SNOMED, LOINC, ICD-10, RxNorm: the vocabularies of medicine, decoded

A woman visits urgent care with a urinary tract infection. One visit, one problem, one prescription. By evening her record holds four codes: an ICD-10-CM code so the visit can be billed, a SNOMED CT concept on her problem list, a LOINC code for the urine culture the clinic ordered, and an RxNorm identifier for the antibiotic she picked up on the way home.

Four code systems for one afternoon. None of them redundant. Each was built by different people to answer a different question — and knowing which question is the difference between reading her record and misreading it.

Four vocabularies, four jobs

Each of the four major vocabularies encodes a different slice of medicine, and none can substitute for another.

ICD is the elder. It descends from international cause-of-death classifications and is maintained by the World Health Organization so that disease statistics can be compared across countries; the United States records and bills diagnoses in ICD-10-CM, a clinical modification with far more detail, mandatory since October 2015. Because reimbursement rides on ICD codes, they are the most complete data in any American record — and the most distorted, because they are shaped by payment rules as much as by medicine.

SNOMED CT is the deep one. Where ICD gives a diagnosis one administrative pigeonhole, SNOMED places it in a web: hundreds of thousands of concepts with formal relationships, so that a disorder knows its parents, its neighbors, and its defining features. It exists to record what clinicians mean, not what billers need.

LOINC codes the question rather than the answer. "Hemoglobin A1c in blood" gets a LOINC code; the 7.2% is a value attached to it. The Regenstrief Institute created LOINC in 1994 so laboratories could exchange results without re-translating one another's local test menus — before it, every lab interface was a bespoke mapping project.

RxNorm, published by the National Library of Medicine, tames drug naming. Brand names, generics, strengths, and dose forms are linked into one graph, so a branded capsule and its generic equivalent resolve to the same clinical drug. Ask "who is taking nitrofurantoin?" without RxNorm and you must already know every name the drug has ever worn.

Why one concept ends up with five codes

The same clinical fact legitimately lives in several vocabularies at once, because each system records a different aspect of it — and the systems do not line up one-to-one.

Take type 2 diabetes. The bill carries an ICD-10-CM code. The problem list carries a SNOMED concept. The supporting evidence is an A1c result under a LOINC code, and the treatment is metformin under RxNorm. Four codes, one disease, no contradiction. The trouble starts when granularity differs: SNOMED frequently distinguishes what ICD lumps together, so mapping in one direction loses information, and mapping back invents it. Local habits make it worse — two hospitals can code the same lab under different LOINC codes, or park a diagnosis at different levels of specificity for years.

For research, this is the quiet killer. A cohort query phrased in one vocabulary silently misses every record coded in another: a patient becomes invisible not because her data is absent but because it is spelled differently. Every computable phenotype that hopes to travel between sites has to speak all four languages or fail without telling anyone.

The mapping machinery

Two pieces of public infrastructure keep the four languages mutually intelligible: the UMLS Metathesaurus and the OMOP standard vocabularies.

The Unified Medical Language System, maintained by the National Library of Medicine, gathers a large set of source vocabularies into a single metathesaurus and assigns shared concept identifiers, so "the same idea" can be recognized across systems. It is the Rosetta stone most other mapping work builds on.

The OHDSI community goes a step further for research. The OMOP Common Data Model designates one standard vocabulary per domain — SNOMED CT for conditions, RxNorm for drugs, LOINC for measurements — and maintains curated mappings from everything else into those standards. When a hospital's data is converted to OMOP, each source code is mapped to its standard concept while the original is preserved alongside it. A study written once against standard concepts can then run at any site on the model, which is the whole point.

Exchange standards lean on the same choices. US Core, the implementation guide that governs US FHIR data, binds problems to SNOMED CT, labs to LOINC, and medications to RxNorm. The research stack and the exchange stack converged on the same answers, which is rarer in health IT than it should be.

Versions drift; pin them

Vocabularies are living publications — codes are added, deprecated, split, and remapped with every release — so an analysis is reproducible only if it can name the exact versions it ran against.

SNOMED CT now publishes monthly international releases (it moved from twice-yearly editions in 2022). RxNorm updates monthly, because drugs come and go that fast. The WHO has already moved on to ICD-11, in effect since 2022, while most US billing still lives in ICD-10-CM — a transition that will take years and quietly break naive code lists along the way. A phenotype that matched ten thousand patients in January can match a different ten thousand in July with nobody touching the query, purely because a vocabulary release moved underneath it. Treating mappings as part of data quality — versioned, validated, auditable — is the only durable defense.

Mithras Health builds on exactly that discipline. The platform keeps each patient's record as FHIR R4 with US Core 6.1.0 profiles as the system of record — SNOMED, LOINC, and RxNorm in their US Core bindings — and projects OMOP views for research, with vocabulary versions pinned so that any cohort can be reconstructed and audited later. Across 28 conditions in 12 categories, that single vocabulary spine is what lets a question asked about interstitial cystitis be re-asked about any other condition without rebuilding the plumbing.

The four vocabularies are usually described, with a sigh, as a mess. They are closer to four precise instruments that refuse to play in unison. The mess is real, but it does not live inside any of them — it lives in the spaces between, and it yields to the unglamorous work of mapping, versioning, and writing down which edition of which language you were speaking when you said what you said.