Common Data Model

A common data model (CDM) is a comprehensive framework that includes standardized definitions, data structures, specifications, and operational rules that allow data from different sources to be organized, presented, and analyzed in a uniform manner. By transforming diverse datasets into a consistent format, a CDM enables interoperability across institutions, databases, and research networks.

Healthcare data are generated and stored in many different formats across institutions, electronic health record systems, claims databases, and research registries. Without standardization, combining data from these diverse sources requires extensive, ad hoc data harmonization for each study, which is time consuming, error prone, and difficult to reproduce. Common data models solve this problem by providing a standardized structure that all participating institutions transform their data into, enabling multicenter studies, distributed analytics, and large scale observational research without requiring raw data to leave the originating institution. Major CDMs include the OMOP Common Data Model, used by the Observational Health Data Sciences and Informatics (OHDSI) network, which spans hundreds of millions of patient records in dozens of countries; the Sentinel Common Data Model, used by the US Food and Drug Administration's Sentinel System, whose data come from the US health care system; and the PCORnet Common Data Model, used by PCORnet, the US National Patient-Centered Clinical Research Network.

A CDM specifies the tables, fields, data types, and standardized vocabularies that constitute the target data format. Each participating institution transforms (or "maps") its local data into the CDM format using an extraction, transformation, and loading (ETL) process. This process maps local data elements to CDM fields, converts hospital specific codes into recognized coding systems, and applies quality checks to ensure consistency. CDMs differ in how far they standardize the codes themselves: the OMOP CDM maps source codes to standard concepts (for example, ICD-10-CM diagnosis codes to SNOMED CT, and NDC drug codes to RxNorm), whereas the Sentinel and PCORnet CDMs store each code together with a field naming its coding system, such as ICD-10-CM or SNOMED CT.

Once data are in the CDM format, standardized analytical tools and queries can be run across all participating sites. In a distributed research network, each site runs the same analysis code on its local CDM database and typically returns only aggregate results, so individual level data need not leave the site, which helps protect patient privacy. This "bring the analysis to the data" approach is fundamental to large scale pharmacovigilance, comparative effectiveness research, and health outcomes studies.

The OMOP CDM, maintained by the OHDSI community, covers clinical data domains including conditions, drug exposures, procedures, measurements, observations, devices, and costs. Standardized vocabularies ensure that the same clinical concept (such as "type 2 diabetes mellitus") is represented consistently across sites, regardless of the local coding system used.

Common data models are appropriate for multicenter observational research, distributed research networks, pharmacoepidemiological studies, health system surveillance, and any project that requires combining or comparing data from multiple heterogeneous sources. They are most valuable when the research question requires large sample sizes, diverse patient populations, or rapid evidence generation across multiple sites. CDMs are less necessary for single site studies using a single data source, though even in those contexts, adopting a CDM can facilitate future data sharing and replication.

A common error is underestimating the complexity and resources required for the ETL process. Mapping local data to a CDM requires deep knowledge of both the source data and the CDM specifications, and it can take months to complete and validate. Inadequate ETL procedures can introduce systematic errors that propagate through all subsequent analyses.

Another mistake is assuming that data in a CDM format are automatically comparable across sites. Even with standardized structures and vocabularies, differences in clinical practice, coding conventions, data capture completeness, and patient populations can affect the comparability of data across institutions. Researchers should conduct site level data characterization and assess data quality before pooling or comparing results. Finally, neglecting to involve clinical domain experts in the ETL mapping process can lead to incorrect or misleading data transformations, particularly for complex clinical concepts that map differently across coding systems.

A hypothetical international research network using the OMOP CDM includes 48 databases across 19 countries, covering over 680 million patient records. A pharmacovigilance study comparing the risk of hepatotoxicity in users of a new anticoagulant with that in users of an established anticoagulant runs the same analysis code at all sites simultaneously. Each site returns only aggregate results (hazard ratios and confidence intervals). The pooled analysis, combining data from 2.3 million anticoagulant users, identifies a statistically significant signal (HR = 1.42, 95% CI: 1.18 to 1.71) that was not detectable in any single database due to insufficient sample size.