Lifebit logo
BlogTechnologyOMOP Tables Explained: What Each Clinical Table Holds

OMOP Tables Explained: What Each Clinical Table Holds

A captivating abstract art piece featuring intertwined 3D shapes on a blue background.
Photo by Steve A Johnson on Pexels

OMOP tables are the set of standard database tables that make up the Observational Medical Outcomes Partnership (OMOP) Common Data Model. Clinical tables such as PERSON, VISIT_OCCURRENCE, CONDITION_OCCURRENCE, DRUG_EXPOSURE and MEASUREMENT hold patient events, while vocabulary tables such as CONCEPT and CONCEPT_RELATIONSHIP give every event a standard code. Once data sits in these tables, the same analysis can run unchanged on any OMOP database in the world.

Why understanding OMOP tables matters

Teams adopt OMOP to make their data comparable with other sources and to use the open tools of the Observational Health Data Sciences and Informatics (OHDSI) community. The model itself is well documented, but the first questions on any mapping project are practical ones. Which table does a lab result go in? What happens to a diagnosis code that maps to a measurement? Why does every table have three different code columns? Getting these answers right early saves weeks of rework and avoids silent errors in later analysis.

This article walks through the OMOP tables by group, explains how they link, and covers the rules that decide where each source record belongs. For the background on what the model is and why it exists, start with our guide to the OMOP CDM. This article assumes version 5.4, the current release.

The OMOP tables, group by group

Person and time

PERSON holds one row per patient, with year of birth, gender, race and ethnicity as standard concepts. Every clinical table links back to it through person_id.

OBSERVATION_PERIOD records the spans of time when a person’s data is expected to be captured, such as the time they were enrolled with an insurer or registered with a practice. Analysis uses it to tell the difference between “no event happened” and “we could not have seen it”.

DEATH records the date and, where known, the cause of death.

Encounters

VISIT_OCCURRENCE holds each encounter with the health system: an inpatient stay, an outpatient visit, an emergency department visit. VISIT_DETAIL breaks a visit into parts, such as moves between wards or an intensive care stay within one admission. Clinical events can point to the visit in which they happened.

Clinical events

  • CONDITION_OCCURRENCE: diagnoses, problems and signs or symptoms.
  • DRUG_EXPOSURE: prescriptions, dispensings and administrations of medicines and vaccines.
  • PROCEDURE_OCCURRENCE: diagnostic and therapeutic procedures.
  • DEVICE_EXPOSURE: implants and other medical devices used on a patient.
  • MEASUREMENT: lab results, vital signs and other values, with numeric results, units and reference ranges.
  • OBSERVATION: clinical facts that fit nowhere else, such as family history, social history or answers to questionnaires.
  • SPECIMEN: biological samples taken from a patient.
  • NOTE and NOTE_NLP: free-text clinical notes and the terms extracted from them by natural language processing.

Health system and cost

LOCATION, CARE_SITE and PROVIDER describe where care happened and who delivered it. PAYER_PLAN_PERIOD records insurance coverage over time, and COST holds charges and payments linked to clinical events.

Derived tables

CONDITION_ERA, DRUG_ERA and DOSE_ERA roll individual records into continuous spans, such as a period of continuous treatment with one drug ingredient. They are built from the event tables, usually by standard scripts, and make many analyses much simpler. Version 5.4 also added EPISODE and EPISODE_EVENT for disease phases and treatment lines, mainly used in oncology.

Vocabulary tables

The vocabulary tables are what make OMOP work. CONCEPT lists every code from every supported vocabulary, each with a concept_id, its domain and whether it is a standard concept. CONCEPT_RELATIONSHIP links concepts, including the “Maps to” links from source codes to standard ones. CONCEPT_ANCESTOR holds the hierarchy, so a query for a broad condition also finds its more specific children. Supporting tables include VOCABULARY, DOMAIN, CONCEPT_CLASS, RELATIONSHIP, CONCEPT_SYNONYM, DRUG_STRENGTH and SOURCE_TO_CONCEPT_MAP, which holds local mappings a site adds itself.

Metadata and results

CDM_SOURCE and METADATA describe the database itself: where the data came from, when it was extracted and which vocabulary release was used. COHORT and COHORT_DEFINITION store groups of patients defined for analysis.

