Skip to main content

Medical terminology: one standard code for every clinical concept

The Medical Terminology Server in Patient Journey Intelligence resolves every clinical concept, whether it comes from a structured EHR field or from text the Information Extraction pipeline found in a note, to a standard vocabulary code such as SNOMED CT, RxNorm, LOINC, ICD-10-CM, or CPT. It runs in the background of extraction, OMOP CDM transformation, and cohort building. When an extraction pipeline leaves its findings uncoded, IE Coding adds the codes afterwards.

A cardiologist documents MI, a hospitalist writes AMI, and the discharge summary says myocardial infarction. Medications and labs vary the same way: a pharmacy system lists ACEI-10 while the EHR says lisinopril 10mg tabs, and lab feeds label the same test HGB-A1C, A1c, and Hemoglobin A1c. A cohort query over raw strings counts each variant as a different concept, so it undercounts patients by an amount nobody can measure. One standard code per concept removes that error before anyone runs a query.

Medical Terminology in Patient Journey Intelligence

A two-minute tour: search NIDDM in SNOMED CT and ICD-10-CM, map ICD-10-CM codes to SNOMED CT, set up a coding profile, and code the 1,402 findings of an oncology extraction job with IE Coding.

Why cohorts need one vocabulary layer​

Take a cohort query for patients with type 2 diabetes. Against unnormalized data, the query has to list every variant clinicians have used: diabetes mellitus type 2, T2DM, NIDDM, DM2, adult-onset diabetes, non-insulin dependent diabetes mellitus, and the local codes behind problem lists and billing records. A missed variant drops patients from the cohort, and the size of the loss is unknown because nobody can list what they did not search for.

With terminology normalization, those variants resolve to SNOMED CT concept 44054006, Diabetes mellitus type 2. Searching the demo Terminology Server for NIDDM returns that SNOMED CT concept and ICD-10-CM code E11.9 with a confidence score of 1.0, so the abbreviation and the full term resolve to the same codes. The same principle applies to every drug, lab test, procedure, and finding in the dataset, and it is what lets an OMOP dataset take part in multi-site research networks, regulatory reporting, and AI agents that reason across patient populations.

Medical Terminology Default Search with SNOMED and ICD10CM selected, showing results for NIDDM: ICD10CM E11.9 and SNOMED 44054006 with confidence 1.0000, followed by related concepts such as NIDDM in obese, E66.9 and 81531005

Searching for the abbreviation NIDDM in SNOMED CT and ICD-10-CM returns SNOMED CT 44054006 and ICD-10-CM E11.9 as exact matches, ranked above related concepts.

Supported standard vocabularies​

The Terminology Server normalizes clinical concepts to five core terminologies, each the reference standard for its clinical domain.

SNOMED CT

Clinical terminology for conditions, findings, procedures, and anatomical structures. The broadest clinical vocabulary and the foundation of OMOP concept mapping.

Example: "Congestive heart failure" → SNOMED 42343007

RxNorm

Standardized drug names, ingredients, strengths, and dose forms maintained by the National Library of Medicine. The standard for medication interoperability.

Example: "Lisinopril 10 MG Oral Tablet" → RxNorm 314076

LOINC

Universal codes for laboratory tests, clinical measurements, and observations. Required for most lab result interoperability and quality reporting.

Example: "Hemoglobin A1c/Hemoglobin.total in Blood" → LOINC 4548-4

ICD-10-CM

Diagnosis codes for regulatory reporting, billing interoperability, and population health. The required standard for most US payer and regulatory submissions.

Example: "Type 2 diabetes mellitus with diabetic neuropathy, unspecified" → ICD-10-CM E11.40

CPT

Procedure codes maintained by the AMA for billing, quality measurement, and outcomes tracking. Essential for surgical and procedural research.

