CN111105317B - A medical insurance fraud detection method based on drug purchase records - Google Patents

Abstract

The invention provides a medical insurance fraud detection method based on a medicine purchase record, belongs to the field of medicine fraud detection methods, and provides the medical insurance fraud detection method based on the medicine purchase record, which can accurately extract medical insurance fraud information, is convenient to operate and has high applicability. In the invention, a fraudster classification model is constructed through a machine learning algorithm; inputting patient information and medicine purchasing information into a model, and establishing a patient-medicine bipartite graph; according to the patient-medicine bipartite graph, a medicine single-mode projection graph is established to form a medicine chain; dividing a medicine chain into a normal chain and an abnormal chain by using a correlation chain algorithm; calculating the similarity of the normal chain and the abnormal chain through cosine similarity formulas respectively; retaining a comparison combination of the abnormal chain and the normal chain with the similarity of not 0; removing the same products in the abnormal chain and the normal chain in the combination, and retaining other medicines; and synthesizing the rest medicines into a fraud chain, and outputting the fraud chain. The invention is mainly used for detecting the fraudulent behavior of fraudulent patients.

Description

A medical insurance fraud detection method based on drug purchase records

Technical Field

The invention belongs to the field of drug fraud detection methods, and in particular relates to a medical insurance fraud detection method.

Background Art

Although medical insurance fraud is not common, medical insurance fraud often corresponds to abnormal drug purchase records. At the same time, medical insurance fraud cases often have the following characteristics:

(1) Uncommon: Fraud incidents are rare but costly, so the distribution between normal patients and fraudsters is extremely unbalanced.

(2) Knowledge sharing: Fraudsters are often influenced by their allies and contacts, who in turn influence others. Fraud knowledge is transferred and occurs in medical procurement behavior patterns.

(3) Behavioral imitation: Fraudulent patients will also imitate the drug purchasing behavior of normal insured persons to conceal their fraudulent goals and try their best to make their drug purchasing behavior look "normal."

Therefore, there is a need for a medical insurance fraud detection method based on drug purchase records that can accurately extract medical insurance fraud information, is easy to operate, and has strong applicability.

Summary of the invention

In view of the problems that existing medical insurance fraud patterns are diverse, fraud information cannot be accurately determined, and manual extraction of fraud information is cumbersome, the present invention provides a medical insurance fraud detection method based on drug purchase records that can accurately extract medical insurance fraud information, is easy to operate, and has strong applicability.

The technical solution of a medical insurance fraud detection method based on drug purchase records involved in the present invention is as follows:

The present invention relates to a medical insurance fraud detection method based on drug purchase records, which comprises the following steps:

Step S1, constructing a fraudster classification model through a machine learning algorithm;

Step S2, inputting patient information and drug purchase information into the model to establish a patient-drug bipartite graph, wherein the patient information includes normal patients and fraudulent patients;

Step S3: According to the patient-drug bipartite graph, a drug single-mode projection relationship is established to form a drug chain;

Step S4, using an associative chain algorithm to divide the drug chain described in step S3 into a normal chain and an abnormal chain;

Step S5, calculating the similarity of the normal chain and the abnormal chain respectively by using the cosine similarity formula;

Step S6, remove the normal chain with a similarity of 0, and retain the comparison combination of the abnormal chain and the normal chain with a similarity not equal to 0;

Step S7, remove the same products in the abnormal chain and the normal chain in the combination, and keep other drugs;

Step S8: synthesize the remaining drugs into a fraud chain and output the fraud chain.

Further: in step S1, the patient information is integrated, and the feature vector of the patient information is extracted using a machine learning algorithm. The information content IV of each feature is calculated using the supervised screening algorithm smbinning for the feature vector, and the features with information content greater than IV are extracted and input into the machine learning algorithm to obtain a fraudster classification model.

Further: in step S2, the drug purchasing information of normal patients and the drug purchasing information of fraudulent patients with fraudulent behavior are set as patient nodes and drug nodes, and a drug-patient undirected bipartite graph of fraudulent patients and a drug-patient undirected bipartite graph of normal patients are constructed respectively; the first round of derivative feature extraction is performed on the patient-drug bipartite graph, and the features extracted in the first round include the total number of types of drugs used and the total amount of drugs used, and a drug single-mode projection relationship is established based on its derived features.

