Outlier Detection Based Medicare Anomalous Data Mining

With the development of medical insurance in China,the coverage and scales of medical insurance systems are increasing,enlarging the effects on people's lives.However,while the expenditure of medicare fund increases due to the aging of population,medicare fraud activities(e.g.,excessive medical treatment,decomposing hospitalization,illegal prescribing,etc.)occur more frequently.Facing this problem,traditional medicare fraud detection methods usually rely on rules which are based on the experience of domain experts.The shortcoming is obvious: it is difficult to ensure the completeness of rules and lacks flexibility.Outlier detection aims to find the data instances that are far away from the majority.This technique has been widely applied in many real-world domains,such as credit card fraud detection,network intrusion detection,abnormal medical case detection,etc.In the medicare field,fraud medicare records tends to be outliers.This paper proposes several anomalous medicare data mining methods based on outlier detection.The main contributions of this paper are as follows:· Context based abnormal treatment and medication detection.The context of a patient,such as the state of illness or the grade of the hospital,has a large effect on the treatment and medication.Traditional rule-based methods have difficulty in dealing with the variety of contexts.A data-driven method is proposed,using deep neural networks to learn the embedding of contexts.Next,it employs Mahalanobis metric learning to detect outliers by comparing an instance with instances of similar contexts;· Density and local outlier based hospitalization fee anomaly detection.Local outlier factor detects outliers by the relative density of an instance.This paper theoretically analyses its limitation: this method can not find out self-clustered outliers.To handle this issue,a density based clustering algorithm(i.e.,DBSCAN)is used to adjust anomaly scores of data instances;· Medicare anomalous data analysis and demonstration system.Based on the research above,a visualization system is designed and implemented,consisting of an offline module and an online module.The functions of the offline module are data storage,cleaning and management.The online module displays the outputs of models,showing why an instance is anomalous from the perspectives of probability and statistics.