Comparative Analysis of Unsupervised Learning Approaches for Anomalous Billing Pattern Detection in Healthcare Payment Integrity

Abstract

Healthcare payment integrity faces substantial challenges from anomalous billing patterns that undermine financial sustainability and compromise resource allocation effectiveness. This research develops a systematic comparative framework evaluating five unsupervised learning algorithms—Isolation Forest, Local Outlier Factor, DBSCAN, One-Class SVM, and Autoencoder—for detecting aberrant billing behaviors within medical claims databases. Through empirical analysis of Medicare Part B data spanning 142,738 provider records, we quantify detection accuracy, computational efficiency, and pattern recognition capabilities across distinct algorithmic approaches. Isolation Forest demonstrates superior performance with 0.847 F1-score and 3.2-second processing time per 10,000 claims, while Autoencoders reveal 23.6% higher sensitivity to complex multivariate anomalies. The analysis identifies critical tradeoffs between precision-recall balance and scalability constraints, establishing quantitative benchmarks for algorithm selection in operational fraud detection systems. Our findings indicate that ensemble configurations combining density-based and reconstruction-error methodologies yield 15.8% improvement over single-algorithm deployments.