Published Paper Details:

Health insurance fraud results in significant financial losses and necessitates scalable, intelligent detection methods. This work presents a modular and unsupervised framework for fraud detection in healthcare claims by integrating semantic feature extraction, heterogeneous graph modeling, and anomaly detection. Structured claim metadata and unstructured clinical narratives are embedded using transformer-based models to capture rich contextual information. These embeddings are incorporated into a heterogeneous graph that connects patients, providers, and claims, enabling the use of graph neural networks to learn complex relational patterns indicative of fraudulent behavior. A graph-based representation is constructed using synthetic yet realistic healthcare data generated in standard clinical formats. Contextual node embeddings are learned, and clustering methods are applied to identify latent behavioral patterns. Unsupervised anomaly detection techniques, including tree-based and distance-based models, are employed to flag suspicious entities. A provider-level risk scoring mechanism is introduced to prioritize investigation efforts. This framework is designed to operate without reliance on labeled data, ensuring adaptability to evolving fraud strategies and generalizability across diverse claim scenarios. Experimental evaluation shows the system's effectiveness in uncovering subtle fraud signatures, highlighting its potential as a robust alternative to traditional rule-based approaches.