] 2. The present invention utilizes the fusion of multiple machine learning algorithms to cross-identify medical insurance fraud users. The accuracy and credibility of medical insurance fraud users screened and identified by this technology have been significantly improved. Different machine learning algorithms have different advantages when analyzing and processing different data distributions and feature types. In view of the complexity of user medical behavior and the large number and quantity of user information, this technology does not rely solely on a single machine learning algorithm for identification, but uses the fusion of multiple machine learning algorithms for cross-identification. Such fusion technology can give full play to the advantages and strengths of various machine learning algorithms in processing different types of data and features, and realize the complementary advantages of multiple algorithms. It can process linear and nonlinear, high-dimensional data, as well as sparse data or capture local features. This method integrates multiple machine learning algorithms for cross-identification, which helps to optimize model performance, improve the adaptability and robustness of the model when processing complex data, and maintain accuracy in data identification even when facing new unknown data.

[0035] 3. This method can effectively deal with the problem of label imbalance in medical fraud detection and can more accurately distinguish real fraudulent behavior. The medical behavior of users is complex and profound, which leads to extremely unbalanced labels of exported medical user information. The number of people involved in some medical behavior characteristics is less than 10%. Conventional machine learning classification models focus on majority class data and ignore minority class data, which may lead to poor generalization ability of prediction models on minority class data and reduce the accuracy of prediction results. The cross-recognition model that integrates multiple machine learning algorithms can use different algorithms to sample and process different types of data through cross-sampling, so that the model can be trained on different data sets. This method can classify and process samples based on different features and data and different sampling methods, avoiding overfitting and underfitting problems caused by label imbalance during data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0037] Figure 1 The basic technical roadmap for constructing the system model of the present invention.

[0038] Figure 2 This is a schematic diagram of the isolation forest theory in the present invention.

[0039] Figure 3 This is a flow chart of the K-nearest neighbor algorithm used in the present invention.

[0040] Figure 4 Schematic diagram of the machine learning algorithm type in the present invention.

[0041] Figure 5 This is a schematic diagram of the contribution value of each feature to the medical fraud result in the present invention. DETAILED DESCRIPTION

[0042] In order 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. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0043] Example 1

[0044] This embodiment provides a medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms, specifically a method for preprocessing data sets consisting of user medical information, mainly including the following steps:

[0045] S1, export the user's medical information data from the medical system (taking the provided data set as an example, including 82 features of various types such as amount, date, and medicine) to form the original data set D0;

[0046] S2, calculate the correlation coefficient between features, construct the correlation matrix, and draw the heat map of Pearson correlation coefficient;

[0047] S3, eliminating irrelevant features related to medical fraud (eliminating 6), eliminating features with too high correlation (correlation coefficient greater than 0.9) (eliminating 33), the elimination results are shown in Tables 1 and 2;

[0048] S4, after elimination, forms a new data set D 1。

[0049] Table 1

[0050]

[0051] Table 2

[0052]

[0053]

[0054] Example 2

[0055] This embodiment provides a medical insurance fraud cross-recognition method based on the fusion of multiple machine learning algorithms, including the following steps:

[0056] S1. Use eight machine learning algorithms (ABOD outlier measurement algorithm, CBLOF anomaly detection method, FeatureBagging feature integration calculation method, HBOS anomaly detection method, Isolation Forest anomaly detection method, K nearest neighbor algorithm, LOF local outlier factor detection method, principal component analysis method) to detect outliers in the new data set D1. The output result of abnormal users is 1, and the output result of normal users is 0. The above algorithms are used to detect the data using eight machine learning algorithms. The eight algorithms derive the number of fraudulent users as follows: Figure 3 ;

[0057] S2. Construct a model for cross-identification of medical insurance fraud that integrates multiple machine learning algorithms. Among the eight machine learning output results, when the number of result 1 is greater than or equal to 6, preliminarily determine that the user is suspected of medical fraud, set the amount and payment standard requirements for re-screening, and finally determine that the users who meet the requirements are suspected of medical fraud. These users are labeled "F" for suspected fraud; users who are not involved in fraud are labeled "T", forming the result data set D3;

[0058] S3, read the data set D3, divide 85% into training set D test , 15% is recorded as the test set D train ;

[0059] S4, training set D test Input the decision tree prediction model and the random forest prediction model to build a prediction model for medical insurance fraud;

[0060] Using two prediction models to test the set D train After testing, the accuracy is calculated and output. The output accuracy of the decision tree prediction model and the random forest prediction model after verification is shown in Table 3 (taking reference data as an example), which proves the rationality of the prediction model.

[0061] Table 3

[0062] Decision Tree Prediction Model 95.57％ Random Forest Prediction Model 95.46％

[0063] S5. Input the dataset D3 into the random forest model, according to the formula Calculate the importance contribution of each feature to the outcome of medical fraud, and select the first, second, third, ..., mth (m≤P) principal components corresponding to the feature values ​​exceeding 80% to obtain the cumulative contribution rate. The i-th principal component: F i =a 1i X1+a 2i X2+…+a pi X p After the contribution values ​​of (i=1,2,…,m) are arranged from large to small, the top 20 features with the highest contribution values ​​are listed in Table 4, and the pie chart of the generated feature contribution values ​​is shown in Table 4. Figure 5 .

