AI-Driven Healthcare Fraud Detection in Emerging and Advanced Markets: A Comparative Framework for Adapting ML Models to U.S. Medicare and Medicaid Compliance Environments

Abstract

Background: Healthcare fraud imposes substantial financial burdens worldwide, but fraud patterns and detection infrastructure differ markedly between emerging healthcare markets — fragmented data, partial coding standardization, evolving regulation — and advanced markets such as U.S. Medicare and Medicaid, which operate under standardized ICD-10/CPT/HCPCS coding and continuous CMS/HIPAA compliance monitoring. Machine learning fraud detectors trained in one environment are not guaranteed to transfer to the other.

Objective: This study develops and empirically tests a comparative AI-driven framework for adapting machine learning-based fraud detection models from an emerging-market claims environment to a U.S. Medicare/Medicaid-style advanced-market environment, testing whether regulatory mapping and feature harmonization improve cross-market transfer relative to naive, unadapted transfer.

Methods: Two reproducible synthetic claims environments were constructed — an Emerging Market (heterogeneous coding, frequency-driven fraud) and an Advanced Market (standardized coding, amount- and mismatch-driven fraud) — with distinct fraud-rate and fraud-mechanism distributions. Four models (XGBoost, LightGBM, Random Forest, Isolation Forest) were evaluated across three scenarios on a common Advanced-Market holdout: (A) native Advanced-Market training, (B) naive cross-market transfer, and (C) the proposed harmonized adaptation using peer-provider deviation scoring and regulatory-maturity features.

Results: Naive transfer (B) retained most discriminative capacity relative to the native upper bound (A) (ROC-AUC within 0.001–0.024) but at markedly lower precision. Harmonized adaptation (C) improved F1-score for XGBoost (0.539→0.598) and LightGBM (0.548→0.579) relative to naive transfer, at a modest cost to ROC-AUC and PR-AUC; Random Forest and Isolation Forest showed a less favorable trade-off. SHAP analysis showed peer-provider claim-timing deviation, provider billing-intensity percentile, and diagnosis–procedure compatibility as dominant predictors.

Conclusions: Regulatory mapping and peer-group feature harmonization measurably improve cross-market transfer for gradient-boosting models on decision-relevant metrics, though not uniformly across all models, indicating cross-market AI adaptation is model-dependent rather than universally beneficial. Validation on real cross-market claims data is the essential next step.