CN109801175A - A kind of medical insurance fraudulent act detection method and device - Google Patents

Abstract

Embodiments of the present invention provide a method and device for detecting medical insurance fraud, wherein the method includes: determining a plurality of medical bill data to be analyzed, and the medical bill data to be analyzed includes a plurality of evaluation indicators; Divide to obtain multiple data subsets; for each data subset, determine a subset weight set corresponding to the data subset; and according to the subset weight set corresponding to the data subset, the medical bills to be analyzed included in the data subset are determined The data is clustered to obtain the cluster members corresponding to the data subset; the ensemble weight set corresponding to the ensemble composed of all the medical bill data to be analyzed is determined; all cluster members are fused according to the ensemble weight set; The medical bill data to be analyzed is isolated, and the isolated medical bill data to be analyzed is regarded as suspicious medical bill data. In this way, the manual workload can be reduced.

Description

A kind of medical insurance fraud detection method and device

technical field

The invention relates to the field of computer technology, in particular to a method and device for detecting medical insurance fraud.

Background technique

As we all know, my country's current medical subsidies are very strong, and the level of people's medical insurance benefits is also increasing. At the same time, some problems in the medical insurance system are becoming more and more obvious. One of the key issues is medical insurance fraud.

Regarding medical insurance fraud, it mainly refers to the acts of citizens, legal persons or other organizations that intentionally fabricate facts, falsify, conceal the true situation, etc., causing losses to medical insurance funds in the process of participating in medical insurance, paying medical insurance premiums, and enjoying medical insurance benefits.

The micro-mechanism of medical insurance fraud is due to the uncertainty of health and disease risks and highly specialized medical services, resulting in serious information asymmetry between consumers and medical service providers. This information asymmetry makes the medical service provider lack the inherent cost restraint mechanism and incentive mechanism, resulting in induced demand, so that the trend of medical expenses cannot be effectively controlled.

In view of the existing research on medical insurance fraud incidents and the analysis of actual data, medical insurance fraud generally manifests in the following three aspects:

1) The price of a single bill is too high: the drug bill does not match the doctor's order (a large number of expensive drugs are prescribed for a minor illness), high-priced bills caused by behaviors such as a large number of purchases of the same drug.

2) The same type of small bills appeared many times: collusion between doctors and patients, splitting a complete and continuous medical service item into multiple service items for charging separately, splitting the bills with excessive fraudulent insurance amount into multiple small prescriptions, etc. This resulted in repeated taking of medicines in a short period of time.

3) Fraudulent use of others' medical insurance cards: Use others' medical insurance cards to handle your own medical insurance business.

In recent years, medical insurance, which should be used as a citizen welfare program, has been abused, causing huge losses to the country. Therefore, through the documents that the insured pays after seeing a doctor in the hospital, their fraudulent behaviors are discovered, and timely accountability and later prevention are carried out to avoid medical treatment. The loss of insurance funds has become an extremely important issue at this stage.

For these aspects, there are many ways to process medical insurance bill data, one of which is to detect the medical insurance bill data through a trained model to obtain whether the medical insurance bill data is suspicious medical bill data, among which, the suspicious medical bill data It is medical billing data that may have medical insurance fraud. In the existing method, in the process of training the model, a large amount of sample medical bill data needs to be manually marked. Specifically, multiple sample documents are obtained, and the sample documents are manually marked. For example, the sample document is marked as a suspicious medical bill. data or not suspicious medical billing data, and then train a model for detection based on multiple sample bills after labeling. It can be seen that, in the existing method, the marking process consumes too much manpower, and the manual workload is large.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present invention is to provide a method and device for detecting medical insurance fraud, so as to reduce the manual workload. The specific technical solutions are as follows:

In a first aspect, an embodiment of the present invention provides a method for detecting medical insurance fraud, including:

determining a plurality of medical bill data to be analyzed, the medical bill data to be analyzed including a plurality of evaluation indicators;

Divide multiple medical bill data to be analyzed to obtain multiple data subsets;

For each data subset, determine the subset weight set corresponding to the data subset; and cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset to obtain the data subset the cluster member corresponding to the set;

Determine the corpus weight set corresponding to the corpus composed of all the medical bill data to be analyzed;

All cluster members are fused according to the corpus weight set, wherein all the cluster members are composed of cluster members corresponding to each data subset respectively;

Determine the isolated medical bill data to be analyzed obtained after the fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data.

In a second aspect, an embodiment of the present invention provides a medical insurance fraud detection device, including:

a first determining module, configured to determine a plurality of medical bill data to be analyzed, wherein the medical bill data to be analyzed includes a plurality of evaluation indicators;

A division module, which is used to divide a plurality of medical bill data to be analyzed to obtain a plurality of data subsets;

a second determination module, configured to determine, for each data subset, a subset weight set corresponding to the data subset;

a clustering module, configured to cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset, and obtain the cluster members corresponding to the data subset;

The third determining module is used to determine the corpus weight set corresponding to the corpus composed of all the medical bill data to be analyzed;

a fusion module, configured to fuse all cluster members according to the corpus weight set, wherein all the cluster members are composed of cluster members corresponding to a plurality of data subsets respectively;

The fourth determination module is configured to determine the isolated medical bill data to be analyzed obtained after fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data.

The medical insurance fraud detection method and device provided by the embodiments of the present invention do not require manual marking during the medical insurance fraud detection process, which can reduce the manual workload. At the same time, the computational efficiency can be improved. Moreover, subjective interference is reduced, and the calculation accuracy can be further improved. Of course, it is not necessary for any product or method to implement the present invention to simultaneously achieve all of the advantages described above.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

1 is a schematic flowchart of a method for detecting a medical insurance fraudulent behavior provided by an embodiment of the present invention;

2 is a schematic flowchart of a fusion process in an embodiment of the present invention;

Figure 3 (a) is a schematic diagram of a subsystem in an embodiment of the present invention;

Figure 3(b) is a schematic diagram of a medical data collection and arrangement module in an embodiment of the present invention;

3(c) is a schematic diagram of a medical data preprocessing module in an embodiment of the present invention;

Figure 3 (d) is a schematic diagram of a module for formulating the weights of evaluation indicators for TCM insurance fraud according to an embodiment of the present invention;

Fig. 3 (e) is the schematic diagram of the cluster member generation module in the embodiment of the present invention;

3(f) is a schematic diagram of a clustering and fusion module in an embodiment of the present invention;

Fig. 3 (g) is the schematic diagram of the output module of TCM insurance fraud result according to the embodiment of the present invention;

Figure 3 (h) is a schematic diagram of a cluster member storage module in an embodiment of the present invention;

4(a) is a schematic flowchart of a specific embodiment provided by an embodiment of the present invention;

Figure 4(b) is a schematic diagram of data flow in a specific embodiment of the present invention;

5 is a schematic structural diagram of a medical insurance fraud detection device provided by an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a medical insurance fraud detection device provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The method for detecting medical insurance fraud provided by the embodiment of the present invention can be applied to an electronic device. Specifically, the electronic device may include a terminal, a server, a processor, and the like.

