Advanced Neural Networks for Detecting Medical Billing Anomalies and Predictive Fraud Prevention

Abstract

Healthcare fraud, particularly through aberrant medical billing, imposes substantial financial burdens on public health

systems and compromises patient trust. Traditional rule-based audit systems fail to capture sophisticated, evolving

anomalous patterns. This study addresses the gap in dynamic, scalable fraud detection by proposing a deep learning

framework combining autoencoders and long short-term memory (LSTM) networks. The primary purpose is to

develop and validate an advanced neural network model capable of identifying billing anomalies in real time and

predicting fraudulent claims before payment disbursement. Using a quantitative research design, we analyzed a

synthetic dataset mimicking 500,000 Medicare claims, incorporating features such as procedure codes, patient

demographics, and provider behavior patterns. The methodology involved training a hybrid autoencoder-LSTM

model for unsupervised anomaly detection and supervised classification. Key findings indicate that the proposed

model achieves a 96.4% fraud detection recall and an area under the curve (AUC) of 0.98, significantly

outperforming logistic regression and random forest baselines. The conclusion underscores that integrating predictive

business analytics with neural networks can reduce improper payments by an estimated 32%, offering a scalable

solution for US public health systems and private payers.