Further: in step S3, a second round of derivative feature extraction is performed on the abnormal chain, and the features extracted in the second round include the abnormal rate of type, the abnormal rate of quantity and the abnormal drug usage rate in the abnormal chain.

Further: In step S4, the associative chain algorithm is specifically as follows: sort the corresponding matrix of the bipartite graph by edge weight, start with the drug combination corresponding to the highest edge weight, as the starting drug combination in the abnormal chain, further search for the drugs connected by the second highest edge weight in the combined drugs, search in sequence, connect the drug chains together, input the edge weight adjacency matrix, and output a chain.

；其中a、b、c分别为正常链或异常链。Further: In step S5, the cosine similarity formula is

; where a, b, and c are normal chains or abnormal chains respectively.

Further: in step S8, a third round of derivative feature extraction is performed on the synthesized fraud chain, and the features extracted in the third round include the abnormal rate of type, the abnormal rate of quantity and the abnormal drug usage rate in the abnormal chain.

The beneficial effects of the medical insurance fraud detection method based on drug purchase records involved in the present invention are:

The present invention relates to a medical insurance fraud detection method based on drug purchase records. It uses a bipartite graph and a unimodal modal projection relationship derived therefrom, and applies an association chain algorithm to extract fraud pattern transfers and hidden drug purchase targets. It has the advantages of being fast and accurate in terms of business logic and is easy to apply. At the same time, the extraction of fraud chains can help regulatory agencies establish regulatory rules to avoid fraudulent activities and prevent malicious fraud activities of fraudulent patients. The medical insurance fraud detection method analyzes drug purchase records in medical insurance data, uses graph theory algorithms to construct effective derivative features, has a high accuracy in fraud judgment, and can effectively detect variable medical insurance fraud patterns.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a medical insurance fraud detection method based on drug purchase records of the present invention;

FIG2 is a flow chart of a medical insurance fraud detection method based on drug purchase records according to Example 2;

FIG3 is a bipartite graph of medicines for normal patients in Example 2;

FIG4 is a bipartite graph of drugs for fraudulent patients in Example 2.

Implementation

The technical solution of the present invention is further described below in conjunction with the embodiments, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Example 1

This embodiment is described in conjunction with FIG. 1. In this embodiment, a medical insurance fraud detection method based on drug purchase records involved in this embodiment includes the following steps:

，对上述特征使用有监督特征筛选算法smbinning对每个特征的信息量IV进行计算，并提取信息量IV大于0.05的特征投入机器学习算法，获得有效的针对欺诈者的分类模型。Step S1: Construct a fraudster classification model through a machine learning algorithm; integrate patient information, use a machine learning algorithm to extract the feature vector of the patient information, use the supervised screening algorithm smbinning to calculate the information volume IV of each feature of the feature vector, and extract the features with information volume greater than IV and input them into the machine learning algorithm to obtain a fraudster classification model. Use a machine learning algorithm to complete the construction of a fraud detection model. Integrate the feature vector corresponding to the patient and the fraud mark Y corresponding to the patient to obtain

, the supervised feature screening algorithm smbinning is used to calculate the information volume IV of each feature, and the features with information volume IV greater than 0.05 are extracted and input into the machine learning algorithm to obtain an effective classification model for fraudsters.

Step S2, inputting patient information and drug purchase information into the model, establishing a patient-drug bipartite graph, wherein the patient information includes normal patients and fraudulent patients; setting the drug purchase information of normal patients and the drug purchase information of fraudulent patients with fraudulent behavior as patient nodes and drug nodes, respectively constructing a drug-patient undirected bipartite graph of fraudulent patients and a drug-patient undirected bipartite graph of normal patients; performing a first round of derivative feature extraction on the patient-drug bipartite graph, wherein the features extracted in the first round include the total number of types of drugs used and the total amount of drugs used, and establishing a drug single-mode projection relationship based on its derivative features.

