NLP Finds Words. Clinical Ontology Understands Medicine.
Standard NLP in risk adjustment reads clinical text and recognizes medical terms. It identifies that a note contains “diabetes mellitus type 2,” “chronic kidney disease,” or “congestive heart failure.” That’s text recognition. It tells you what words appear in the document. It doesn’t tell you what those words mean in clinical context.
Clinical ontology is the knowledge layer that gives NLP clinical understanding. It knows that “A1C 8.2” is a monitoring element for diabetes. It knows that “GFR 42” indicates CKD Stage 3b. It knows that “increased furosemide to 40mg” is a treatment decision for heart failure. It understands the clinical relationships between findings, conditions, and management actions that determine whether documentation constitutes genuine evidence of active care.
Without ontological depth, NLP produces identification without comprehension. The system finds “diabetes” but can’t evaluate whether the surrounding documentation proves the condition is being managed. With ontological depth, the system understands not just that diabetes is mentioned but whether the note contains the clinical elements that constitute management evidence under MEAT criteria.
Why This Distinction Matters for Audit Defensibility
MEAT criteria require evidence of clinical management, not just disease mention. Monitoring means specific clinical activities (labs, vitals, imaging) relevant to the condition. Assessment means the provider’s clinical judgment about the condition’s status. Treatment means documented management decisions. Each element requires clinical understanding to evaluate, not just text pattern matching.
A system without clinical ontology can identify that a note mentions “CKD.” A system with clinical ontology can determine that the note contains a GFR result (monitoring), a staging assessment (assessment), a nephrology referral (treatment), and a medication adjustment for renal dosing (treatment). The first system finds the condition. The second system evaluates the evidence. The second system’s output is what auditors look for and what coders need to make defensible decisions.
OIG audits consistently find that diagnoses fail not because they’re absent from charts but because the surrounding documentation doesn’t prove active management. That’s an ontological failure, not an NLP failure. The text recognition worked. The clinical comprehension didn’t exist.
What Ontological Depth Looks Like in Practice
A clinically-aware system processes a chart and produces output like: “HCC 85 (CHF): Identified. Monitoring: BNP 450 pg/mL referenced (elevated, consistent with active disease). Assessment: ‘volume overload improving’ documented. Treatment: furosemide increased from 20mg to 40mg; cardiology follow-up scheduled. MEAT: 3/4 elements present. Defensibility: strong.” Each element reflects clinical understanding, not just text extraction.
Compare this to standard NLP output: “CHF identified. Confidence: 94%.” The confidence score reflects the model’s certainty that CHF appears in the text. It says nothing about whether the documentation proves active management. The coder receiving this output still has to read the full chart to make the defensibility assessment the system should have provided.
The Capability Test
Plans evaluating risk adjustment software should test for ontological depth, not just NLP accuracy. Feed the system a chart where a condition appears in the problem list but has no management evidence in the encounter note. If the system recommends the code with high confidence, it’s doing text recognition without clinical understanding. If it flags the MEAT gap and recommends against submission, it understands the difference between mention and management. That difference determines audit outcomes.
