Explainable AI for Institutional Fraud Decisions: A Cross-Sector Empirical Study Using Public Healthcare and Financial Transaction Data

Abstract

Institutional fraud detection operates at the intersection of predictive analytics, regulatory accountability, and human judgment. While machine learning models can identify fraudulent behavior, their organizational value depends on their ability to support institutional decision-making through interpretable and stable explanations. In regulated environments, analytical outputs must be justified and integrated into investigative workflows. This study empirically examines explainable machine learning as a decision support mechanism across two institutional domains: public healthcare payment systems and financial transaction systems. Using the CMS-derived Healthcare Provider Fraud Detection dataset and the UCI Credit Card Fraud dataset, logistic regression, random forest, and gradient boosting models were developed and evaluated under realistic class imbalance conditions. Explainability was assessed using SHAP-based feature attribution to examine explanation stability and institutional interpretability. Results show that healthcare fraud models produce stable and institutionally meaningful explanations aligned with billing behavior, while financial fraud models generate accurate predictions but less stable explanations. These findings indicate that explainability is shaped by institutional data structure rather than model architecture alone. The study contributes to business analytics research by demonstrating how explanation stability influences decision relevance in institutional fraud detection.