Medical Code Mapping

The foundational layer identifies and extracts discrete medical concepts from unstructured text. A transformer-based NER model pinpoints spans of text corresponding to diagnoses, procedures, anatomical sites, and medications.

• Detects multi-word concepts like "acute myocardial infarction" as a single entity
• Distinguishes between similar terms based on context (e.g., "cold" as temperature vs. illness)
• Handles abbreviations, acronyms, and clinical shorthand
• Outputs character-level offsets for audit trail traceability

This component normalizes extracted clinical concepts to standard terminologies. It maps a physician's narrative description to a SNOMED CT concept ID or a LOINC code before translating to billing code sets.

• Uses dense vector embeddings for semantic similarity matching against terminology databases
• Resolves synonyms: "heart attack" → 22298006 | Myocardial infarction (disorder) |
• Maintains mappings between SNOMED CT, ICD-10-CM, CPT, and HCPCS code systems
• Handles post-coordinated expressions where a single concept requires multiple codes

Once a concept is normalized, the system must select the most specific billable code. ICD-10-CM codes exist in parent-child hierarchies, and the AI must navigate to the highest level of specificity.

• Traverses code trees: I21 (STEMI) → I21.0 (Anterior wall) → I21.01 (Left main coronary artery)
• Validates against Excludes1 and Excludes2 rules to prevent mutually exclusive code pairs
• Applies coding clinic guidelines and official conventions as deterministic constraints
• Flags codes requiring laterality, episode of care, or encounter type modifiers

Every mapped code must be defensibly linked to supporting clinical evidence in the source documentation. This component creates auditable traceability from the assigned code back to the specific sentence or finding that justifies it.

• Generates a provenance chain: Code → Normalized Concept → Extracted Entity → Source Text Span
• Highlights missing evidence that would support a more specific or higher-reimbursing code
• Provides clinical reviewers with direct navigation to the originating documentation
• Essential for payer audits and coding compliance reviews

A deterministic rules layer enforces payer-specific and regulatory coding requirements that go beyond clinical accuracy. This includes National Correct Coding Initiative (NCCI) edits, Local Coverage Determinations (LCDs), and payer-specific bundling rules.

• Automatically unbundles procedures when documentation supports modifier usage (e.g., -59, -XS)
• Applies Medically Unlikely Edits (MUEs) to flag excessive units
• Validates code pairs against procedure-to-procedure edit tables
• Ensures compliance with CMS and commercial payer billing guidelines

Not every mapping is high-confidence. This component assigns a probabilistic confidence score to each code assignment and routes low-confidence cases to certified coders for manual review.

• Scores reflect semantic ambiguity, missing evidence, or conflicting documentation
• Routes cases below a configurable threshold to a human-in-the-loop review queue
• Learns from reviewer corrections to improve future mapping accuracy
• Provides explainability: "Code I21.01 assigned at 68% confidence due to missing laterality documentation"