Hi all,
Sharing findings from a systematic evaluation we ran at Corewell Health. We’re testing whether OMOP vocabulary tables can serve as a code discovery tool for querying Epic EHR source data directly — and we want the community’s perspective on whether the gaps we found are expected limitations or structural issues that also affect standard OMOP CDM queries.
What We Were Trying to Do
Our honest broker team builds research cohorts from Epic Clarity and Caboodle. Investigators typically arrive with ICD-10 code lists derived from protocols or literature. We wanted to test whether CONCEPT_ANCESTOR could replace or augment manual ICD-10 enumeration — specifically:
- Anchor at a SNOMED concept (e.g., “Congenital heart disease”)
- Traverse CONCEPT_ANCESTOR to find all descendant SNOMED concepts
- Reverse-map descendants to ICD-10-CM via CONCEPT_RELATIONSHIP (‘Mapped from’)
- Match ICD-10 codes against Epic’s DiagnosisTerminologyDim to find patients
If this worked reliably, it would give us hierarchy-based cohort definitions that self-maintain across ICD-10-CM fiscal year updates — a significant advantage over static code lists that silently stale.
We recognize this is not the primary intended use of the vocabulary (which is standard-concept traversal within a CDM). We’re asking: is this a reasonable secondary use case, or are the gaps we found expected limitations when operating outside the standard OMOP CDM query pattern?
How We Tested It
Three clinical domains at increasing complexity. For each, we compared:
- Hierarchy approach: CONCEPT_ANCESTOR from a SNOMED anchor → ICD-10 code expansion → patient lookup
- Manual approach: ICD-10 codes a competent analyst would enumerate from clinical knowledge
Patient-level concordance measured on a 4.5M patient EHR.
Results
Domain 1: Parkinson’s Disease (simple)
| Metric | Hierarchy | Manual |
|---|---|---|
| SNOMED descendants | 12 | — |
| ICD-10 codes produced | 8 | 8 |
| Patients found | 209 | 209 |
| Concordance | 100% |
No divergence. Both approaches find exactly the same patients. (One known mapping gap — PSP/G23.1 — is being addressed in the February release per a separate issue.)
Domain 2: Congenital Heart Disease (complex)
Anchor: 312723 (“Congenital heart disease”)
| Metric | Hierarchy | Manual (Q20-Q24) |
|---|---|---|
| SNOMED descendants | 928 | — |
| ICD-10 codes produced | 54 | ~80 |
| Patients found | 108,322 | 129,759 |
| Concordance | ||
| Both approaches | 108,092 | |
| Hierarchy-only | 230 | |
| Manual-only | 21,667 |
The hierarchy misses 21,667 patients. Root cause: Atrial Septal Defect (4289309) and Ventricular Septal Defect (434462) are not descendants of “Congenital heart disease” (312723). They sit under “Cardiac septal defects” → “Structural disorder of heart” → “Heart disease” — a parallel branch.
The 230 hierarchy-only patients are genuine discoveries: Eisenmenger’s syndrome (I27.83), Williams syndrome (Q93.82), and neonatal CHD screening (P09.5).
Domain 3: Venous Thromboembolism (complex, dual failure)
Anchors: 444247 (“Venous thrombosis”) + 440417 (“Pulmonary embolism”)
| Metric | Hierarchy | Manual (I26, I80-I82) |
|---|---|---|
| SNOMED descendants | 472 | — |
| ICD-10 codes produced | 212 | ~100 |
| Patients found | 255,343 | 261,764 |
| Concordance | ||
| Both approaches | 199,901 | |
| Hierarchy-only | 55,442 | |
| Manual-only | 61,863 |
VTE shows failure in both directions:
Under-capture (61,863 patients missed): The most common DVT code — I82.409 (“Acute embolism and thrombosis of unspecified deep veins of lower extremity,” 112K patients in our system) — maps to SNOMED 40481089 (“Embolism from thrombosis of vein of lower extremity”), which is NOT a descendant of “Venous thrombosis” (444247). The FY2023 peroneal vein additions (I82.451/452) map to “Disorder of vein of lower extremity” — also outside the hierarchy.
Over-capture (55,442 patients included): The hierarchy captures hemorrhoids (K64, 17.5K patients), retinal vein occlusion (H34, 10.7K patients), and obstetric venous complications (O22, 7K patients). Ontologically correct — hemorrhoids ARE venous thrombosis — but no VTE researcher would include them.
What We Learned
- Simple, well-coded domains (Parkinson’s): Hierarchy works perfectly. Manual and hierarchy produce identical results.
- Complex domains (CHD, VTE): Hierarchy is unreliable as a standalone approach. Miss rates of 17-24% on the most common codes. The hierarchy reflects SNOMED’s biological taxonomy, which doesn’t always align with how diseases are grouped for clinical research.
- The failure isn’t in the reverse-mapping step. The missed ICD-10 codes DO have SNOMED mappings — just to SNOMED concepts that sit outside the expected ancestor hierarchy. This suggests the same gap would affect standard OMOP CDM queries using CONCEPT_ANCESTOR (the
condition_concept_idfor these patients would still not be a descendant of the anchor). - The hierarchy IS valuable as a discovery tool. In every domain it found clinically relevant codes that manual enumeration missed (Eisenmenger’s in CHD, obstetric VTE). The problem is completeness, not relevance.
Our Questions for the Community
- Is this a known limitation? Do experienced ATLAS users expect CONCEPT_ANCESTOR-based cohort definitions to have this level of incompleteness for complex domains? Is there guidance on which domains are “safe” for hierarchy-only definitions?
- Would these gaps also affect standard OMOP CDM queries? If a patient’s condition_occurrence record has condition_concept_id = 4289309 (ASD), a query anchored at ancestor 312723 (CHD) would miss them in the CDM too — correct?
- Is there a recommended multi-anchor strategy? For CHD, should we combine 312723 + 318549 (Cardiac septal defects)? Is there documentation or community guidance on selecting anchor concepts for complex domains?
- Is using the vocabulary for ICD-10 code discovery (our use case) reasonable, or is this an unsupported pattern? If the vocabulary team’s position is “CONCEPT_ANCESTOR is only guaranteed complete within standard-concept space, not for reverse-mapping to source vocabularies,” that’s a useful answer that sets expectations.
Our Current Approach
Based on these findings, we’ve adopted a hybrid methodology:
- Use CONCEPT_ANCESTOR as a discovery and validation layer (finds codes you didn’t know about, flags gaps in your manual list)
- Maintain manual ICD-10 code lists as the primary cohort definition (informed by clinical judgment)
- Compare the two and investigate discrepancies before finalizing
This works for us operationally, but we’d like to understand whether the community considers the hierarchy gaps fixable (vocabulary improvements over time) or inherent (a characteristic of how SNOMED’s ontology maps to clinical research groupings).
Technical Details
- Vocabulary version: v20260227 (February 2026 Athena release)
- EHR: Epic (Caboodle data warehouse)
- Matching path: CONCEPT_ANCESTOR → CONCEPT_RELATIONSHIP (‘Mapped from’) → ICD-10-CM concept_code → DiagnosisTerminologyDim.VALUE
- System: 4.5M patients, Corewell Health (21-hospital system)
Happy to share the SQL or additional detail on any domain. We plan to continue testing additional domains and can report back.
Roger Carlson Corewell Health Research Institute