：使用药品种类总量 (2)

：使用药品总量；The training data is divided into patients with fraudulent behavior and patients with normal performance. Each patient corresponds to multiple medication records. The multiple medication records of the same patient are sorted out. Based on graph theory, patients and drugs are set as two different types of nodes. The patient node (IDj) and the drug node (Dk) of the medication record are linked together to construct the drug-patient undirected bipartite graph of fraudulent patients and the drug-patient undirected bipartite graph of normal patients. For the constructed drug-patient undirected bipartite graph of fraudulent patients and the drug-patient undirected bipartite graph of normal patients, the first round of features of each patient's drug purchasing behavior are extracted: (1)

：Total number of drugs used (2)

: Total amount of drugs used;

The drug unimodal projection relationships corresponding to the drug-patient bipartite graph of fraudulent patients and the drug-patient bipartite graph of normal patients are deduced respectively, and the associative chain algorithm is used to represent the drug purchasing behavior of fraudulent patients as abnormal chains; the drug purchasing behavior of normal patients is represented as normal chains.

The single-mode projection algorithm of the bipartite graph is mainly used to study the relationship between nodes of the same type. By using the characteristic that one type of node in the bipartite graph is connected through another type of node, the nodes of the same type are directly associated and clustered, thereby generating a growing network graph containing nodes of a single type. The single-mode projection relationship depends on the above-mentioned bipartite graph. First, the drug nodes (Dk) to be studied need to be added to the new network one by one. Taking the added drug nodes as the starting point, other drug nodes connected to the drug nodes through the patients are searched in the bipartite graph according to the node edge weight from low to high, and the drug nodes obtained in the search process are connected to the starting point node. This process is repeated until all drug nodes are connected, forming a new single-mode projection relationship. The two projection relationships are processed using the association chain algorithm to obtain the abnormal chain and normal chain corresponding to the patient's behavior.

The second round of derived feature extraction is performed on the chain. The features extracted in the second round include the abnormal rate of type, the abnormal rate of quantity and the abnormal drug usage rate in the abnormal chain. The second round of patient behavior features are derived based on the abnormal chain corresponding to the fraudulent patient behavior. For each abnormal chain, the following three features are extracted for each patient: (1) the ratio of the number of types of drugs used by each patient that is the same as the number of drugs in the fraudulent chain to the total number of drug types used by each patient; (2) the ratio of the total amount of drugs used by each patient in the fraudulent chain to the total amount of drugs used by each patient; (3) the ratio of the total amount of drugs used by each patient in the fraudulent chain to the types of drugs used by the patient in the abnormal chain.

Step S4, using an associative chain algorithm to divide the drug chain described in step S3 into a normal chain and an abnormal chain; the associative chain algorithm is specifically as follows: sorting the corresponding matrix of the bipartite graph by edge weight, starting from the drug combination corresponding to the highest edge weight, as the starting drug combination in the abnormal chain, further searching for drugs connected by the second highest edge weight in the combined drugs, searching in sequence, connecting the drug chains together, inputting the edge weight adjacency matrix, and outputting a chain.

；其中a、b、c分别为正常链或异常链。Step S5: Calculate the similarity of the normal chain and the abnormal chain respectively using the cosine similarity formula; the cosine similarity formula is:

; where a, b, and c are normal chains or abnormal chains respectively.

Step S6: remove the normal chains with a similarity of 0, and retain the comparison combinations of abnormal chains and normal chains with a similarity of not 0; calculate the cosine similarity between the abnormal chain corresponding to each fraudulent patient behavior and the normal chain corresponding to each normal patient behavior, and remove the comparison combinations of abnormal chains and normal chains with a similarity of 0. Retain other comparison combinations, remove the drugs on the abnormal chain that are the same as the normal chain in the combination, and combine the remaining drugs on the abnormal chain into a fraudulent chain.

Step S7, remove the same products in the abnormal chain and the normal chain in the combination, and keep other drugs;

Step S8, synthesize the remaining drugs into a fraud chain, and perform a third round of derivative feature extraction on the synthesized fraud chain, wherein the features extracted in the third round include the abnormal rate of type, the abnormal rate of quantity, and the abnormal drug usage rate in the abnormal chain; output the fraud chain.