An embodiment of the present invention provides a method for detecting medical insurance fraud, as shown in FIG. 1 , which may include:

S101. Determine a plurality of medical bill data to be analyzed.

The medical bill data to be analyzed includes multiple evaluation indicators. Evaluation indicators may include total billing price, type of medicine, number of times of taking medicine, type of doctor's order, number of the doctor who issued the doctor's order, number of medical insurance booklet, executive department, number of executive doctor, patient's death sign and/or patient's ID number, and so on.

Specifically, determining a plurality of medical bill data to be analyzed may include:

A plurality of original medical bill data are obtained; for each original medical bill data, the original medical bill data is preprocessed to obtain medical bill data to be analyzed corresponding to the original medical bill data.

The raw medical billing data may be medical billing data obtained from a hospital database. In an implementation manner, after obtaining the original medical bill data, a data sorting process may also be included, such as sorting the original medical bill data obtained from the hospital database into data in a predetermined format, so as to facilitate subsequent processing.

Preprocessing can include a series of operations such as medical data cleaning, medical data integration, medical data dimensionality reduction, and medical data standardization.

S102: Divide a plurality of medical bill data to be analyzed to obtain a plurality of data subsets.

Divide a plurality of medical billing data to be analyzed into a plurality of data subsets.

In an implementation manner, a plurality of medical bill data to be analyzed are randomly sampled to generate data subsets of the same magnitude. If a total of 150 medical bill data to be analyzed are obtained, and each 50 of them is divided into a data subset, 3 data subsets can be obtained.

S103, for each data subset, determine a subset weight set corresponding to the data subset; and cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset to obtain the The cluster member corresponding to the subset of data.

Determining the subset weight set is to determine the weight value for each evaluation index of the medical bill data to be analyzed in the data subset. In an implementation manner in the embodiment of the present invention, the weight value corresponding to each evaluation index is randomly determined; or the manually determined weight value corresponding to each evaluation index can be obtained. In another implementation manner, the subset weight set corresponding to the data subset may be determined adaptively based on the medical bill data to be analyzed included in the data subset.

The process of clustering data subsets can be understood as clustering according to the similarity between the medical bill data to be analyzed included in the data subsets. For example, it can be understood that the similarity is relatively large, such as greater than the similarity threshold. A cluster member is obtained by clustering multiple medical bill data to be analyzed.

S104: Determine a corpus weight set corresponding to a corpus composed of all medical bill data to be analyzed.

The process of determining the corpus weight set is similar to the process of determining the subset weight set corresponding to the data subset, and the specific determination process may refer to the process of determining the subset weight set corresponding to the data subset. The difference is that the process of determining the corpus weight set is based on the corpus composed of all the medical billing data to be analyzed. In an optional embodiment of the present invention, a corpus weight set corresponding to the corpus may be determined adaptively based on a corpus composed of all medical bill data to be analyzed.

S105, fuse all cluster members according to the corpus weight set.

Among them, all cluster members are composed of cluster members corresponding to each data subset respectively.

In order to improve the disadvantages of using a single clustering algorithm, a fusion process is introduced. Specifically, the similarity between the cluster members may be calculated, and the similarity greater than the threshold is fused according to the similarity between the cluster members. Wherein, the similarity between cluster members can be calculated by calculating the similarity of the cluster centers of the cluster members, or the similarity of the edge data of the cluster members.

S106: Determine the isolated medical bill data to be analyzed obtained after the fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data.

Since the proportion of medical insurance fraud is much smaller than the normal use of medical insurance, the isolated points in the results obtained after clustering and fusion are used as regulations for suspected medical insurance fraud, and the isolated medical bill data to be analyzed after fusion is used as the Suspicious medical billing data.

After determining the isolated medical bill data to be analyzed obtained after the fusion, and using the isolated medical bill data to be analyzed as suspicious medical bill data, it may further include:

Display suspicious medical billing data.

In addition, the suspicious medical bill data can be manually reviewed, and the persons involved in the actual medical insurance fraud can be held accountable, so as to realize the whole-process and all-link intelligent detection of medical insurance fraud.

In the embodiment of the present invention, manual marking is not required in the process of detecting medical insurance fraud, which can reduce manual workload. At the same time, the computational efficiency can be improved. Moreover, subjective interference is reduced, and the calculation accuracy can be further improved.

In addition, in the embodiment of the present invention, for the medical insurance fraud scenario, the large data set is divided into multiple small data subsets to perform multi-threaded parallel computing, so as to speed up the computing process.

In an optional embodiment of the present invention, S103: Determine the subset weight set corresponding to the data subset, which may include:

According to each medical bill data to be analyzed included in the data subset, a first evaluation index weight function is constructed; according to the first evaluation index weight function, a particle swarm optimization algorithm is used to determine a subset weight set corresponding to the data subset.

The subset weight set includes first weight values corresponding to each evaluation index respectively.

Specifically, the data subset is set as X _a ={X _a1 , X _a2 ,...,X _an }, (1≤i≤n), where X _a1 , X _a2 ,..., X _an are the The n medical bill data to be analyzed included in the data subset, X _ai = (x _ai1 , x _ai2 , ..., x _aim ), x aik ( _k∈ [1, m]) is the kth evaluation of X _ai index. The weight ω _a = (ω _a1 , ω _a2 , . . . , ω _am ) of the evaluation index is introduced, where ω _ak (k∈[1, m]) represents the weight of the k-th dimension evaluation index in the a-th data subset. The corresponding Euclidean distance calculation formula is obtained to represent the similarity degree _daij between the medical bill data to be analyzed X _ai and the medical bill data X _aj to be analyzed on the data subset.

Then, a function S _aij representing the degree of similarity is introduced, and the similarity function is defined as:

where the value of γ is determined by:

Among them, when _Saij approaches 1, it means that the degree of similarity is greater, and the distance between two points is smaller; when _Saij is close to 0, it means that the similarity of data is lower, and the distance between two points is larger; When _Saij is around 0.5, it means that the ambiguity is large.

In order to make the clustering result have the property of relatively less ambiguity, by adjusting the attribute weight, the distance between similar data is reduced, and the distance between dissimilar data is increased, that is, an attribute evaluation function is found to comprehensively evaluate the distance between each point. similarities, so that the overall ambiguity is minimized. The similarity relationship matrix is optimized so that the similarity relationship function S _aij of data with greater similarity tends to be close to 1, and similarly, the similarity relationship function of data with small similarity is made to approach 0. For this purpose, the first evaluation index weight function is introduced, which is defined as follows:

Among them, when Weight _a (ω) approaches 0, it means that the ambiguity is the smallest.