Example: "Coronary artery bypass, using arterial graft(s); single arterial graft" → CPT 33533

40+ additional vocabularies

Specialized terminologies including NDC, HCPCS, MedDRA, HPO, UMLS, NCI Thesaurus, MeSH, ICD-O-3, ICD-10-PCS, ATC, and CVX, covering oncology, genomics, rare disease, immunization, and global reporting.

Where terminology runs in the platform​

Medical terminology is rarely a manual step. It is built into the places where clinical concepts are created, stored, and queried.

Information Extraction

Entities the NLP pipeline extracts from clinical text can be normalized to standard vocabularies as part of extraction and enrichment, which makes free-text findings usable in cohort, registry, and curation workflows. When a pipeline does not code its findings, IE Coding codes them afterwards.

OMOP CDM transformation

OMOP-modeled facts rely on standard concept identifiers, vocabulary IDs, domains, and relationships. Normalization routes each clinical fact to the correct OMOP domain and keeps querying consistent across OMOP-based analytics.

Cohort Builder

Cohort definitions are more reliable when criteria use standard concepts instead of raw source text or local codes. Terminology mapping lets cohort logic include synonyms, descendants, and equivalent source values where supported.

Medical Reasoning

The inference engine uses concept relationships from the Terminology Server to identify drug classes, disease hierarchies, and related conditions when it deduplicates entities and resolves conflicts across sources.

Depending on permissions and deployment configuration, terminology appears as a direct search tool on the Medical Terminology page, as lookup fields inside other modules, or as a background mapping service. Every Cohort Builder query, research export, and AI agent prompt then operates on standard vocabulary codes, however the source data was documented.

How concept resolution works​

Concept resolution and terminology enrichment, step 4 of the NLP pipeline

Step 4 of the NLP pipeline resolves extracted medical phrases to standard vocabularies using cache pre-warming, entity-type classification, terminology server lookups, and batched processing. Enriched annotations are written back for OMOP modeling and downstream analytics.

Every clinical concept passes through the same four-step sequence, whether the NLP pipeline extracted it from a note or a structured field delivered it from an EHR:

1

Receive the clinical concept

NLP extracts 'metformin 1000mg BID' from a clinical note, or a structured field holds the local code 'MET-1000' from the pharmacy system.

↓
2

Search the terminology server

The server searches the vocabularies by name, synonym, abbreviation, or code. Entity-type classification narrows the search: a drug phrase queries RxNorm first, a diagnosis phrase queries SNOMED CT and ICD-10-CM.

↓
3

Resolve to a standard code

The best match is selected by semantic similarity and clinical context. Result: RxNorm 861004, metformin hydrochloride 1000 MG Oral Tablet.

↓
4

Store in OMOP CDM

The resolved concept_id is stored in the OMOP CDM with its provenance: source value, vocabulary version, and resolution confidence.

During ingestion this runs in the background, so concepts are normalized before a researcher opens the platform.

Look up and map codes yourself​

The Medical Terminology page under Data Enrichment is where you work with the Terminology Server directly. Use it when you need to:

  • search for a standard code before defining a cohort criterion
  • check whether an extracted entity was normalized correctly
  • understand why several source values appear under the same concept
  • review vocabulary, concept ID, domain, or mapping metadata
  • confirm that local codes have a standard representation before downstream analysis

Default Search finds concepts by name, code, synonym, or clinical shorthand, and understands abbreviations and spelling variations without exact matches. Choose the code systems on the left to narrow the search. Each result shows the code system, a confidence score, the concept name and code, whether the concept is valid and standard, and its OMOP domain.

Example queries:
  • Search hypertension to see SNOMED CT concepts for essential, secondary, and pulmonary hypertension, ranked by relevance.
  • Search A1c to see the LOINC codes for hemoglobin A1c across specimens and measurement methods.
  • Search ace inhibitor to see the drug class and the individual ACE inhibitor products, which supports class-level queries without listing every drug.

Concept mapping​

Your institution's internal codes stay in the data. They are mapped to standard concepts so both representations coexist, and a query on the standard concept also returns records coded with the local code.