Taking the fraud chain as the benchmark, the third round of patient behavior characteristics is derived. For each fraud chain, the following three characteristics are extracted for each patient: (1) the ratio of the number of drug types used by each patient that is the same as the number of drug types in the fraud chain to the total number of drug types used by each patient; (2) the ratio of the total amount of drugs used by each patient in the fraud chain to the total amount of drugs used by each patient; (3) the ratio of the total amount of drugs used by each patient in the fraud chain to the types of drugs used by the patient in the fraud chain.

Example 2

This embodiment is explained in conjunction with Figures 2, 3 and 4 as well as Example 1. In this embodiment, a medical insurance fraud detection method based on drug purchase records involved in this embodiment uses the medical insurance drug purchase data published by a certain city as a case, and its corresponding medical insurance fraud detection method establishes corresponding patient-drug bipartite graphs for fraudulent patients and normal patients, respectively.

Patients with fraudulent behavior and normal patients are segmented from the training data. Each patient corresponds to multiple medication records, totaling 1,368,148 medication records for 15,000 people. Multiple medication records of the same patient in the records are sorted out. Based on graph theory, patients and drugs are set as two different types of nodes. The patient node (IDj) and the drug node (Dk) are linked together through the medication record to construct the drug-patient undirected bipartite graph of fraudulent patients and the drug-patient undirected bipartite graph of normal patients. In the undirected bipartite graph constructed by medication records, patient nodes can only be connected to each other through drug nodes, not directly to each other; drug nodes can also only be connected to each other through patient nodes, not directly. There is only a simple purchase relationship between drug nodes and patient nodes. Therefore, an undirected bipartite graph is used to represent drug purchasing behavior. The edge weight in the undirected bipartite graph is the number of times the patient purchases the corresponding drug.

：使用药品种类总量，

表示患者节点j连接的药品节点数量；(2)

：使用药品总量，

表示患者节点j连接的药品节点的权重之和。Based on the constructed drug-patient undirected bipartite graph of fraudulent patients and the drug-patient undirected bipartite graph of normal patients, the first round of features are extracted for each patient’s drug purchasing behavior: (1)

: Total number of types of drugs used,

represents the number of drug nodes connected to patient node j; (2)

: Total amount of drugs used,

Represents the sum of the weights of the drug nodes connected to patient node j.

The drug unimodal projection relationships corresponding to the drug-patient bipartite graph of fraudulent patients and the drug-patient bipartite graph of normal patients are deduced respectively, and the associative chain algorithm is used to represent the drug purchasing behavior of fraudulent patients as abnormal chains; the drug purchasing behavior of normal patients is represented as normal chains.

The single-mode projection algorithm of the bipartite graph is mainly used to study the relationship between nodes of the same type. By using the characteristic that one type of node in the bipartite graph is connected through another type of node, the nodes of the same type are directly associated and clustered, thereby generating a growing network graph containing nodes of a single category. The single-mode projection relationship depends on the above-mentioned bipartite graph. First, the drug nodes (Dk) to be studied need to be added to the new network one by one. Taking the added drug nodes as the starting point, other drug nodes connected to it through patients are searched in the bipartite graph from low to high according to the node edge weight. The drug nodes obtained in the search process are connected to the starting point node. This process is repeated until all drug nodes are connected to form a new single-mode projection relationship. Finally, the newly formed single-mode projection relationship is represented by a matrix. The edge weight corresponding to the edge formed by the node connection is the number of patients who use two types of drugs together. The corresponding matrix form is as follows:

药品和

药品的患者数量为

。Among them, m is the number of drugs, and

Medicines and

The number of patients for the drug is

.

Based on this, the drug unimodal projection relationship corresponding to the drug-patient bipartite graph of fraudulent patients and the drug-patient bipartite graph of normal patients is obtained, and the two projection relationships are processed using the associative chain algorithm to obtain the abnormal chain and normal chain corresponding to the patient behavior.