In order to minimize the Weight _a (ω) function, the particle swarm optimization algorithm is used, and the particle swarm optimization algorithm with slower convergence speed but obvious characteristics is used in parallel for the Weight _a (ω) generated by the data subset with a small amount of data. The process is as follows:

(1) Initialization stage: First, the dimension of the target space is defined as the dimension k of the weight of the medical insurance fraud evaluation index, the initial position and initial velocity of the particles are the default values, and finally the size of the particle swarm is set according to the number of data subsets. the maximum number of iterations.

(2) Calculate the individual value and the global optimal solution: the individual extreme value is the optimal solution found by each particle, and a global value is found from these optimal solutions, which is called this global optimal solution. Compare with the historical global optimum and update it.

(3) Update particle velocity and position.

(4) Judging whether the termination condition is reached: when the maximum number of iterations is reached, the iteration is stopped, and the result is output.

Through the above-mentioned particle swarm optimization algorithm flow, the minimum value of the Weight _a (ω) function can be obtained, and then the weight ω _a of each evaluation index to be optimally assigned can be obtained.

In an optional embodiment of the present invention, S104: Determine a corpus weight set corresponding to a corpus composed of all medical bill data to be analyzed, which may include:

According to all the medical bill data to be analyzed, a second evaluation index weight function is constructed; according to the second evaluation index weight function, a corpus weight set corresponding to the corpus is determined through a differential evolution algorithm.

Wherein, the ensemble weight set includes second weight values corresponding to each evaluation index respectively.

After merging each data subset, the complete set is X={X ₁ , X ₂ ,..., X _n }, (1≤i≤n), X _i =(x _i1 , x _i2 ,...,x _im ), x _ik is the k-th evaluation index of Xi _, and the evaluation index weight Ω=(Ω ₁ , Ω ₂ , ..., Ω _m ) is introduced, where Ω _k (k∈[1, m]) represents medical insurance The weight of the k-th dimension evaluation index in the complete set of billing data, and finally the weight function of the second evaluation index is obtained as:

in,

For the Weight _a (ω) generated by the data set with a large amount of data, the differential evolution algorithm with fast convergence speed and easier to achieve the global optimal solution is used. The process is as follows:

(1) Set the basic parameters including population size as 100, scaling factor as 0.5 and crossover probability as 0.8.

(2) Initialize the population, set the dimension as the dimension k of the weight of the medical insurance fraud evaluation index, and the initialization algebra is 1.

(3) Calculate the population fitness value.

(4) When the termination condition is not satisfied, a loop is performed, and mutation, crossover, and selection operations are performed in sequence until the operation is terminated.

Finally, the weight Ω of the evaluation index based on the data set is obtained.

In the embodiment of the present invention, two targeted algorithms, particle swarm optimization and differential evolution algorithm, are respectively used for a data subset with a small amount of data and a complete data set with a large amount of data, so that the data with a small amount of data can be highlighted. Data characteristics, data with a large amount of data can quickly converge, and the obtained evaluation index weight sets are respectively applied to cluster member generation and cluster fusion to improve the final cluster fusion performance.

In an optional embodiment of the present invention, S103: Cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset, and obtain the cluster members corresponding to the data subset , which can include:

A1, based on each evaluation index, respectively determine the sub-similarity between each pair of medical bill data to be analyzed included in the data subset.

A2, according to the first weight value corresponding to each evaluation index included in the subset weight set, weight the sub-similarities corresponding to each evaluation index determined based on each evaluation index respectively, and obtain each pair of medical bill data to be analyzed. the total similarity between them.

A3, according to the total similarity, perform clustering on each medical bill data to be analyzed included in the data subset to obtain cluster members corresponding to the data subset.

Specifically, the weighted Canopy rough clustering is mainly used in parallel for each data subset, and then the weighted agglomeration hierarchical algorithm is performed according to the obtained cluster center points and the number of clusters.

Use Canopy coarse clustering, define two thresholds T ₁ and T ₂ , and let T ₁ >T ₂ ;

(1) Randomly select a point x from the medical bill data to be analyzed, and calculate the distance _daij from the point in the data subset a to other medical bill data to be analyzed by the following weighted Euclidean distance.

(2) If it is judged that _daij <T ₁ , it means that it is a weak correlation, and these points are classified into one category;

(3) Continue to judge, if _daij <T ₂ , it means that there is a strong correlation, and it is not necessary to perform calculation again to remove these points from the data subset.

(4) Repeat the above three steps until the data subset is empty. At this time, the number of categories W _a and the cluster center C _aw of each category can be quickly obtained.

Then a weighted agglomerative hierarchical clustering algorithm is performed on a parallel computing unit to perform deep clustering on the data.

(1) Consider each data in the sample space as a class cluster, and set a total of n classes at this time, and take the average weighted distance between class p and class q as the distance between the two classes.

n _ap , n _aq represent the amount of data contained in class p and class q in data subset a, respectively.

(2) In each iteration, merge the two classes into one class. The two selected classes are the ones with the smallest average connection. That is, according to our chosen distance metric, the two classes have the smallest d(p,q) distance between them, and thus are the most similar, and are merged together.

(3) Repeat the above steps continuously, and finally stop the clustering after obtaining the number of Canopy coarse clustering categories W _a , a parallel unit finally obtain a cluster member M _a .

(4) Integrate the W _a clusters in the a cluster members Ma obtained after deep clustering and the cluster center C _aw in a one-to _- one correspondence, and take the cluster center C _aw as the center of the cluster W _a where it belongs. , if there are multiple centers C _aw in the same cluster or not in any cluster, which cannot correspond, return the corresponding cluster data and cluster center to the coarse clustering unit to perform coarse clustering again to determine the cluster center c _aw and the number of clusters w _a are returned to the current agglomerative hierarchical clustering unit to perform deep-level clustering, and the w _a clusters obtained from the clustering results correspond to the clustering center c _aw again, and this step is repeated continuously until it is completed. Clusters correspond one-to-one with cluster centers. Finally, the results are combined to obtain W _a clusters with C _aw as the cluster center in a cluster member M _a .

(4) Each point in the cluster is X _awi , which represents the i-th data in the k-th cluster in the a-th cluster member, X _awi =(x _awi1 , x _awi2 ,..., x _awim ), x _awik (k∈[1, m]) is the k-th evaluation index of X _awi , and calculates the distance d _awi from each cluster center C _aw to the cluster center C aw , which represents the a-th cluster member in the The distance of the i-th data in the k-th cluster from the cluster center in its cluster.