OMOP tables at a glance

TableWhat it holdsUsual standard vocabularyExample record
PERSONOne row per patientOMOP gender, race and ethnicity conceptsPatient born 1962
VISIT_OCCURRENCEEncountersOMOP visit conceptsEmergency department visit
CONDITION_OCCURRENCEDiagnoses and symptomsSNOMED CTType 2 diabetes
DRUG_EXPOSUREMedicines and vaccinesRxNorm, RxNorm ExtensionMetformin 500 mg tablet
PROCEDURE_OCCURRENCEProceduresSNOMED CT, CPT-4, HCPCSKnee replacement
MEASUREMENTLabs and vital signsLOINC, SNOMED CTHbA1c of 7.2 percent
OBSERVATIONOther clinical factsSNOMED CT, LOINCFamily history of breast cancer
CONCEPTEvery code in every vocabularyAllAn ICD-10-CM code and its SNOMED CT match

How the tables link

Every clinical record carries a person_id that points to PERSON. Most also carry a visit_occurrence_id that ties the event to an encounter, and a provider_id for the clinician involved. Each code column points to a concept_id in CONCEPT. FACT_RELATIONSHIP links records that belong together across tables, such as a measurement and the condition it confirms.

The three code columns

Each event table stores a code three ways. The concept_id column holds the standard concept used for analysis. The source_value column keeps the original code or text exactly as it arrived. The source_concept_id column holds the OMOP concept for that original code, if one exists. Keeping all three means analysts get a standard view while the original is never lost. A fourth column, type_concept_id, records where the record came from, such as an EHR problem list or an insurance claim.

Where each source record belongs

The rule that surprises most teams is that the domain of the standard concept, not the source table, decides where a record goes. A code that arrived in a diagnosis list but maps to a standard concept in the Measurement domain belongs in MEASUREMENT. A procedure code that maps to an Observation concept belongs in OBSERVATION. Mapping tools look up the domain in CONCEPT and route each record accordingly.

When a source code has no standard mapping, the record still goes in, with concept_id set to 0 and the original code kept in source_value. That keeps the record visible and makes gaps easy to measure. Our OMOP implementation checklist covers how to track and close those gaps.

A worked example

Take a hospital record for one outpatient appointment. The appointment itself becomes a row in VISIT_OCCURRENCE. The ICD-10-CM diagnosis code on the bill is looked up in CONCEPT, followed through its “Maps to” link to a SNOMED CT concept, and written to CONDITION_OCCURRENCE with the original ICD-10-CM code kept in source_value. The blood pressure taken at the start of the appointment goes to MEASUREMENT with its value and unit. The new prescription goes to DRUG_EXPOSURE, mapped to RxNorm. All four rows share the same person_id and visit_occurrence_id, so an analyst can see them as one encounter.

Common mistakes with OMOP tables

  • Routing by source table. Putting every diagnosis-list code in CONDITION_OCCURRENCE regardless of its domain.
  • Dropping unmapped records. Losing data instead of loading it with concept_id 0.
  • Skipping OBSERVATION_PERIOD. Without it, analyses cannot tell missing data from absent events.
  • Mixing vocabulary releases. Mapping some data with one release and some with another. Record the release in CDM_SOURCE.
  • Losing units. Loading lab values without unit concepts, which makes results impossible to compare.

OMOP tables in a federated setting

OMOP is especially useful when data cannot be pooled. If every site maps its data to the same OMOP tables, an approved analysis can be sent to each site and run there, with only aggregate results returned. The data never leaves the source, and the analysis code does not change from site to site. That is how federated networks run studies across many hospitals and countries. A Trusted Research Environment provides the secure workspace and output checks around that analysis, and our guide to data harmonization explains how source data gets into a common model in the first place.

Practical next steps

Before mapping, profile your source data so you know which tables each source will feed. Load the vocabulary tables from a single, recorded release. Map the person, observation period and visit tables first, since everything else depends on them. Then map clinical events domain by domain, measuring the share of records with concept_id 0 as you go. Finally, run the OHDSI data quality checks and fix what they find before anyone runs a study.

Frequently asked questions

What are OMOP tables?

They are the standard tables of the OMOP Common Data Model: clinical tables for patients and their health events, health system and cost tables, derived era tables and vocabulary tables that hold standard codes.

Which OMOP table holds lab results?

MEASUREMENT holds lab results and vital signs, including numeric values, units and reference ranges.

What is the difference between CONDITION_OCCURRENCE and OBSERVATION?

CONDITION_OCCURRENCE holds diagnoses, problems and symptoms. OBSERVATION holds other clinical facts, such as family or social history, that do not fit a more specific table.

What does concept_id 0 mean?

It means the source code had no standard mapping. The record is kept, with the original code in source_value, so the gap can be measured and fixed.

Which OMOP version should I use?

Version 5.4 is the current release and is supported by the main OHDSI tools.

Why do OMOP tables help federated analysis?

When every site uses the same tables and vocabularies, one analysis can run unchanged at each site, so only results need to travel.