AI-Driven Fraud, Waste, and Abuse Detection in Medicaid Claims Using Graph Neural Networks

Abstract

Medicaid is one of the largest public health insurance programs in the United States, processing millions of claims annually and managing billions
of dollars in healthcare expenditures. The scale and complexity of its multi-entity billing ecosystem make it highly susceptible to fraud, waste,
and abuse, including phantom billing, upcoding, identity misuse, and coordinated provider networks. Traditional detection approaches rely on
rule-based audits and tabular machine learning models that analyze claims as independent records. These methods often fail to capture hidden
relational patterns and interconnected fraud schemes embedded within provider–patient–facility networks.
This study proposes an AI-driven fraud detection framework using Graph Neural Networks to model Medicaid claims as a heterogeneous graph
structure. Providers, beneficiaries, claims, and facilities are represented as nodes, while billing interactions, referrals, and shared identifiers form
edges. The model leverages graph-based message passing to capture relational dependencies and detect coordinated fraud behavior. To address
class imbalance, imbalance-aware learning strategies are incorporated during training. Baseline comparisons include conventional machine
learning classifiers and network embedding techniques.
Experimental results demonstrate significant improvements in precision, recall, F1-score, and ROC-AUC compared to traditional approaches.
The proposed framework also integrates explainability mechanisms to identify influential subgraphs and high-risk billing patterns, supporting
transparency in fraud investigations. Overall, this research presents a scalable, interpretable, and policy-relevant solution for enhancing Medicaid
program integrity.