Secure and explainable fraud detection in healthcare claims using blockchain-based machine learning

Healthcare insurance fraud causes substantial financial losses, operational inefficiencies, and reduced trust among patients, providers, and insurers. Conventional fraud detection approaches often rely on centralized infrastructures and opaque machine learning models, limiting transparency, adaptability, and data integrity. To address these challenges, this paper proposes a secure and explainable healthcare fraud detection framework that integrates blockchain technology with machine learning. A permissioned blockchain provides decentralized, tamper-resistant storage and an immutable audit trail for claim-related evidence, while a stacking ensemble architecture based on LightGBM and XGBoost identifies suspicious provider behavior from healthcare claims data. Experiments conducted on a real-world dataset comprising more than 500,000 outpatient claims, 40,000 inpatient claims, and 138,000 beneficiary records demonstrate effective performance under severe class imbalance. Given the imbalanced nature of fraud detection, Average Precision (AP) was adopted as the primary evaluation metric. Among the evaluated models, the Voting classifier achieved the highest AP score of 0.746, indicating the strongest precision-recall balance for identifying fraudulent providers. The proposed stacking framework achieved competitive overall performance with an accuracy of 0.940, AUC of 0.951, AP of 0.733, precision of 0.702, recall of 0.630, and F1-score of 0.664. Model predictions are further interpreted using SHAP (SHapley Additive exPlanations), providing stable feature-level explanations across multiple validation runs. The blockchain layer securely anchors hashed prediction artifacts and explanation records with minimal latency, enabling verifiable and tamper-resistant auditing. These findings demonstrate the potential of integrating explainable machine learning with blockchain-based auditability to support trustworthy healthcare fraud detection.