A Hybrid LLM–Graph Architecture for Laboratory Billing Discrepancy and Overpayment Risk Detection

Abstract

Mismatches between physician orders, specimen collection records, and billed procedure codes are a persistent and under-examined source of financial leakage in clinical laboratory and microbiology billing. Deterministic rule engines, such as the National Correct Coding Initiative (NCCI) maintained by the Centers for Medicare & Medicaid Services (CMS) to promote correct coding and reduce improper payments, reliably catch exact code-pair combinations that should not be billed together, but they have nothing to say about the free-text, clinically worded order language that microbiology testing routinely produces. We propose a hybrid architecture that pairs a deterministic rule-matching layer and a canonical order-specimen-test-claim linkage graph with a retrieval-constrained large language model (LLM) reconciliation component built to close that gap. The architecture spans seven layers: data ingestion and normalization; a linkage graph in which discrepancies surface as graph anomalies (orphan claims, unbundling, duplicate specimen use, temporal inversion); a Policy, Terminology & Clinical Knowledge Layer spanning coding rules, coverage policies, and clinical knowledge; a hybrid detection engine that combines rule matching, LLM-based semantic reconciliation, and statistical anomaly scoring; a retrieval-augmented generation (RAG) explanation layer that grounds every flagged case in cited rule or guideline text; a human review workflow; and a feedback loop that recalibrates detection thresholds and prompts from auditor decisions. The design rests on one non-negotiable rule: the LLM component never determines a billing code, and it is never asked to. Its narrowest and most consequential task is to answer a single bounded question, whether the observed, performed test semantically corresponds to the ordered test, while a deterministic crosswalk and validation pipeline separately generates every candidate code and finalizes every outcome, which reduces the risk of unsupported code generation while preserving traceability to validated catalog entries. The system is designed to detect laboratory billing discrepancies and prioritize potential overpayment and compliance risk for human review, not to detect fraud: a flag indicates that two or more records disagree in a way that affects payment, and any determination of intent remains a judgment for human reviewers and, where warranted, compliance or legal staff to make. We present this architecture as a design-science contribution, walk through the reconciliation component's prompt and output schema in detail, and discuss what we expect in terms of precision, explainability, and deployment trade-offs. Empirical validation against production claims data has not yet been carried out and is identified as the necessary next step.