Improving Accuracy of Fraud Detection Models in Health Insurance Claims Using Deep Learning / AI

Abstract

This paper proposes a novel approach of deep learning towards the improvement of accuracy in health insurance fraud detection. To capture complicated patterns in claims data, we consider a hybrid architecture which consists of BiLSTM networks with attention mechanisms and graph neural networks (GNNs). Our model achieves an accuracy of 94.2% on a large dataset of 1.2 million health insurance claims, which is a 15% increase in terms of accuracy over traditional machine learning methods. In addition, the attention visualization and feature importance allow for interpretation of the more complex architecture.