Research On Abnormal Co-occurrence Medical Visit Fraud Behaviors Identification Method Based On Bi-clustering

China has entered the era of universal medical insurance,with the coverage of the medical insurance system expanding and the number of beneficiaries increasing.This brings convenience to people's medical treatment,people can bring medicare card to medical treatment,instant settlement.However,there are also some criminals who engage in health insurance fraud.There are many kinds of medical insurance frauds,among which the abnormal co-occurrence medical fraud visit behavior is a common insurance fraud,which is usually manifested as the fraudsters illegally use the medical insurance cards of multiple others and illegally get the medical insurance fund by reselling drugs for many times.This fraud has two characteristics:one is that the fraudsters occupy the health insurance card,many times at the same time in the same place;second is that fraudsters usually buy similar drugs to facilitate drug reselling Generally speaking,the likelihood of a health care provider's involvement,or the openness of health care policies,makes such behavior often similar to that of a regular insured person,making it difficult to identify.However,abnormal co-occurrence medical fraud visit behavior has brought more and more huge losses to the medical insurance fund.For example,in 2017,drug dealers in Ningbo city colluding with dozens of retired insured workers to resell drugs and damage the medical insurance fund by more than 900,000 yuan,so the identification of fraudulent behavior of abnormal co-occurrence of medical treatment becomes increasingly urgent.Currently,there are some fraud identification methods,such as clustering,frequent pattern mining.These methods usually fail to fully take into account the two characteristics of abnormal co-occurrence medical fraud visit behavior,misjudge the normal medical treatment behavior,resulting in the failure to accurately identify the fraudulent behavior.Therefore,how to accurately identify abnormal co-occurrence medical fraud visit behavior is a challenge.In order to identify abnormal co-occurrence medical fraud visit behavior as accurately as possible,without missing detection or misjudgment,the excavated patient population must simultaneously conform to the two characteristics of the fraud,that is,frequent medical treatment at the same time and at the same place,and purchase similar drugs.Clustering is a commonly used method for data analysis Outliers are obtained by clustering as outliers to identify fraud.Previous clustering methods usually carry out global clustering based on the characteristic dimension of data objects.For example,patients can generate one-hot vector representation according to the drugs they bought,and the patients clustered in a cluster have similar medical prescriptions.The bi-clustering method can cluster data objects and feature dimensions at the same time.For example,the rows represent patients and the columns represent drugs in the matrix.Each element in the matrix represents the number of drugs purchased by patients.The rows of these sub-matrices correspond to patients with similar prescriptions whose prescriptions contain drugs corresponding to the columns of the sub-matrix.Similarly,the bi-clustering method can mine the patient groups that frequently seek medical treatment at the same time and place,as long as the matrix of rows representing patients and columns representing the dimension of time and place of medical treatment is constructed.Therefore,two kinds of fraud detection methods based on bi-clustering,single-view bi-clustering fraud detection method and multi-view bi-clustering fraud detection method are proposed Single-view bi-clustering fraud detection method(Biclustering-sim),first built in patients with medical treatment time place dimension matrix,by bi-clustering method for mining frequent same time same place for medical treatment of patients with suspected groups and their suspicious of medical records,and then calculated according to medical records of the suspected patients with suspected the medicine prescription similarity between groups,the patients with other medicine prescription is not similar in patients with normal filtering,eventually get fraud patients more accurately.However,there are shortcomings in this method.The bi-clustering may gather fraudsters from different groups or normal patients together,and the suspicious medical records excavated are not the medical records of fraudsters,which may affect the calculation of similarity of subsequent medical prescriptions and may omit fraudsters.Multi-view bi-clustering fraud detection method(Multi-view biclustering-sim)is an extension of single-view bi-clustering fraud detection method,which makes up for the shortcomings of single-view bi-clustering fraud detection method.By constructing patient-visit time and place dimensional matrix view and patient-drug matrix view,it conducts a bi-clustering in these two views to obtain a cross-view consistent patient group,that is,a fraudulent patient group with frequent visits to the same place at the same time and similar medical prescriptionsMedical insurance data from Laiwu city in Shandong province were used,including more than 7,000 patients and 190,000 medical records.According to the real abnormal co-occurrence medical fraud visit behavior cases,the medical treatment records conforming to the abnormal co-occurrence medical fraud visit behavior were synthesized and inserted into the real data,and a number of synthetic data were simulated to evaluate the effectiveness of the two fraud detection methods.In order to make the synthesized data more authentic,part of the fraudulent data will be deleted randomly according to a certain proportion for each fraud patient.In the synthetic data experiment,compared with the three comparison methods,the single-view bi-clustering fraud detection method filters out the normal patients who frequently seek medical treatment at the same time and at the same place,which can identify the fraudsters more accurately and reduce the misjudgment rate.The multi-view bi-clustering fraud detection method is better than the single-view bi-clustering fraud detection method,because the former makes up the deficiency of the latter.Finally,the method in this paper was applied to real medicare data and four suspected patient populations were mined.