Mapping examples:

Local pharmacy codes:

  • ACEI-10mg → RxNorm 314076 (Lisinopril 10 MG Oral Tablet)
  • STATIN-HIGH → RxNorm 259255 (Atorvastatin 80 MG Oral Tablet)

Local diagnosis codes:

  • HTN-benign → SNOMED 59621000 (Essential hypertension)
  • CHF-systolic → SNOMED 441481004 (Chronic systolic heart failure)

Local lab codes:

  • HGB-A1C → LOINC 4548-4 (Hemoglobin A1c/Hemoglobin.total in Blood)
  • TROP-I → LOINC 10839-9 (Troponin I.cardiac [Mass/volume] in Serum or Plasma)

Map Codes translates standard codes between vocabularies. Pick a concept map on the left, for example ICD10CM to SNOMED, and enter one or more codes. The mapping is rarely one to one: ICD-10-CM groups conditions more coarsely than SNOMED CT, so in the example below E11.9 maps to 59 SNOMED CT concepts, including 44054006.

Map Codes tab with the ICD10CM to SNOMED concept map selected, mapping I50.22, E11.9, E83.52 and C50.911 to their SNOMED target codes

Map Codes translates ICD-10-CM codes for chronic systolic heart failure (I50.22), type 2 diabetes (E11.9), hypercalcemia (E83.52), and right breast cancer (C50.911) to SNOMED CT.

Concept relationships​

Clinical concepts are organized in hierarchies. Metformin is a biguanide, which is an antidiabetic agent, so a query for patients on antidiabetic agents should include metformin without the researcher listing it. The Terminology Server encodes these relationships for cohort queries and AI agents.

Relationship types:

Hierarchical (IS-A):

  • Metformin IS-A Biguanide IS-A Antidiabetic agent
  • A query for antidiabetic agents includes every metformin product.

Cross-vocabulary (Maps-To):

  • SNOMED 44054006 (Diabetes mellitus type 2) maps to ICD-10-CM E11.9 (Type 2 diabetes mellitus without complications).
  • One source fact supports both a SNOMED-based cohort and an ICD-10-CM report.

The same structure supports queries such as patients on any ACE inhibitor or ARB, and it lets the Medical Reasoning module detect drug class interactions, contraindications, and treatment gaps.

Vocabulary versions​

Vocabularies update on their own schedules. ICD-10-CM changes every October 1, LOINC publishes twice a year, and RxNorm updates monthly. The Terminology Server keeps version metadata for every resolved concept, so historical analyses stay reproducible after an update.

  • SNOMED CT: US Edition, updated annually
  • RxNorm: monthly updates from the National Library of Medicine
  • LOINC: two releases a year
  • ICD-10-CM: annual updates, effective October 1
  • CPT: annual updates from the AMA

When an update retires a concept or maps it to a new code, existing data keeps its original code and the new code is available for forward-looking queries.

Code an Information Extraction job after the fact​

Some extraction pipelines return entity labels and text spans without codes. IE Coding, a tab of the Medical Terminology page, adds standard codes to the findings of a completed Information Extraction job, so those findings can feed cohorts, registries, and curation like any coded fact.

The walkthrough below uses a completed job from the demo environment: Hypercalcemia of Malignancy NER Extraction, which ran the OMOP NER Assertion Pipeline over 128 oncology documents and returned 3,108 findings across 41 entity labels. An oncology data team wants its diagnoses coded in SNOMED CT and ICD-10-CM, its drugs in RxNorm, and its tumors in ICD-O-3 as well.

Step 1: create a coding profile​

A coding profile records, for each entity label, which code systems to code it in. Open IE Coding, then Coding Profiles, then New Profile, and pick the job under Add labels from an IE job. The editor lists every label of the job with its number of findings and a starting set of code systems.

