Reliability‑Aware Multimodal Intelligence for Healthcare Insurance Fraud Detection: A Trimodal Data Perspective

Abstract

Healthcare insurance fraud has emerged as a significant challenge due to the rapid growth of digital healthcare systems, increasing volumes of electronic insurance claims, and the evolving complexity of fraudulent activities. Conventional fraud detection approaches primarily rely on manually defined rules, single-modal data, or traditional machine learning algorithms, which often struggle to capture complex relationships among heterogeneous healthcare information and provide reliable predictions. To address these limitations, this study proposes a Reliability-Aware Multimodal Intelligence Network (RAMI-Net) for healthcare insurance fraud detection from a trimodal data perspective. Initially, healthcare insurance claims are preprocessed and transformed into three complementary perspectives: patient structured data, claim-related textual information, and healthcare relationship graphs. These perspectives are processed using TabNet, ClinicalBERT, and Graph Attention Network (GAT) to extract modality-specific feature representations. A Reliability Estimation Module evaluates the confidence of each modality before a Reliability-Aware Cross-Modal Attention Fusion mechanism integrates reliable features while minimizing noisy information. The fused representations are classified using an Explainable Fraud Detection Network, and SHAP provides transparent interpretations by identifying influential fraud-related features. Experimental results demonstrate that the proposed framework achieves 99.95% accuracy, 99.98% precision, 99.97% recall, 99.95% F1-score, and 99.63% ROC-AUC, outperforming conventional machine learning and deep learning approaches. The proposed RAMI-Net provides an accurate, robust, reliable, and interpretable solution for intelligent healthcare insurance fraud detection, supporting effective real-world healthcare insurance claim management.