Search for a command to run...
Payer denials for high-cost oncology drugs are frequently driven not by coding errors butby missing or ambiguous clinical documentation, particularly around biomarker thresholdsand treatment sequencing. In prior work we introduced the CXR documentation reasoningmodule, a policy-aware engine that uses large language models (LLMs), uncertainty detection,and policy rules to assess whether unstructured oncology notes support claim admissibility. Inthis paper we describe a focused, synthetic proof-of-concept for pre-submission documentationreadiness: given a planned oncology claim, determine whether the current documentation islikely to support payment and explain why.We construct three tiers of synthetic non-small cell lung cancer (NSCLC) datasets (A/B/C)with increasing realism and noise—from ideal, perfectly documented cases to highly chaoticscenarios with missing laboratory results, qualitative-only PD-L1 statements, temporalgaps, and conflicting evidence. We then implement a lightweight documentation readinessengine driven by a simple PD-L1 policy (“PD-L1 ≥ 1% within 180 days”) and align itwith the production FactExtractor and EnhancedComplianceChecker used in CXR. Thereadiness engine prefers LLM-derived PD-L1 percentages when available, falls back to lab/noteparsing when necessary, and produces structured outputs (READY/NOT extunderscoreREADY/HIGH extunderscore RISK) with explicit missing elements, ambiguities, timingflags, and plain-language explanations.On Tier A (ideal data), the readiness engine and full CXR stack achieve 100% agreementon claim admissibility. On Tier B (best-case real-world imperfections with 30% qualitative-only PD-L1 documentation and modest lags), they agree on approximately 82% of claims; theremaining 18% disagreements are concentrated in edge cases where PD-L1 is described onlyas “positive” or “high” without a numeric threshold, highlighting the impact of policy choiceson how strictly to treat qualitative biomarker evidence. On Tier C (chaotic data with missingand stale evidence), both systems consistently mark all claims as not ready/non-compliantonce shared timing checks are applied. We argue that this alignment, especially in the mostchallenging tier, demonstrates that a fact-extraction-aligned readiness layer can faithfullyapproximate the behavior of a richer documentation reasoning engine while remaining simpleenough for rapid POC work and downstream integration.