A Deep Learning Approach for Healthcare Insurance Fraud Detection

Healthcare fraud is a global financial challenge affecting economic stability and trust in services, with traditional machine learning models struggling to accurately capture its complexity and adaptive nature. This study investigates the application of three deep learning (DL) models, which are artificial neural networks (ANN), convolutional neural networks (CNN) and long-short-term memory networks (LSTM) for healthcare fraud detection. This study used healthcare claim data, including patient demographics, claim amounts, diagnostic codes, and procedure types, to analyse healthcare service usage and identify fraudulent activity. To enhance the interpretability of these models, locally interpretable model-agnostic explanations (LIME) were used. The evaluation results demonstrated that the ANN was the best performer with an accuracy of 0.94, precision of 0.78, recall of 0.45, and F1-score of 0.57. While CNN excelled in accuracy, the LSTM was more effective in reducing false negatives. The LIME for ANN shows the prediction of a claim to be non-fraudulent with a high probability of 0.96, as opposed to a 0.03 probability of being fraudulent with ‘PotentialFraud', as a driving feature, the evaluation metrics show that it is good at correctly identifying fraudulent cases. This study highlights the efficacy of integrating deep learning models with explainable AI (XAI), contributing to the growing research body in healthcare insurance fraud detection.