Research On Health Insurance Fraud Detection Based On Dynamic Heterogeneous Information Network

With the development of medical insurance,the problem of medical insurance fraud is emerging.Every year,a large amount of health insurance funds are cheated.In the process of health insurance business,a large amount of health insurance related data have been accumulated.The health insurance data contains information about the patients themselves,information about going to the hospital for treatment,and insurance participation payment.Detecting abnormal data that may be fraudulent is an important means to control the phenomenon of health insurance fraud.The existing medical insurance fraud detection methods generally determine the fraud related patterns and indicators through prior knowledge,and then use outlier detection to detect anomalies based on these fixed patterns and indicators.However,the fraud pattern is concealed and changeable.This method that relies on prior knowledge often lags behind the development of fraud models,and cannot detect fraud records in a new fraud model in a timely manner.At the same time,the health insurance data is accumulated over the years.The data are time series,mainly reflected in the timeliness.The climate,policy,medical level and economic development of different years all have some influence on the data of health insurance,such as sudden changes of various indicators caused by breakthrough of treatment methods,and the incidence rate of some diseases in different seasons.Therefore,there is sometimes a lack of comparability between data in different time periods,but many researches have not paid attention to this feature.Therefore,the traditional methods of mining anomalies from fixed patterns and other methods that ignore timing characteristics are difficult to meet current needs.In view of the above problems,this paper mainly makes the following contributions:1.In this paper,a model of health insurance fraud detection based on dynamic heterogeneous information network(HIN)is proposed.Using the rich expression ability of HIN,the complex relationship between entities and entities involved in the field of health insurance is modeled,and the health insurance business representation model is established.Based on the health insurance business representation model,all possible business patterns,interrelated business combination patterns and related indicators in the health insurance field are mined in an unsupervised way.The dynamic HIN is constructed according to the time of data occurrence,and anomaly detection is carried out from both horizontal and vertical perspectives.Horizontal detection is in the same time period,to find out the data with abnormal external environment compared with other records;the vertical detection spans multiple time periods,to find out the data with great changes over time compared with itself.2.In this paper,we propose an optimization method of fraud detection based on the law of disease evolution.Health insurance fraud detection model does not take into account the normal change of some data of the disease over time in the longitudinal detection,we learn the law of disease evolution to optimize it.We propose the corresponding model of disease evolution.First,the graph convolution network is used to process the data of a single time,embed the neighbor information in the network,and add Attention to weight different neighbor nodes.Then,a self auto-encoder is constructed to reduce the dimension of time series data and cluster it.At last,we put forward an index to measure the probability of normal evolution of each cluster,and learn the law of disease development from the normal evolution of clusters.Adjust the longitudinal inspection to the data whose change does not conform to the normal evolution law.In this paper,the above models are tested on real datasets to evaluate the effect.The experiments show that the models proposed in this paper have good performance.Through the above research,we can narrow the scope of health insurance records suspected of fraud,improve the efficiency of detection,find some new fraud patterns in time,and reduce the corresponding fraud losses.