By comparing the similarity of the edge distances of the same cluster, the points with farther edge distances and a small number of points are defined as b isolated points Slolitary _ab . Each class in each cluster member can get a nearest point distance d _awmin and farthest edge distance d _awmax to prepare for the subsequent fusion strategy.

In the embodiment of the present invention, the Canopy rough clustering and agglomerative hierarchical clustering algorithms are used for the clustering training of the cluster members for the medical insurance fraud scenario, so that the cluster center and the number of clusters can be quickly obtained, and there is no need to pre-set the clustering. The number of clusters, step-by-step iteration produces the best visible effect. Moreover, the cluster members generated for the data subset have a hierarchical structure and can restore its hierarchical structure, so that each iteration process in the clustering can be restored, which is convenient for later analysis.

In an optional embodiment of the present invention, in S103: for each data subset, determine a subset weight set corresponding to the data subset; and according to the subset weight set corresponding to the data subset, the data subset includes After clustering the to-be-analyzed medical bill data to obtain the cluster members corresponding to the data subset, the method further includes:

Save the cluster members corresponding to each data subset.

For the newly generated small data sets that need to be discovered for medical insurance fraud information, the original training set of the same structure and type can be used to train cluster members for joint fusion, which saves the process of retraining all data sets, and has better performance. Extensibility.

In an optional embodiment of the present invention, in S105, all cluster members are fused according to the corpus weight set, which may include:

B1, determine the fusion strategy.

The process of determining the fusion strategy can also be understood as the process of determining the consensus function. The embodiment of the present invention does not limit the manner of determining the fusion strategy, and any manner that can realize the fusion of each cluster member falls within the protection scope of the embodiment of the present invention.

B2, based on each evaluation index, determine the similarity between each cluster center pairwise, wherein each cluster center corresponds to each cluster member in all cluster members respectively.

Specifically, in the process of determining the cluster members, for each cluster member among all the cluster members, the cluster center of the cluster member is determined.

B3, according to the second weight value corresponding to each evaluation index included in the ensemble weight set, and the second similarity between each cluster center pairwise, all cluster members are fused through a fusion strategy.

In order to improve the disadvantages of using a single clustering algorithm, a clustering fusion algorithm is introduced, as shown in Figure 2. Specifically, the consensus function strategy is determined as follows:

(1) For a cluster member M _a and m medical insurance fraud evaluation index weights Ω=(Ω ₁ , Ω ₂ , ..., Ω _m ) based on the complete set of medical insurance billing data from the cluster member generation module , where Ω _k (k∈[1, m]) represents the weight of the k-th medical insurance fraud evaluation index, and calculates the distance d _egfh between the cluster centers C _eg and C _fh respectively (here a takes e and f to represent two Cluster member, w takes g and h represents the cluster cluster in the two cluster members).

If the distance d _egfh between the two cluster centers is less than one of the closest point distances d _ewmin and d _fwmin between the two cluster centers, the two clusters are classified as a class of clusters, and one of the cluster centers is defined as The fused cluster centers, and the closest point distance and the farthest edge distance are updated according to the new cluster centers. If the conditions are not met, no fusion is performed, and the process is repeated until all clusters are merged to obtain new multiple cluster centers Cnew _w , distances from the nearest point dnew _wmin and distances from the farthest edge dnew _wmin .

(2) For the isolated points Slolitary _ab that appear in the cluster members, calculate the distance d _awb for each new cluster center Cnew _w respectively, indicating the distance between the isolated point b in the previous cluster member a and the cluster center of the new cluster cluster w distance.

For outliers whose distance is less than the distance dnew _wmin of the farthest edge, it is classified into this cluster. If the condition is not met, it is still regarded as an isolated point, and the final isolated point Snew _b is obtained. The isolated point can also be understood as isolated medical bill data. In this way, the final fusion result can be obtained.

In the embodiment of the present invention, a weighted clustering fusion algorithm is introduced in the process of medical insurance fraud detection. Compared with the traditional clustering algorithm that only considers the clustering algorithm, the isolated point and cluster edge data can be re-analyzed. The stability is better.

In the embodiment of the present invention, the above-mentioned process of detecting medical insurance fraud may be implemented through different subsystems, wherein different subsystems may include modules.

Specifically, as shown in FIG. 3( a ), it may include a data acquisition subsystem, a data analysis subsystem, a result display terminal, and a cluster member storage area.

The data collection subsystem mainly includes a module: the medical data collection and arrangement module, which is responsible for the collection and arrangement of medical insurance documents into a specific data format for subsequent data processing and use.

The data analysis subsystem mainly includes four modules: the medical data preprocessing module mainly performs medical data preprocessing. The medical insurance fraud index weight formulation module performs adaptive weight division for each evaluation index after preprocessing, and completes the weight distribution for different evaluation indexes. The generating module of cluster members is responsible for the function of generating multiple cluster members in the fusion algorithm. The cluster fusion module completes the fusion operation in the cluster fusion algorithm for multiple cluster members, and obtains the final data cluster fusion result.

The result display terminal mainly includes one module: the medical insurance fraud result output module, which screens the outliers in the clustering fusion results, and displays the outlier information in the data clustering results and the relevant medical insurance document regulations on the terminal as a manual review. The basis for the prosecution of the persons involved.

The cluster member storage area mainly includes a module: the cluster member storage module, which is responsible for storing the cluster members, interacting with the cluster member generation module, and completing the reuse of the cluster members.

After completing the basic setting, the whole system can realize the whole-process and whole-link intelligent medical insurance fraud detection, and the data exchange between the four subsystems is carried out through the data interface.

The different parts are described below. For simplicity of introduction, the medical billing data is simply referred to as medical data or data in the following description.

As shown in Figure 3(b), the data acquisition subsystem includes one module: the medical data acquisition and arrangement module. This module mainly extracts data from medical billing databases without human intervention and converts it into a format suitable for input.

This module is mainly composed of 4 functional units, which are hospital database interface, data interaction control unit, medical form data sorting and format conversion unit and specified format data interface.

Hospital database interface: responsible for encapsulating the hospital database, shielding the differences in data management methods of different databases; connecting with the hospital database, providing data reading and query functions for the data sorting unit in the form of a simple interface, and using services for the service request unit Data is provided easily.

Data interaction control unit: responsible for controlling and coordinating other units to jointly complete the data interaction function. As the control core of this module, it controls the hospital database interface to send service requests and interactive data to the hospital database, and forwards the data obtained from the hospital database to the data sorting and format conversion unit, and controls the data sorting and format conversion unit to process the data. Format conversion.

Medical data sorting and format conversion unit: Responsible for sorting and converting the data format according to the subsequent preprocessing requirements, ensuring the consistency of the data format, and outputting multiple data in a fixed format. This unit accepts the control instructions of the data exchange unit, and organizes and converts the data sent to the data preprocessing module.

