Hasan, Asif and Nasir, Abu and Sultana, Nasrin (2025) AI-Driven Eligibility Verification in U.S. Medicaid: A T-MSIS V4-Aligned Real-Time Data Matching Framework for Reducing Improper Payments and Protecting Public Program Integrity. International Journal of Computational and Experimental Science and Engineering, 11 (1). ISSN 2149-9144

Text final+5495.pdf Download (862kB)

Abstract

Background: Medicaid experiences billions of dollars in annual improper payments arising from eligibility determination errors, inaccurate beneficiary information, and fraudulent enrollment. Fragmented state eligibility systems and sequential manual verification hinder timely detection. The CMS Transformed Medicaid Statistical Information System Version 4 (T-MSIS V4) provides standardized national data infrastructure, but current implementations support reporting rather than intelligent verification. Objective: This study proposes and empirically evaluates MedVerify-AI, an artificial intelligence-driven eligibility verification framework aligned with T-MSIS V4 that combines multi-source probabilistic record linkage with AI-based anomaly scoring for real-time detection of improper enrollments. Methods: The framework links applications to external verification sources (identity, income, death, residency records) using Fellegi–Sunter probabilistic record linkage, computes verification-layer scores across identity, financial, status, and duplicate dimensions, and fuses evidence with a supervised meta-learner. Evaluation used a reproducible synthetic dataset of 61,096 Medicaid applications over 24 months containing 5,902 anomalous applications (9.7%) across five documented anomaly types, with a temporal train/test split and four baselines: rule-based verification, logistic regression, isolation forest, and random forest. Results: Probabilistic linkage achieved F1 of 0.998, recovering most identity links missed by deterministic matching (recall 0.996 vs. 0.988) at 1.18 ms per record. The proposed framework achieved the highest F1-score (0.803) and ROC-AUC (0.884) of all methods, improving recall by 13.5 percentage points over rule-based verification (0.727 vs. 0.593) at a 0.91% false-positive rate. Deceased-enrollee, income-misreporting, and residency anomalies were detected with complete recall. Conclusions: Combining standardized T-MSIS data structures, probabilistic linkage, and AI-based evidence fusion materially outperforms rule-based verification for pre-enrollment screening. Validation on operational Medicaid data is the essential next step.

Item Type:	Article
Uncontrolled Keywords:	Medicaid, T-MSIS, eligibility verification, artificial intelligence, record linkage, program integrity
Subjects:	H Social Sciences
Depositing User:	admin eprints
Date Deposited:	20 Oct 2026 08:02
Last Modified:	20 Oct 2026 08:02
URI:	http://eprints.umsida.ac.id/id/eprint/17222

Actions (login required)

View Item