Taking the process of obtaining an abnormal chain by processing the drug-patient bipartite graph of fraudulent patients as an example, first sort the matrix corresponding to the bipartite graph according to the edge weight, and start from the drug combination corresponding to the highest edge weight as the corresponding starting drug combination in the abnormal chain, and further retrieve the drugs connected by the second highest edge weight corresponding to the drugs in the combination. If they exist, connect the retrieved drugs to both sides of the starting drug combination of the abnormal chain according to the corresponding relationship until the connection drug positions corresponding to the second highest edge weight of the drugs on both sides of the chain cannot be retrieved without duplication. Repeat the above steps until all drugs in the single-mode projection relationship are traversed, and finally obtain an abnormal chain combination that carries fraudulent patient information and does not contain duplicate drugs.

The processing method of the drug-patient single-mode projection relationship of normal patients is consistent with the above method, and finally a normal chain combination carrying normal patient information and not containing duplicate drugs is obtained.

条，则对应每个患者可以获得

个衍生特征，标记为

。Based on the abnormal chain corresponding to the fraudulent patient behavior, the second round of patient behavior features is derived. For each abnormal chain, the following three features are extracted for each patient: (1) the ratio of the number of types of drugs used by each patient that are the same as the types of drugs in the fraudulent chain to the total number of types of drugs used by each patient; (2) the ratio of the total amount of drugs used by each patient in the fraudulent chain to the total amount of drugs used by each patient; (3) the ratio of the total amount of drugs used by each patient in the fraudulent chain to the types of drugs used by the patient in the abnormal chain. If the number of abnormal chains obtained is

Each patient can obtain

derived features, labeled

.

Calculate the cosine similarity between the abnormal chain corresponding to each fraudulent patient behavior and the normal chain corresponding to each normal patient behavior, and remove the comparison combinations of abnormal chains and normal chains with a similarity of 0. Keep other comparison combinations, remove the drugs on the abnormal chain that are the same as the normal chain in the combination, and combine the remaining drugs on the abnormal chain into a fraudulent chain. Perform the above operation on each group of abnormal chains and normal chains with a non-zero similarity to solve the corresponding fraudulent chain.

个衍生特征，标记为

。Based on the fraud chain, the third round of patient behavior features is derived. For each fraud chain, the following three features are extracted for each patient: (1) the ratio of the number of types of drugs used by each patient that are the same as the types of drugs in the fraud chain to the total number of types of drugs used by each patient; (2) the ratio of the total amount of drugs used by each patient in the fraud chain to the total amount of drugs used by each patient; (3) the ratio of the total amount of drugs used by each patient in the fraud chain to the types of drugs used by the patient in the fraud chain. If the number of fraud chains obtained is r, then each patient can obtain

derived features, labeled

.

Complete the construction of fraud detection models using machine learning algorithms.

，对上述特征使用有监督特征筛选算法smbinning对每个特征的信息量IV进行计算，并提取信息量IV大于0.05的特征投入机器学习模型，本实施例以逻辑回归算法为例。smbining算法：R语言下的分类处理方法，在本文中目的是对数据集的信息量分类，去除信息量过低的特征向量与欺诈标记的匹配首先，对上述数据集进行十折交叉验证，针对训练集数据使用逻辑回归算法建立关于分类系数的凸优化目标，利用梯度下降法对凸优化目标进行迭代更新，使用ROC、AUC作为模型性能表现的评估变量，比较结果如下表1所示，获得相对性能最佳分类系数向量t，其中，In summary, the feature vector corresponding to the patient and the fraud mark Y corresponding to the patient are integrated to be

, the supervised feature screening algorithm smbinning is used to calculate the information content IV of each feature for the above features, and the features with information content IV greater than 0.05 are extracted and put into the machine learning model. This embodiment takes the logistic regression algorithm as an example. Smbining algorithm: a classification processing method under R language. The purpose of this article is to classify the information content of the data set and remove the feature vectors with too low information content and the matching of fraud marks. First, a ten-fold cross-validation is performed on the above data set. The logistic regression algorithm is used for the training set data to establish a convex optimization target for the classification coefficient. The gradient descent method is used to iteratively update the convex optimization target. ROC and AUC are used as evaluation variables for model performance. The comparison results are shown in Table 1 below. The classification coefficient vector t with the best relative performance is obtained, where

。

.