Labels of the John Snow Labs models start with the usual code systems for their kind:

Kind of labelStarting code systemsLabels in the hypercalcemia job
ConditionSNOMED CT, ICD-10-CMDISEASE SYNDROME DISORDER (319 findings), SYMPTOM (226), HEART DISEASE (56), KIDNEY DISEASE (32)
TumorSNOMED CT, ICD-10-CM, ICD-O-3CANCER DX (69), ONCOLOGICAL (38)
DrugRxNorm, RxNorm Extension, ATCDRUG (271), SUBSTANCE (9)
ProcedureSNOMED CT, CPT4, HCPCS, ICD-10-PCSPROCEDURE (220), IMAGING TECHNIQUE (6)
Lab test or vital signLOINC, SNOMED CTPULSE (29), BLOOD PRESSURE (26), TEMPERATURE (11), WEIGHT (9)

The kind of each label comes from its definition on the Models Hub, and the info icon next to a label shows that definition. Other labels start with the code systems of the Entity Mapping list. Labels such as directions, section headers, and names of people, and labels found in neither list, start unchecked: in this job DATE (349 findings), DOCTOR (285), and FREQUENCY (203) stay uncoded.

Uncheck a label or remove a code system you do not want, and add code systems to any label. Codes per code system sets how many codes are kept for each finding in each code system. The best one is shown and the rest are kept as alternatives.

New Coding Profile editor for the profile Oncology notes - conditions, drugs, procedures, built from the Hypercalcemia of Malignancy NER Extraction job, listing DATE and DOCTOR as not coded, DISEASE SYNDROME DISORDER and SYMPTOM with SNOMED and ICD10CM, DRUG with RxNorm, RxNorm Extension and ATC, and PROCEDURE with SNOMED, CPT4 and HCPCS

The profile editor starts each label of the job with the usual code systems for its kind and leaves labels such as DATE and DOCTOR uncoded.

When the profile is made from an Information Extraction job, Suggest code systems asks the decision model to classify each label. It reads some findings of the label, plus the label's Models Hub definition when there is one, and decides whether the label names a condition, a tumor, a drug, a procedure, a test, a body part, a device, or an observation such as a lifestyle factor. Each kind gets its usual code systems, and labels that are none of these, such as directions or dates, are left uncoded. Check the suggestions before you save. Suggestions and decision model scoring both need a model on the Entity Resolver → Decision Model slot of the Service Configuration page.

Step 2: start a coding run​

Click New Coding Run, pick the completed Information Extraction job, pick a saved profile, choose whether to use the decision model, and start the run from the Review step. You can also start from the job's page with Code Terms. If no profile fits yet, New Profile on the profile step opens the editor with the job's labels and brings you back with the new profile selected.

Before you start, the profile step shows how the selected profile will code the job, label by label. For the hypercalcemia job, the profile codes 24 of 41 labels and skips the other 17, which leaves 1,402 findings to code.

Preview on the Coding Profile step of a new coding run: How Oncology notes - conditions, drugs, procedures codes this job, 24 labels coded and 17 not in the profile, with ADMISSION DISCHARGE and AGE not in the profile, ALCOHOL coded in SNOMED, BLOOD PRESSURE in LOINC and SNOMED, and CANCER DX in SNOMED, ICD10CM and ICDO3

The Coding Profile step previews which labels the profile will code and in which code systems before the run starts.

Step 3: choose decision model scoring​

The Terminology Server returns the codes whose names are closest to each finding. It does not know whether a code is the right kind of concept for that sentence. Score the codes with the decision model, off by default, asks the decision model to read each finding marked in its document and choose which returned code of each code system the finding refers to, or none of them. In a very long document it reads the 600 characters on each side of the finding.

  • The same words can get different codes in different places. A term shows the average score of its findings.
  • Codes are ordered by that score, so the best fit becomes the main code. Every code is kept.
  • Codes scored below the threshold you set are not used as the term's code. The code details show them greyed out, with a DM score next to Similarity, which only says how close the code name is to the finding text.
  • Runs take longer with the decision model. The run page shows how many findings the decision model did not score, and those keep their search order. If it scored fewer than half of the findings, the run fails.

