Models for Detecting Frauds in Medical Insurance

Abstract

Health insurance is important for many people, but unfortunately it is susceptible to frauds, therefore expenditures for covering the funds show exponential growth. The victims of this kind of scams are not only the institutions that provide the funds and treatments, but also are the ones who really need that help, except they have lost their priority due to a committed fraud. In order to rationally provide funds and minimize losses, there is a need for fraud detection systems. In this paper, this issue is considered as a binary classification problem, using data inherent in the nature of the field. The whole data science pipeline process is considered in order to elaborate our results that are higher than the published ones on the same problem: 0.95, 0.96 and 0.98 AUC scores with different models. The data is integrated from three interconnected databases, which are pre-processed and then their cross-section is undertaken. The dataset is unbalanced concerning the records of both classes, therefore certain balancing techniques are applied. Several models are built using traditional Machine Learning models, classifiers with Deep Neural Networks and ensemble algorithms and their performance is validated according to several evaluation metrics.