Specified format data interface: responsible for exchanging information with the data preprocessing module, accepting control instructions from the data interaction control unit, and sending data in specified format to the preprocessing module of the data analysis subsystem.

Among them, the format conversion of data can also be understood as a process in the process of data preprocessing.

The data analysis subsystem includes four modules: medical data preprocessing module, medical insurance fraud evaluation index weight formulation module, cluster member generation module and cluster fusion module.

As shown in Figure 3(c), the medical data preprocessing module mainly completes the preprocessing function of medical data. The machine learning preprocessing algorithm is used to convert the original data into multiple data subsets divided according to a single standardized low-dimensional medical form data. for processing by subsequent modules.

This module is mainly composed of 7 functional units, namely service and data interface, data interaction control unit, medical data cleaning unit, medical data integration unit, medical data standardization unit, medical data dimensionality reduction unit and random sampling unit.

Service and data interface: responsible for the interaction between this module and the data collection and sorting module of the data collection subsystem, and transmits multiple obtained medical data in a fixed format to the medical data cleaning unit for subsequent processing.

Data interaction control unit: responsible for the control service and data interface to receive structural data from the data acquisition and sorting module, and forward the structural data to the medical data cleaning unit for data preprocessing. Preprocessing mainly includes cleaning, integration, standardization and dimensionality reduction of medical data.

Medical data cleaning unit: responsible for cleaning the data, and remove the data with a large number of missing evaluation indicators in the obtained data; remove data indicators such as home address, place of origin, age that are useless for this scenario; remove the format error indicators appear data; for repeated data, its redundancy is removed, the data scale is reduced, and multiple complete non-redundant medical billing data tables are obtained and transferred to the medical data integration unit.

Medical data integration unit: responsible for data integration, merging and integrating the data of multiple related forms, merging the data of the same billing number and the same patient, using the billing number as the unique key, and obtaining a single data containing all evaluations. The data form of the indicator information and passed into the medical data standardization unit.

Medical data standardization unit: responsible for the standardization of data, standardize the deviation of the data, and standardize the interval to be greater than or equal to 0 and less than or equal to 1 to prevent data weight differences caused by different dimensions. Finally, a standardized medical data table is obtained and passed into the medical data dimension reduction unit.

Medical data dimensionality reduction unit: responsible for the dimensionality reduction of data. Here, principal component analysis is mainly used to obtain the covariance matrix of medical data. For the items with the most information, the dimensionality reduction work is completed, and finally a low-dimensional data form is obtained. Finally, the low-dimensional single medical insurance data form is passed into the random sampling unit.

Random sampling unit: responsible for generating a data subset of the same magnitude from the received low-dimensional single medical insurance data form through random sampling, and finally passing the entire pre-processed multiple data subsets into subsequent medical insurance fraud evaluations The index weight formulating module and the cluster member generating module.

As shown in Figure 3(d), the medical insurance fraud evaluation index weight formulation module mainly completes the adaptive formulation of the weights of various medical insurance fraud evaluation indicators, which is convenient for the subsequent generation of cluster members and the use of cluster fusion.

This module is mainly composed of three functional units, namely, the generation unit of the medical insurance fraud evaluation index weight function, the evaluation index weight function multi-subset particle swarm optimization unit, and the evaluation index weight function global differential evolution unit.

Medical insurance fraud evaluation index weight function generation unit: responsible for receiving standardized low-dimensional medical bill data subsets preprocessed by multiple medical data preprocessing modules, and using each dimension in the above low-dimensional data form as evaluation indicators. Multi-threaded parallel In multiple data subsets, multiple health insurance fraud evaluation index weight functions are generated, which can obtain the best index weight distribution when the minimum value is obtained. In addition, the data subsets are merged into a collection, and based on the above evaluation indicators, the medical insurance fraud evaluation index weight function based on the whole set is generated, and multiple medical insurance fraud evaluation index weight functions generated by the data subset are input into the evaluation index weight function. The multi-subset particle swarm optimization unit inputs the medical insurance fraud evaluation index weight function generated by the complete set into the evaluation index weight function complete set differential evolution unit for optimal solution processing.

Evaluation index weight function multi-subset particle swarm optimization unit: responsible for multi-threaded parallel use of particle swarm optimization algorithm to obtain multiple evaluation indicators according to multiple data subsets after receiving multiple medical insurance fraud evaluation index weight functions generated by data subsets A weight set is obtained, and multiple medical insurance fraud evaluation index weight sets obtained based on the data subset are transmitted to the cluster member generation module to be used as the corresponding weight value of each evaluation index in the weighted clustering algorithm.

The differential evolution unit of the evaluation index weight function ensemble: After receiving the evaluation index weight function generated by the ensemble, the differential evolution algorithm is used to process the evaluation index weight function on the basis of it, and the global maximum value of the index weight function is adaptively obtained. The optimal solution is obtained to obtain the optimal weight of each evaluation index of medical insurance fraud based on the complete set, and transmit it to the cluster fusion module as the weight for calculating the weighted distance during cluster fusion.

As shown in Figure 3(e), the cluster member generation module mainly completes the generation and storage steps of the cluster members of the cluster fusion algorithm, and can also automatically read the cluster members generated by the previous data for cluster fusion use.

The module is mainly composed of 4 functional units, namely coarse clustering unit, agglomerative hierarchical clustering unit, data interaction control unit and cluster member read and write interface.

Rough clustering unit: responsible for receiving multiple data subsets from the medical data preprocessing module and the weights of each evaluation index for multiple data subsets received from the medical insurance fraud evaluation index weight formulating module and set the weights of multiple evaluation indexes Corresponding to the data subset, and then multi-threaded parallel weighted Canopy rough clustering operation to obtain the cluster center and the number of clusters of the cluster, and transmit the rough clustering result information containing the original information to the agglomerative hierarchical clustering unit for processing. At the same time, receive the cluster data that the cluster centers from the subsequent agglomerative hierarchical clustering unit cannot be integrated with the clusters obtained after deep clustering, and perform the weighted Canopy rough clustering operation again to obtain new cluster centers. And the number of clusters is returned to the agglomerative hierarchical clustering unit.