To add scores later, open a completed run and click Score with Decision Model. That starts a new run that scores the codes the run already has, without searching the Terminology Server again.

Step 4: review the codes​

Follow the run on its detail page. A run codes about one unique finding text per second, so a large job can take a while. The review step estimated up to about 23 minutes for the hypercalcemia job; the run coded all 1,402 findings in 3 minutes 37 seconds with no failures, because those findings share 276 unique terms.

Completed coding run for the Hypercalcemia of Malignancy NER Extraction job: 1,402 of 1,402 findings processed, 1,402 coded, 0 failed, duration 3m 37s, decision model off, with Open IE Job and Score with Decision Model buttons

The completed run: 1,402 findings coded in 3 minutes 37 seconds. Score with Decision Model can rescore these codes later without searching the Terminology Server again.

When the run finishes, open the job's Terminology coding tab to see the codes each term received. For the hypercalcemia job, the most frequent terms came back as:

TermEntity labelFindingsCodes
malignancyCANCER DX69 in 57 documentsICD-10-CM C80.1, ICD-O-3 8000/3, SNOMED CT 1240414004
hypercalcaemiaDISEASE SYNDROME DISORDER45 in 39 documentsICD-10-CM E83.52, SNOMED CT 66931009
myocardial infarctionHEART DISEASE36 in 35 documentsICD-10-CM I21.9, SNOMED CT 22298006
metastatic breast carcinomaONCOLOGICAL34 in 26 documentsICD-10-CM C79.81, ICD-O-3 8010/6-C50.9, SNOMED CT 1264495000
pulmonary embolismDISEASE SYNDROME DISORDER27 in 27 documentsICD-10-CM I26, SNOMED CT 59282003
Terminology coding tab of the Hypercalcemia of Malignancy NER Extraction job, listing coding run 218638c2 as completed and the coded terms malignancy (CANCER DX, 69 findings, ICD10CM C80.1, ICDO3 8000/3, SNOMED 1240414004) and hypercalcaemia (45 findings, ICD10CM E83.52, SNOMED 66931009)

The Terminology coding tab lists the job's coding runs and the codes each term received in the selected run.

The same table shows why the decision model exists. Without it, the code with the closest name wins. The ENTITY label was coded in RxNorm among other systems, so the term breast, a body part, received RxNorm 1486137, a product named Breast Essence II. The term sepsis received ICD-10-CM P36.9, Bacterial sepsis of newborn, unspecified, in a corpus of adult oncology notes. Narrowing the code systems of a broad label like ENTITY, or scoring the run with the decision model, removes both errors.

Open a document to see the codes of each finding. In the discharge summary below, myocardial infarction is coded ICD-10-CM I21.9 and SNOMED CT 22298006, and the extraction's assertion, family, appears next to the codes, so a reviewer sees that this is family history and not the patient's own diagnosis.

Document view on the Clinical Findings tab, with the finding myocardial infarction (HEART DISEASE, assertion family, confidence 62%) showing Terminology Server codes ICD10CM I21.9 and SNOMED 22298006 and the context sentence There is a family history of myocardial infarction

In a document, each finding shows its Terminology Server codes alongside the extraction's assertion and the sentence it came from.

A few rules decide which codes you see:

  • A job shows the codes of its last completed run. A cancelled or failed run leaves them as they were.
  • Every run keeps its own codes. The Terminology coding tab shows the latest completed run; pick another run above the table to see its codes. A document shows the codes of the latest completed run.
  • Labels the extraction pipeline already codes are skipped.
  • A job can have one run in progress at a time, and a run codes at most 100,000 findings.
  • Deleting an Information Extraction job also deletes its coding runs and their codes.