1 2 3 4 5 6 7 8 9 10 Training AUC 0.86 0.86 0.82 0.85 0.85 0.86 0.86 0.86 0.86 0.85 Test AUC 0.84 0.79 0.8 0.78 0.82 0.8 0.78 0.78 0.81 0.86

Table 1

；According to the classification coefficient vector t, the fraud probability is

;

使用logistic函数计算病患对应的欺诈概率，完成对病患是否为欺诈病患的判定。建立新的信用模型的实现方式有多种。为了得到可行的信用评分模型，本实施例中新的信用模型的属性集为属性集可行域的子集。基于属性集性质进一步确定新信用评分模型使用的算法。目前，可应用于生成信用评分模型的算法种类较多。例如：基于逻辑回归，基于随机森林，基于GBDT等等。在本实施例中，进行算法筛选时包含使用算法融合后的新算法，并按照如下策略实现算法的优选。Logistic回归：研究某一事件发生的概率与若干因素间的关系，可得出事件发生的概率。当概率大于0.5时。可认为其发生，小于0.5时，可认为其不发生。get

The logistic function is used to calculate the fraud probability corresponding to the patient, and the determination of whether the patient is a fraudulent patient is completed. There are many ways to implement a new credit model. In order to obtain a feasible credit scoring model, the attribute set of the new credit model in this embodiment is a subset of the feasible domain of the attribute set. The algorithm used by the new credit scoring model is further determined based on the properties of the attribute set. At present, there are many types of algorithms that can be applied to generate credit scoring models. For example: based on logistic regression, based on random forest, based on GBDT, etc. In this embodiment, the algorithm screening includes the use of a new algorithm after algorithm fusion, and the algorithm is optimized according to the following strategy. Logistic regression: Study the relationship between the probability of an event and several factors, and the probability of the event can be obtained. When the probability is greater than 0.5. It can be considered to occur, and when it is less than 0.5, it can be considered not to occur.

Association chain algorithm: Sort the corresponding matrix of the bipartite graph by edge weight (frequency of occurrence of drug combination), start with the drug combination corresponding to the highest edge weight as the starting drug combination in the abnormal chain, and further search for drugs connected by the second highest edge weight in the combination drug. Search in sequence and connect the drug chains in series. Input the edge weight adjacency matrix and output a chain. As shown in Figures 3 and 4, the bipartite graphs A, B, C, D, and E are normal patients, J, G, S, R, and G are abnormal patients, and a, b, c, d, e, and f are drugs.

Single-mode projection relationship

The edge weight adjacency matrix shows the relationship:

Header (beginning of the drug link) Tail (end of the drug link) Rights (frequency of drug links appearing) aDrugs bMedicine 2 aDrugs cDrugs 1 aDrugs d. Drugs 0 bMedicine cDrugs 4 bMedicine d. Drugs 1 cDrugs d. Drugs 1

The edge weight matrix of the bipartite graph is sorted by edge weight. The frequency of drug combinations is a-b 2 times, a-c 2 times, a-d 0 times, b-c 4 times, b-d 1 time, and c-d 1 time. Starting from the drug combination corresponding to the highest edge weight, that is, b-c is the starting drug combination in the abnormal chain, further search for drugs connected by the second highest edge weight in the combination drug, that is, a-b is twice, b-d is once, so take a-b, search in sequence, connect the drug chains together, and output a-b-c-d, that is, a normal chain. That is, the association chain algorithm outputs a normal chain arranged according to the edge weight.

The edge weight adjacency matrix shows the relationship:

Header (beginning of the drug link) Tail (end of the drug link) Rights (frequency of drug links appearing) bMedicine cDrugs 4 bMedicine e-Drugs 1 bMedicine fMedicine 1 cDrugs e-Drugs 3 cDrugs fMedicine 2 e-Drugs fMedicine 2

The edge weight matrix of the bipartite graph is sorted by edge weight. The frequency of drug combinations is b-c 4 times, b-e 1 time, b-f 1 time, c-e 3 times, c-f 2 times, e-f 2 times. Starting from the drug combination corresponding to the highest edge weight, that is, b-c is the starting drug combination in the abnormal chain, and further searching for drugs connected by the second highest edge weight in the combination drugs, that is, c-f 2 times, c-e 3 times, so c-f is selected, searched in sequence, the drug chains are connected in series, and b-c-e-f is output, that is, an abnormal chain. That is, the association chain algorithm outputs an abnormal chain arranged according to the edge weight.