Agglomerative hierarchical clustering unit: responsible for the cluster member generation step in the clustering fusion algorithm, read the original information and the results after rough clustering and the optimal weight value of each index from the rough clustering unit and the index weight formulating module, respectively. Then, on the premise that the cluster centers and the final number of clusters are known, multi-threads use the agglomerative hierarchical clustering algorithm in parallel to obtain multiple cluster members. At the same time, the access of cluster members can be realized by interacting with the read-write interface of cluster members, and the generated cluster members can be stored in the cluster member storage module as historical cluster members for subsequent cluster members that need to be re-clustered and trained. Or the newly imported small amount of data to be verified can be called out for training, improving the accuracy of clustering a small amount of data and reducing the waste of retraining resources. Finally, the cluster members obtained from all data subsets are sent to the cluster. The fusion module performs the subsequent steps of cluster fusion. In addition, if the cluster center from the coarse clustering unit cannot correspond to the cluster in the cluster member, and the cluster member generation fails, the corresponding cluster data and cluster center will be returned to the coarse clustering unit for new Perform the rough clustering operation, and regenerate the cluster members after receiving the cluster center and the number of clusters returned by the rough clustering unit until the generation is successful.

Data interaction control unit: controls the read-write interface of cluster members to send service requests and interaction data to the cluster member storage module, and completes the data interaction forwarding between the data interaction control unit and the cluster member storage module.

Cluster member read and write interface: responsible for the interaction between the agglomerative hierarchical clustering algorithm unit and the cluster member storage module, accept the control instructions of the data interaction control unit, realize the interaction of data between the data interaction control unit and the cluster member storage module, and The data interaction result is fed back to the data interaction control unit for subsequent processing.

As shown in Figure 3(f), the cluster fusion module mainly completes the construction of the consensus function of the cluster fusion algorithm, and then completes the final step of the cluster fusion algorithm to obtain the final data clustering result.

This module is mainly composed of 3 functional units, namely, constructing consensus function unit, data interaction control unit, and data clustering result interface.

Constructing a consensus function unit: responsible for obtaining multiple cluster members generated in the cluster member generating module from the cluster member generating module and obtaining the weights of each evaluation index of medical insurance fraud based on the medical data complete set from the medical insurance fraud evaluation index weight formulating module, and constructing The corresponding consensus function performs weighted fusion of the obtained cluster members, and automatically reprocesses the isolated point data to obtain the final result of cluster fusion of medical form data. It is transmitted to the data clustering result interface for subsequent processing.

Data interaction control unit: controls the data clustering result interface to send service requests and interactive data to the result output module, and completes the data interactive forwarding work between the data clustering result interface and the result output module.

Data clustering result interface: responsible for constructing the interaction between the consensus function unit and the result output module, accepting the control instructions of the data interaction control unit, realizing the interaction of data between the cluster member fusion unit and the result output module, and feeding back the data interaction results to The data interaction control unit performs subsequent processing of outlier screening and display.

The result display terminal includes a module: the medical insurance fraud result output module. As shown in Figure 3(g).

The medical insurance fraud result output module mainly completes the search for isolated points suspected of medical insurance fraud in the data clustering results and the automatic display function of the final result.

This module is mainly composed of 4 functional units, which are service and data interface, data interaction control unit, medical insurance fraud isolated point finding unit and medical insurance fraud suspected regulation display unit.

Service and data interface: responsible for constructing the information interaction between the consensus function module and the result output module, receiving the service request from the data interaction control unit, obtaining the final clustering result generated by the clustering fusion unit, and feeding back the data interaction result to the data interaction control unit for subsequent processing.

Data interaction control unit: responsible for controlling and coordinating other units to jointly complete the data interaction function. The control service and data interface receives the clustering result data from the constructing consensus function module, and forwards the clustering result data to the outlier finding unit, and then controls the outlier finding unit to select outliers in the clustering results as suspected of medical fraud.

Medical insurance fraud isolated point finding unit: responsible for screening out the clustering results as isolated points suspected of medical fraud from the received clustering results, and passing the medical insurance document information contained in the isolated points to the medical insurance fraud suspected regulations display unit for follow-up deal with.

Medical Insurance Fraud Suspected Regulations Display Unit: Receive the medical insurance document information from the isolated point search unit, and display it on the terminal for subsequent manual review and as the basis for accountability of the persons involved.

The cluster member store contains one module: the cluster member store module. As shown in Figure 3(h), the cluster member storage module mainly completes the automatic access function of cluster members, and provides the basis for the extended use provided by the cluster fusion algorithm.

The module is mainly composed of three functional units, namely service and data interface, data interaction control unit, and cluster member database unit.

Service and data interface: responsible for the data interaction of cluster members between the cluster member generation module and the cluster member storage module, accepting service requests from the data interaction control unit, completing the access function of cluster members, and feeding back the data interaction results Perform subsequent processing on the data interaction control unit.

Data interaction control unit: responsible for controlling and coordinating other units to jointly complete the data interaction function. The control service and data interface receives the cluster member data from the cluster member generation module, and can forward multiple clustering result data to the cluster member database unit and control its storage. It can also accept the request of the cluster member generation module from the cluster member. The membership database unit reads the cluster membership data.

Cluster member database unit: responsible for the storage of cluster member data, this unit accepts the control instructions of the data interaction unit, and provides read, write and query functions.

Specifically, as shown in Figure 4(a), the flow of a specific embodiment of the embodiment of the present invention is as follows:

C1, read the medical insurance documents that need to be screened in the medical bill database of the hospital, and organize them into a plurality of medical data forms in a fixed data format for subsequent use in preprocessing.

C2, perform data cleaning, data integration, data standardization and data dimensionality reduction on the read data to obtain a low-dimensional, standardized single medical data form, and obtain multiple medical insurance bill data subsets through random extraction, which is convenient for subsequent processing of data.

C3: Construct multiple index weight functions based on the data subsets according to the obtained preprocessed data subsets, and combine the data into a complete set to construct an index weight function based on the complete data set, and then use the particle swarm optimization algorithm in parallel to make multiple index weight functions based on the data subsets. The index weight function of the data subset obtains the global optimal solution, and then multiple optimally allocated medical insurance fraud evaluation index weight sets based on the data subset are obtained. In addition, the differential evolution algorithm is used to obtain the global optimal index weight function based on the data set. Then, the optimal allocation of medical insurance fraud evaluation index weight sets based on the complete data set is obtained, which are respectively applied to the subsequent weight regulation in cluster member generation and cluster fusion.

C4: First, the weighted Canopy clustering algorithm is used in parallel for the obtained multiple medical insurance bill data subsets and the corresponding medical insurance fraud evaluation index weight sets to perform rough clustering to obtain the number of multiple cluster centers and clusters, and then parallel Using the agglomerative hierarchical clustering algorithm weighted according to the known cluster centers and the number of clusters (the weight of the medical insurance fraud evaluation index), the clustering operation is performed to obtain multiple cluster members. At the same time, the cluster member database is requested, and the cluster members saved in the history are read, which together serve as the basis for the next cluster member fusion. Write the generated cluster members into the cluster member database, and update the cluster members, which is convenient for use when running the scheme next time.

