Paper 01 • System Architecture
AI Medical Coding: Why Evidence Extraction and Code Adjudication Must Be Separated
Authors: HealthyClaim Engineering
•
Reviewed: September 2026
•
Topic: Architecture Isolation
Abstract: Modern Large Language Models (LLMs) excel at natural language comprehension and semantic synthesis. However, deploying them directly as end-to-end medical coding systems introduces critical failure modes: probabilistic hallucinations of unperformed procedures, inability to guarantee compliance with statutory NCCI unbundling rules, and non-deterministic behavior across identical patient charts.
This paper outlines the architectural necessity of isolating the pipeline into two distinct components: a bounded linguistic extraction model that identifies documented clinical observations, and a compiled, deterministic rules engine that performs all subsequent CPT, ICD-10-CM, and modifier adjudication against official CMS and AMA datasets.
Core Finding: Separating extraction from adjudication eliminates direct LLM code hallucination and enables byte-for-byte deterministic replay across identical clinical inputs.
Paper 02 • Comparative Analysis
Deterministic Medical Coding vs. LLM-Only Coding: A Verification Architecture Comparison
Authors: HealthyClaim Engineering & Billing Research
•
Reviewed: September 2026
•
Topic: Empirical Benchmarks
Abstract: We compare two prevailing paradigms in automated revenue cycle management: (1) Single-prompt or multi-agent LLM systems where the model directly produces billing line items, and (2) Dual-stage architectures where the LLM is restricted to fact extraction and deterministic software evaluates NCCI, MUE, and LCD criteria.
Through synthetic case analysis, we demonstrate that LLM-only systems exhibit significant error rates in modifier bypass determination (e.g., inappropriately appending Modifier 59 without anatomical distinctness) and frequently violate MUE unit ceilings. Deterministic verification engines achieve 100% adherence to published NCCI and MUE logic.
Core Finding: Deterministic verification provides complete audit defensibility, because every adjudication decision corresponds to a specific, citeable published rule rather than a latent neural weight.
Paper 03 • Data Theory
Closed-World Verification in Healthcare Revenue Cycle Management
Authors: HealthyClaim Engineering
•
Reviewed: September 2026
•
Topic: Mathematical Verification
Abstract: Under an open-world assumption, an AI model may extrapolate missing clinical facts (e.g., inferring that an injection was administered bilaterally when only a unilateral site is documented). In medical billing, this behavior constitutes unsupported billing and severe audit risk.
We formalize the Closed-World Assumption (CWA) for clinical coding: an assertion is true if and only if it is explicitly grounded in the documented clinical narrative. If vital information (such as laterality, time, or physician involvement) is unstated, the system must fail-closed to a structured provider clarification rather than inferring the fact.
Core Finding: Closed-world verification guarantees zero unsupported billing, transforming automated claim generation from an audit liability into a defensible clinical intelligence tool.
Research Notice & Scope
These papers reflect ongoing research by the HealthyClaim engineering team. Research methodologies and benchmarks utilize synthetic clinical data and published CMS statutory sources. These publications are for technical evaluation and peer discussion and do not constitute billing advice or official legal interpretations.