AI-Driven Modernization of Medicare and Medicaid Enterprise Systems: Interoperability, Claims Analytics, and Fraud Detection Frameworks

Abstract

The Centers for Medicare & Medicaid Services (CMS) is responsible for the provision of healthcare coverage to more than 150 million beneficiaries; however, the enterprise systems of CMS suffer from various issues, namely data interoperability, inefficiency, and fraud, waste, and abuse. Rule-based mechanisms have proven to be inadequate for fraud prevention in the changing environment, whereas the fragmentation of datasets of CMS limits the effectiveness of any analysis. Therefore, this paper aims to propose an integrated AI-based framework for healthcare analytics, which includes interoperability, predictive analytics, anomaly detection, and explainable artificial intelligence for large-scale fraud detection. In particular, the suggested framework includes a novel FHIR-like interoperability module that would allow to align heterogeneous datasets within CMS into a single patient-provider-focused data lake. A hybrid approach to the fraud detection algorithm implementation is introduced based on supervised machine learning methods (Logistic Regression, Random Forest, XGBoost, LightGBM, CatBoost) with the use of imbalance aware training. Moreover, the framework incorporates unsupervised anomaly detection algorithms (Isolation Forest, Local Outlier Factor) and graph-based network analysis of providers to detect relational fraud. Experimental results on a dataset with over 1.1 million samples show that XGBoost outperforms other classifiers, producing the highest accuracy at 99.64%, the highest ROC-AUC at 0.9998, and the highest F1-score at 0.96. Unsupervised models continue to detect anomalous providers at a rate of 2%, whereas graph-based analysis detects suspicious communities among providers at 58% of the total set. Claim frequency and Medicare payment amounts appear as features used most frequently by feature attribution in detecting fraudulent activities.