What standardization changes for your data​

Without terminology normalizationWith terminology normalization
Cohorts miss patients whose records use a different term for the same concept.You query standard concepts instead of every local synonym.
Analytics fragment because each project builds its own mapping logic.Analytics share one vocabulary foundation across platform modules.
Interoperability across registries, sites, and external partners is hard.OMOP and reporting workflows receive clean concept mappings.
Research results are biased by undercounting non-standard or local terms.Vocabulary updates are managed once, without rewriting every downstream workflow.

In practice that means two researchers running the same query at different times get the same cohort, because both query the same concept_id. The OMOP dataset uses the same vocabularies as other OMOP network sites, which supports federated queries and cross-institutional studies. Data is already mapped to ICD-10-CM, CPT, and other required code sets for regulatory submissions. AI agents reason over standard concept_ids, which removes the string variants that cause missed patients.

Common checks when a result looks wrong​

SymptomWhat to check
Search results are too broadAdd vocabulary, domain, or concept type filters where available.
A term does not resolveTry common synonyms, abbreviations, spelling variants, or the source code itself.
A mapped concept looks incorrectReview the source text, source code, vocabulary, domain, and mapping confidence if available.
A cohort count is lower than expectedCheck whether the criterion includes relevant synonyms, descendants, or mapped source concepts.
OMOP output is missing a conceptConfirm that the source value was extracted or ingested and that a standard mapping exists.
Terminology behavior changed over timeReview the vocabulary version metadata.

A successful mapping produces a standardized representation of the source concept: vocabulary, concept identifiers, source provenance, and relationship metadata where available. Those standardized concepts feed Information Extraction, Cohort Builder, Data Curation, Database Explorer, and OMOP-based analytics. For how the rest of the platform turns them into patient journeys, see the Patient Journey Intelligence product page, or schedule a demo to run IE Coding on your own extraction jobs.

Frequently asked questions​

They are preserved. The OMOP CDM stores both the original source value (your local code) and the resolved standard concept_id. Queries using the standard concept also capture data coded with local codes, and you can still filter or audit by source value.

Entity-type classification runs before the vocabulary lookup. The NLP pipeline assigns each extracted entity to a category such as drug, diagnosis, lab test, procedure, or anatomy from the surrounding sentence, and the Terminology Server restricts its search to the vocabularies for that category. Terms that remain ambiguous are flagged with several candidate concepts and a confidence score. In IE Coding, the decision model can also read each finding in its document and pick the code it refers to there.

Create a coding profile on Medical Terminology → IE Coding that lists the code systems for each entity label, then start a New Coding Run on the completed job, or click Code Terms on the job's page. When the run finishes, the job's Terminology coding tab shows the codes each term received.

A run codes about one unique finding text per second, so time depends on how many distinct finding texts the job has. Scoring with the decision model makes the run longer. A run codes at most 100,000 findings, and a job can have one run in progress at a time.

The Terminology Server ranks codes by how close their names are to the finding text. The decision model reads each finding in its document and chooses which returned code of each code system it refers to, or none. Codes are then ordered by that score, and codes below your threshold are not used as the term's code but are kept and shown greyed out.

Yes. For concepts with no standard equivalent, such as institution-specific composite codes, local care protocols, or internal risk categories, you can define custom mappings. They are stored in the OMOP CDM with the non-standard concept mechanism and stay queryable alongside standard concepts.

Every resolved concept_id is stored with its vocabulary version at resolution time. When a vocabulary update remaps a concept, historical records keep their original concept_id. You can query by concept at a historical version, or use the mapping tables to translate historical records to current codes.

The server supports 40+ vocabularies, including NDC, HCPCS, MedDRA, HPO, UMLS, NCI Thesaurus, MeSH, ICD-O-3, ICD-10-PCS, ATC, and CVX. The list follows the OMOP vocabulary catalog maintained by OHDSI.

The server uses cache pre-warming for frequent concepts and batched processing for high-throughput workloads, so terminology resolution adds little to an ingestion pipeline whose time is mostly spent querying source systems and transferring data.