Research And Implementation Of Medicare Fraud Detection Model Based On Outlier Analysis Technology

With the increasing popularity of medical insurance and the continuous improvement of medical informatization,Medicare Fraud has become a social problem.The usual fraud behaviors,such as pharmaceutical companies and medical consumables manufacturers dump drugs and medical consumables,medical institutions forge patient records,doctors use medications without indications,patients misappropriate health insurance cards,etc.These behaviors not only damage the interests of ordinary insureds,but also seriously hinder the implementation and promotion of health insurance policies.Facing with the actual problems of medical insurance business in Weihai and Yantai City,this thesis proposes a medicare data representation and feature extraction method.Based on the data representation,this thesis proposes a medicare fraud detection model.Experiments show that the proposed model and method has a good experimental performance.The main tasks of this thesis are:(1)This thesis proposes a medicare data representation and feature extraction method.The patient is used as the evaluation object.For each patient,the extraction method converts the medical records to a vector,each element represents the usage of a medical item.Considering the correlation of original features,the dimension reduction methods,such as PCA,KPCA and Autoencoder,are used as feature extraction methods.Experiments show that the approach proposed in this thesis,which is applied to the medicare fraud detection task,has a good performance.(2)A hybrid medicare fraud detection model is proposed.This thesis proposes a hybrid model which combines the supervised and unsupervised methods.Unsupervised method is used for outlier analysis of unlabeled data.With the manual annotation for the outlier,label is produced.Therefore,supervised method can be trained with the labeled data.In this way,the backlog of medicare data can be reduced,and the work efficiency of manual auditing and machine auditing can be improved.(3)This thesis designs and implements the proposed model.Each logical unit of the model is encapsulated to achieve its asynchronous call.Users can control the system and manually review the abnormal data through the web page.