Enhancing Fraud Detection in Health Insurance: Deep Neural Network Approaches and Performance Analysis

Abstract

This study develops and examines a comprehensive deep learning framework for the detection of multi-class healthcare fraud in National Health Insurance Scheme (NHIS) claims. We examined 20,388 NHIS healthcare claims revealing four specific fraud patterns: Phantom Billing, Wrong Diagnosis, Ghost Enrollee, and legitimate claims. Four different deep neural network architectures were developed and evaluated: Simple NN, Deep Wide NN, Regularized NN, and Residual NN, in addition to ensemble methods. The Simple Neural Network achieved the highest overall performance, with a test accuracy of 79.84% and an F1-macro score of 77.76%. Despite possessing only 100,324 parameters (five times fewer than the Wide Deep Neural Network), it outperformed more complex designs while achieving the fastest training time of 40.61 seconds. Multiclass analysis demonstrated exceptional performance in Ghost Enrollee detection (97.84% F1-score) and moderate performance in Phantom Billing detection (61.15% F1-score).