[0064] Table 4

[0065]

[0066] Table 5 Symbols

[0067]

[0068]

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A medical insurance fraud cross-recognition method based on the fusion of multiple machine learning algorithms, characterized in that: The following steps are involved: S1, preprocessing the original data set D0; S2. Use the heat map to find the highly correlated feature information, and remove the significantly correlated feature information to reduce the correlation between the features; S3. Construction of a medical insurance fraud detection model based on multiple machine learning algorithms. The exported user information is of unlabeled type and cannot be directly read and identified. The medical information of fraudulent users is an outlier in the data set. Machine learning is used to find outliers or outliers to find relevant data. S4. Model verification: randomly divide the data set containing all user information into training set and test set, use decision tree model and random forest model machine learning algorithm as prediction model, read in the training set, use the test set for detection, and output the results; random forest algorithm sorts the important features that affect the fraud structure, obtains the important features of medical fraud, and detects the medical behavior of fraudsters; S5. The algorithm extracts the main medical features that affect the identification of fraud, sorts them according to their importance, and extracts important features to further analyze and confirm medical insurance fraud.

2. The medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms according to claim 1 is characterized in that: In step S1, information overlap affects the result of fraud factor identification, resulting in inaccurate determination results. The correlation coefficient between features is first calculated, and then a heat map of the Pearson correlation coefficient is drawn. The heat map shows the correlation between multiple variables, and the degree of correlation between different variables is indicated by the depth and change of colors.

3. The medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms according to claim 1 is characterized in that: In step S2, the heat map is drawn using the above algorithm, and feature columns with correlation coefficients greater than 0.9 are found and removed to complete the preprocessing of the data set. The data set after removal is D1.

4. The medical insurance fraud cross-identification method based on the fusion of multiple machine learning algorithms according to claim 1 is characterized in that: The step S3 comprises the following steps: S3.1, perform angle-based outlier detection on the dataset D1, using the ABOD outlier degree algorithm, Where p is the feature point being evaluated, VAR represents the variance of the calculated expression, and x and y are any two points in the data set that form a vector with p. and represents the vector from point x, y to p, The inner product of the pointing quantity reflects their angular relationship. and They represent the square of the vector norm respectively; Formula (1) is used to calculate the angles between points. When the angles formed are a large number of acute angles or obtuse angles, machine learning determines that the point is an outlier. The output result of the outlier point is 1, and the output of the normal point is 0. S3.2, perform outlier detection of local anomaly factors on the data set D1, use the CBLOF anomaly test method, cluster the data into clusters, select the Euclidean distance, select a suitable K value for clustering, substitute the position of the point, calculate the distance to the large cluster, and the point that is too far away is an outlier. The outlier point is recorded as 1 and the normal point is recorded as 0; S3.3, high-dimensional data outlier detection is performed on the data set D1. This detection method is an integrated learning method. The feature is used as the detection object to select the feature and variance by building a decision tree model. The combination function is selected to combine the scores after detection. The outliers are judged based on this. The outliers are recorded as 1 and the normal points are recorded as 0. S3.

4. Perform anomaly detection based on the histogram for the dataset D1. The anomaly score of each sample is calculated to make a judgment. The higher the score, the greater the possibility of being an outlier. The following formula is used to calculate the anomaly score of the sample: P(p)=P1(p)*P2(p)*P3(P)*...*P a (p) (2) Among them, P(p) represents the score of the user's medical sample, P a (p) refers to the score of a characteristic under specific conditions; Specify an anomaly score. For samples with a higher anomaly score, the user is considered to have an anomaly and the output result is 1. For samples with a lower anomaly score, the output result is 0. S3.

5. Perform Isolation Forest Outlier Detection on the dataset D1. This detection is a parameter-free and unsupervised learning algorithm. It divides the feature space and continuously generates multiple trees and methods for anomaly detection. Randomly generate the cutting point from the given data, and generate a random plane based on the cutting point to cut the data space. One cutting will produce two subspaces. The data smaller than the cutting point under the selected dimension is placed on the left branch, and the data larger than the cutting point is placed on the right branch. Repeat the cutting and classification operations on the left and right branches, and continuously construct new leaf nodes until only one data is left on the leaf node and it is impossible to continue cutting. Use the segmented features as the basis for detecting outliers, and record the outliers as 1 and the normal points as 0. S3.

6. Perform K-nearest neighbor algorithm outlier detection on the data set D1. It classifies new samples based on the distance measurement between samples. The category of the new sample is determined by the most common category among its nearest k neighboring points. When there is a sample data set, each data of the sample has a classification label. When new data without labels is input, each feature of the new data is compared with the features of the data in the sample set. Finally, the classification label closest to the feature in the sample set is extracted. Based on this, the outlier is judged, and the outlier is recorded as 1 and the normal point is recorded as 0. S3.

7. Perform local outlier factor detection on the data set D1 and identify outliers through density calculation. The LOF algorithm determines the point by calculating the density deviation between the data point and its neighboring points. When the density of a data point is much lower than that of its neighbors, the disk determines that the point is an outlier. If LOF>1, it is considered an outlier. The distance between the data point and its neighbors is farther than expected. The outlier point is recorded as 1 and the normal point is recorded as 0. S3.8, perform principal component analysis on the data set D1, and express the matrix x, x is an n×p matrix, which means a two-dimensional array with n rows and p columns, used to represent a multidimensional dataset storing the user's medical information; Calculate the eigenvectors and eigenvalues ​​through the matrix: a p represents the columns of the matrix, represents the eigenvector, a ij Represents the element in row i and column j; λ1≥λ2≥···≥λ k ≥0, tr represents the sum of all eigenvalues, R is a semi-positive definite matrix, p is the order of matrix R, and λ k represents the kth eigenvalue of the matrix R; The outliers are identified based on the contribution value of each feature, with abnormal users being marked as 1 and normal users being marked as 0; S3.

9. When six or more algorithms all output 1 and meet the two additional judgment conditions that the pooled payment amount is large and accounts for a large proportion of the total treatment cost, the algorithm will ultimately determine that there is fraudulent behavior and label the data set as fraudulent or non-fraudulent based on this.

Images