AnExplainable AI Engineering Framework for Claims-Only First-Stage Provider Audit Triage Using SHAP-Guided Hybrid Retrieval-Augmented Generation

This study proposes an explainable artificial intelligence (XAI) engineering workflow for provider-level healthcare claim audit prioritization using SHAP-guided hybrid retrieval-augmented generation (RAG). The framework integrates provider-level claim aggregation, tree-based risk screening, SHAP explanation, exploratory group-level SHAP clustering, policy concept retrieval, and constrained large language model audit narrative generation. Experiments on a public Medicare provider fraud dataset use the dataset-provided PotentialFraud label as a weak audit prioritization label rather than a legal determination of fraud. The results show that reimbursement exposure, utilization duration, claim repetition, deductible patterns, and beneficiary case mix contribute to provider-level risk scores. Additional cross-validation, calibration, threshold, and scale-confounding analyses indicate that provider size and financial exposure are important confounders, while non-scale and contextual features also retain predictive information. Beyond prediction, the framework organizes local SHAP drivers into exploratory provider archetypes and maps explanation patterns to audit-relevant policy concepts. Compared with pure embedding retrieval, the SHAP-guided hybrid retriever increases policy concept diversity and explanation alignment, although these retrieval metrics do not replace independent expert audit validation. Because the public dataset does not include referral pathways, inter-facility transfers, provider–network relationships, or care-coordination records, the framework cannot determine whether utilization patterns are explained by clinically appropriate referrals, regional access constraints, or multi-level care pathways. Its current applicability is therefore limited to provider-level audit prioritization using the available claims and beneficiary variables. The proposed system is positioned as a reproducible engineering prototype for cautious, human-reviewed audit support rather than a comprehensive or automated fraud determination system.