C5, after obtaining the cluster members and the medical insurance fraud evaluation index weight set based on the complete set of medical insurance bill data, construct a consensus function, and fuse the cluster members according to the weighted distance between the cluster center and the edge point to obtain the final medical insurance document Cluster fusion results.

C6, find out isolated points in the clustering fusion results as the adjustment of suspected medical insurance fraud, and display the relevant medical insurance document information on the terminal page for manual review, so as to be used as the basis for accountability of the persons involved and the dependence of subsequent medical insurance fraud protection data.

The whole process does not require human intervention, and can realize intelligent data processing in the whole process and in the whole process.

Among them, as shown in Figure 4(b), the data flow is as follows:

The entire data process can be divided into 6 stages: the collection and sorting stage of medical insurance documents, the preprocessing stage of medical data, the stage of formulating the weight of adaptive medical insurance fraud evaluation indicators, the stage of reading, generating and storing cluster members, and the completion of constructing a consensus function Cluster member fusion stage and finding outliers in the results and taking them as medical insurance fraud suspect regulation stage.

The collection and sorting stage of medical insurance documents: mainly processed by the medical document data collection and sorting module, which reads the medical form data including the expense list, the doctor's order form, and the patient's basic information registration form from the hospital billing database, using a fixed format. Convert the data into multiple form data in a fixed format, and pass the data stream into the medical data preprocessing stage.

Medical data preprocessing stage: It is mainly processed by the medical data preprocessing module. After receiving multiple form data streams in a fixed format passed in from the previous stage, data cleaning, integration, standardization and dimensionality reduction are carried out in sequence, and the data is processed. Convert it into a single standardized low-dimensional medical form data, and convert it into multiple data subsets of the same amount through random sampling, and finally pass the data stream into the formulation stage of the adaptive medical insurance fraud evaluation index weight and read, generate and store. Cluster membership stage.

Adaptive medical insurance fraud evaluation index weight formulation stage: It is mainly processed by the medical insurance fraud indicator weight formulation module, and receives multiple data subsets that are divided into a single standardized low-dimensional medical form data from the previous stage, respectively for multiple data subsets. The data subset and the complete set after the data subset are merged to construct the medical index weight evaluation function, and use the particle swarm optimization algorithm and the differential evolution algorithm to process each subset in parallel to obtain the optimal data subset-based medical insurance fraud. The weight of each evaluation index and the weight of each evaluation index of medical insurance fraud based on the complete data set are respectively passed into the reading, generating and storing stage of cluster members and the construction of consensus function to complete the stage of cluster member fusion.

The stage of reading, generating and storing cluster members: it is mainly processed by the cluster member generation module and the cluster member storage module. The cluster member generation module receives the incoming medical insurance fraud in the development stage of the adaptive medical insurance fraud evaluation index weight. The weights of the evaluation indicators and the multiple data subsets that are entered in the medical data preprocessing stage are divided according to a single standardized low-dimensional medical form data, and the Canopy coarse clustering and agglomerative hierarchical clustering algorithms are sequentially used to parallelize the multiple data subsets. processing, transforming the data into multiple cluster members. At the same time, it interacts with the cluster member storage module to complete the storage of new cluster members and the reading of historical cluster members. Finally, the data stream is passed into the construction consensus function to complete the cluster member fusion stage.

Constructing a consensus function to complete the cluster member fusion stage: it is mainly processed by the cluster fusion module, and the incoming cluster members and the weights of the adaptive medical insurance fraud evaluation index passed in the reading, generating and storing stage of the cluster member are processed. Construct a consensus function fusion strategy based on the weights of each evaluation index of the complete set, perform clustering fusion, and convert the data into the final medical bill data clustering result. And pass the data stream into the finding outliers in the results and use them as the regulation stage of suspected health insurance fraud.

Find outliers in the results and use them as suspected medical insurance fraud regulations Stage: It is mainly processed by the medical insurance fraud result output module. Convert it to isolated point data suspected of medical insurance fraud, and transmit the data stream to the final display terminal for display. Complete the entire data flow.

The present invention adopts the adaptive weight clustering fusion method for the medical bill data of the hospital in the embodiment, which can realize the detection of intelligent medical insurance fraud in the whole link and the whole process. It reduces the situation of human subjectivity intervention, making the final result more accurate.

An embodiment of the present invention provides a medical insurance fraud detection device, as shown in FIG. 5 , including:

The first determination module 501 is configured to determine a plurality of medical bill data to be analyzed, and the medical bill data to be analyzed includes a plurality of evaluation indicators;

A dividing module 502, configured to divide a plurality of medical bill data to be analyzed to obtain a plurality of data subsets;

The second determining module 503 is configured to, for each data subset, determine a subset weight set corresponding to the data subset;

The clustering module 504 is configured to cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset, and obtain the cluster members corresponding to the data subset;

The third determining module 505 is configured to determine the corpus weight set corresponding to the corpus composed of all the medical bill data to be analyzed;

A fusion module 506, configured to fuse all cluster members according to the corpus weight set, wherein all cluster members are composed of cluster members corresponding to multiple data subsets respectively;

The fourth determination module 507 is configured to determine the isolated medical bill data to be analyzed obtained after fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data.

In the embodiment of the present invention, manual marking is not required in the process of detecting medical insurance fraud, which can reduce manual workload. At the same time, the computational efficiency can be improved. Moreover, subjective interference is reduced, and the calculation accuracy can be further improved.

Optionally, the second determination module 503 is specifically configured to construct a first evaluation index weight function according to each to-be-analyzed medical bill data included in the data subset; A subset weight set corresponding to the data subset, wherein the subset weight set includes first weight values corresponding to each evaluation index respectively.

Optionally, the clustering module 504 is specifically configured to, based on each evaluation index, respectively determine the sub-similarity between each pair of medical bill data to be analyzed included in the data subset; and respectively according to each evaluation included in the subset weight set. The first weight value corresponding to the index, weights the sub-similarities corresponding to each evaluation index determined based on each evaluation index, and obtains the total similarity between each pair of medical bill data to be analyzed; Each medical bill data to be analyzed included in the data subset is clustered to obtain cluster members corresponding to the data subset.

Optionally, the third determination module 505 is specifically configured to construct a second evaluation index weight function according to all the medical bill data to be analyzed; according to the second evaluation index weight function, through a differential evolution algorithm, determine the universe weight set corresponding to the universe, Wherein, the full set weight set includes the second weight values corresponding to each evaluation index respectively.

Optionally, the fusion module 506 is specifically used to determine a fusion strategy; based on each evaluation index, determine the similarity between each cluster center pairwise, wherein each cluster center is each cluster in all cluster members. The members correspond respectively; according to the second weight value corresponding to each evaluation index included in the corpus weight set, and the second similarity between each cluster center, all cluster members are fused through a fusion strategy.

