Soumik, MD Shadman and Babu, Rasel Hossain and Sarkar, Mrinmoy (2026) GovFraudNet: A Novel AI-Powered Multi-Source Fraud Intelligence Fusion Model for Cross-Agency Medicaid Fraud and Public Benefits Misuse Detection in the USA Using Machine Learning, Cybersecurity, and Advanced Data Analytics. American Journal of Current Tendency and Innovation, 3 (4). ISSN 2996-945X

Text AJCTI_GovFraudNet+A+Novel+AI-Powered.pdf Download (1MB)

Abstract

Medicaid and other public benefit programs in the United States lose tens of billions of dollars each year to fraud, waste, and misuse, and the fragmented structure of the agencies that administer them allows sophisticated schemes to hide in the gaps between data silos. This article proposes GovFraudNet, a multi-source fraud intelligence fusion model that combines four heterogeneous evidence streams, namely Medicaid claims, enrollment and eligibility records, provider network relationships, and cybersecurity telemetry, into a single calibrated risk score that can be shared across agencies. The model trains a specialist gradient-boosted expert on each source, converts each expert's output to log-odds evidence, and fuses the evidence with a second-stage meta-learner that also receives cross-source interaction features designed to expose coordinated fraud rings. Because real cross-agency data are not publicly available at the record level, the framework was evaluated on a controlled simulation of 60,000 beneficiary-provider records with a 3.5% fraud prevalence, in which each source reveals only part of the fraud. On a held-out test set of 18,000 records, GovFraudNet reached a ROC-AUC of 0.821, compared with 0.802 for a random forest on concatenated features and 0.717 for the best single source. Under a fixed investigative cost model, it produced the highest peak net recovery, about 1.71 million U.S. dollars in the test sample, which is 46.5% more than a claims-only system. The article also presents the mathematical formulation, an ablation study, a cost-benefit calculation, and a governance and cybersecurity framework for privacy-preserving deployment. The results are simulation-based and should be interpreted as a proof of concept rather than evidence of production performance.

Item Type:	Article
Uncontrolled Keywords:	Medicaid fraud, public benefits integrity, multi-source data fusion, machine learning, anomaly detection, cybersecurity analytics, graph learning
Subjects:	H Social Sciences
Depositing User:	admin eprints
Date Deposited:	27 Oct 2026 09:31
Last Modified:	27 Oct 2026 09:31
URI:	http://eprints.umsida.ac.id/id/eprint/17284

Actions (login required)

View Item