CHIMERA: A Multi-Agent Graph-Temporal Framework for Interpretable Medical Insurance Fraud Detection

Medical insurance fraud imposes significant financial burdens globally, yet traditional detection methods struggle with the complexity of heterogeneous claims data and lack interpretability. We propose CHIMERA, a Collaborative Hierarchical Intelligent Multi-Expert Reasoning Architecture. This framework integrates a Large Language Model backbone with a Graph-Augmented Temporal Transformer (GATT) to capture relational and sequential dynamics. A Cross-Modal Semantic Fusion (CMSF) module aligns structured features with semantic embeddings, while an uncertainty-aware ensemble combines unsupervised detectors with domain rules. Extensive experiments on CMS synthetic data demonstrate that CHIMERA achieves an AUC-PR of 0.782, significantly outperforming state-of-the-art baselines. Furthermore, the multi-agent system provides investigator-ready explanations, bridging the gap between detection accuracy and operational transparency.