Optionally, the first determining module 501 is specifically configured to acquire multiple original medical bill data; for each original medical bill data, preprocess the original medical bill data to obtain a medical bill to be analyzed corresponding to the original medical bill data. data.

Optionally, the device further includes: a display module for displaying suspicious medical bill data.

Optionally, the device further includes: a saving module, configured to save the cluster members corresponding to each data subset respectively.

It should be noted that the medical insurance fraud detection device provided by the embodiment of the present invention is a device applying the above medical insurance fraud detection method, and all the embodiments of the above medical insurance fraud detection method are applicable to the device, and can achieve the same or similar beneficial effects.

An embodiment of the present invention also provides a medical insurance fraud detection device, as shown in FIG. 6 , including a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 pass through the The communication bus 604 accomplishes the mutual communication.

a memory 603 for storing computer programs;

The processor 601 is configured to implement the method steps of the above medical insurance fraud detection method when executing the program stored in the memory 603 .

In the embodiment of the present invention, manual marking is not required in the process of detecting medical insurance fraud, which can reduce manual workload. At the same time, the computational efficiency can be improved. Moreover, subjective interference is reduced, and the calculation accuracy can be further improved.

The communication bus mentioned by the medical insurance fraud detection device may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus or an Extended Industry Standard Architecture (Extended Industry Standard Architecture, EISA) bus or the like. The communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

The communication interface is used for communication between the medical insurance fraud detection device and other devices.

The memory may include random access memory (Random Access Memory, RAM), and may also include non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

The above-mentioned processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; may also be a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

An embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the method steps of the above medical insurance fraud detection method are implemented.

In the embodiment of the present invention, manual marking is not required in the process of detecting medical insurance fraud, which can reduce manual workload. At the same time, the computational efficiency can be improved. Moreover, subjective interference is reduced, and the calculation accuracy can be further improved.

It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article, or device that includes the element.

Each embodiment in this specification is described in a related manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the apparatus, device, and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference may be made to some descriptions of the method embodiments for related parts.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims (10)

1. a medical insurance fraud detection method, is characterized in that, comprises: determining a plurality of medical bill data to be analyzed, the medical bill data to be analyzed including a plurality of evaluation indicators; Divide multiple medical bill data to be analyzed to obtain multiple data subsets; For each data subset, determine the subset weight set corresponding to the data subset; and cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset to obtain the data subset the cluster member corresponding to the set; Determine the corpus weight set corresponding to the corpus composed of all the medical bill data to be analyzed; All cluster members are fused according to the corpus weight set, wherein all the cluster members are composed of cluster members corresponding to each data subset respectively; Determine the isolated medical bill data to be analyzed obtained after the fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data. 2. The method according to claim 1, wherein the determining the subset weight set corresponding to the data subset comprises: constructing a first evaluation index weight function according to each medical bill data to be analyzed included in the data subset; According to the first evaluation index weight function, a particle swarm optimization algorithm is used to determine a subset weight set corresponding to the data subset, wherein the subset weight set includes first weight values corresponding to each evaluation index respectively. 3 . The method according to claim 2 , wherein the medical bill data to be analyzed included in the data subset is clustered according to the subset weight set corresponding to the data subset, and the corresponding data subset is obtained. 4 . The cluster members of , including: Based on each evaluation index, respectively determine the sub-similarity between each pair of medical bill data to be analyzed included in the data subset; Weighting the respective sub-similarities corresponding to each evaluation index determined based on each evaluation index according to the first weight value corresponding to each evaluation index included in the subset weight set, respectively, to obtain each pair of medical bill data to be analyzed. the total similarity between According to the total similarity, each medical bill data to be analyzed included in the data subset is clustered to obtain a cluster member corresponding to the data subset. 4. The method according to claim 1, wherein the determining a corpus weight set corresponding to a corpus composed of all medical bill data to be analyzed comprises: constructing a second evaluation index weight function according to all the medical bill data to be analyzed; According to the second evaluation index weight function, a differential evolution algorithm is used to determine a corpus weight set corresponding to the corpus, wherein the corpus weight set includes a second weight value corresponding to each evaluation index. 5. The method according to claim 1, characterized in that, merging all cluster members according to the corpus weight set, comprising: Determine the fusion strategy; Based on each evaluation index, the similarity between each cluster center is determined respectively, wherein each cluster center is corresponding to each cluster member in all cluster members; According to the second weight value corresponding to each evaluation index included in the corpus weight set, and the second similarity between each cluster center, all cluster members are fused through the fusion strategy. 6. The method according to any one of claims 1 to 5, wherein the determining a plurality of medical bill data to be analyzed comprises: Obtain multiple raw medical billing data; For each original medical bill data, the original medical bill data is preprocessed to obtain the to-be-analyzed medical bill data corresponding to the original medical bill data. 7. The method according to any one of claims 1 to 5, wherein the isolated medical bill data to be analyzed obtained after fusion is determined, and the isolated medical bill data to be analyzed is regarded as suspicious medical bills After the data, the method further includes: The suspicious medical billing data is displayed. 8. The method according to any one of claims 1 to 5, wherein, for each data subset, a subset weight set corresponding to the data subset is determined; and according to the subset weight set corresponding to the data subset The set weight set performs clustering on the medical bill data to be analyzed included in the data subset, and after obtaining the cluster members corresponding to the data subset, the method further includes: Save the cluster members corresponding to each data subset. 9. A medical insurance fraud detection device, characterized in that, comprising: a first determining module, configured to determine a plurality of medical bill data to be analyzed, wherein the medical bill data to be analyzed includes a plurality of evaluation indicators; A division module, which is used to divide a plurality of medical bill data to be analyzed to obtain a plurality of data subsets; a second determination module, configured to determine, for each data subset, a subset weight set corresponding to the data subset; a clustering module, configured to cluster the medical bill data to be analyzed included in the data subset according to the subset weight set corresponding to the data subset, and obtain the cluster members corresponding to the data subset; The third determining module is used to determine the corpus weight set corresponding to the corpus composed of all the medical bill data to be analyzed; a fusion module, configured to fuse all cluster members according to the corpus weight set, wherein all the cluster members are composed of cluster members corresponding to a plurality of data subsets respectively; The fourth determination module is configured to determine the isolated medical bill data to be analyzed obtained after fusion, and use the isolated medical bill data to be analyzed as suspicious medical bill data. 10. The apparatus according to claim 9, wherein the second determination module is specifically configured to construct a first evaluation index weight function according to each medical bill data to be analyzed included in the data subset; The evaluation index weight function determines the subset weight set corresponding to the data subset through the particle swarm optimization algorithm, wherein the subset weight set includes the first weight values corresponding to each evaluation index respectively.