Because the cosine similarity after vectorization is not 0. Therefore, after removing the drugs b and c in the abnormal chain that are the same as the normal chain, a fraud chain e-f is formed.

Cosine similarity definition: Cosine similarity, also known as cosine similarity, evaluates the similarity of two vectors by calculating the cosine value of the angle between them. Cosine similarity plots the vectors into the vector space according to the coordinate values. The cosine value of the angle between two vectors in the vector space is used as a measure of the size of the difference between two individuals. The closer the cosine value is to 1, the closer the angle is to 0 degrees, that is, the more similar the two vectors are. Conversely, the closer it is to 0, the lower the similarity between the two vectors. This is called "cosine similarity".

，其中a,b,c为正常链与异常链；formula:

, where a, b, c are normal chains and abnormal chains;

Among them, the a vector is [x ₁ , y ₁ ], the b vector is [x ₂ , y ₂ ], the a vector is the vectorization of the normal chain, and the b vector is the vectorization of the abnormal chain. Thus, vectors with a similarity of 0 are removed.

。

.

Claims (3)

1. A medical insurance fraud detection method based on drug purchase records, characterized in that it includes the following steps: Step S1, constructing a fraudster classification model through a machine learning algorithm; Step S2, inputting patient information and drug purchase information into the model to establish a patient-drug bipartite graph, wherein the patient information includes normal patients and fraudulent patients; Step S3: According to the patient-drug bipartite graph, a drug single-mode projection relationship is established to form a drug chain; Step S4, using an associative chain algorithm to divide the drug chain described in step S3 into a normal chain and an abnormal chain; Step S5, calculating the similarity of the normal chain and the abnormal chain respectively by using the cosine similarity formula; Step S6, remove the normal chain with a similarity of 0, and retain the comparison combination of the abnormal chain and the normal chain with a similarity not equal to 0; Step S7, remove the same products in the abnormal chain and the normal chain in the combination, and keep other drugs; Step S8, synthesizing the remaining drugs into a fraud chain, and outputting the fraud chain; In step S2, the drug purchase information of normal patients and the drug purchase information of fraudulent patients with fraudulent behavior are set as patient nodes and drug nodes, and a drug-patient undirected bipartite graph of fraudulent patients and a drug-patient undirected bipartite graph of normal patients are constructed respectively; the first round of derived feature extraction is performed on the patient-drug bipartite graph, and the features extracted in the first round include the total amount of types of drugs used and the total amount of drugs used, and a drug single-mode projection relationship is established based on its derived features; In step S3, a second round of derivative feature extraction is performed on the abnormal chain, and the features extracted in the second round include the abnormal rate of type, the abnormal rate of quantity, and the abnormal drug usage rate in the abnormal chain; In step S4, the association chain algorithm is specifically as follows: sorting the corresponding matrix of the bipartite graph by edge weight, starting from the drug combination corresponding to the highest edge weight as the starting drug combination in the abnormal chain, further searching for drugs connected to the second highest edge weight in the drug combination, searching in sequence, connecting the drug chains together, inputting the edge weight adjacency matrix, and outputting a chain; In step S8, a third round of derivative feature extraction is performed on the synthesized fraud chain, and the features extracted in the third round include the abnormal rate of type, the abnormal rate of quantity and the abnormal drug usage rate in the abnormal chain. 2. According to claim 1, a medical insurance fraud detection method based on drug purchase records is characterized in that, in step S1, patient information is integrated, a machine learning algorithm is used to extract a feature vector of the patient information, the supervised screening algorithm smbinning is used to calculate the information volume IV of each feature of the feature vector, and the features with information volume IV greater than the IV are extracted and input into the machine learning algorithm to obtain a fraudster classification model. 3. A medical insurance fraud detection method based on drug purchase records according to claim 2, characterized in that in step S5, the cosine similarity formula is

; where a, b, and c are normal chains or abnormal chains respectively.

Images