Introduction

Background and Motivation

The availability of very large-scale clinical databases in electronic form has opened the possibility to generate systematic and large-scale evidence and insights about healthcare. This discipline is called Observational Outcome Research, and it uses longitudinal patient level clinical data in order to describe and understand the onset of disease and the effect of other clinical events as well as treatment interventions on the progression of the disease. Often, this research constitutes secondary use of the data, as they are being collected for purposes other than research: administrative data such as insurance reimbursement claims and Electronic Health or Medical Records (EHR, EMR).

Because of the collection purpose for primary use, the format and representation of the data follows that primary use. It also introduces artifacts and bias into the data. In addition, all source datasets differ from each other in format and content representation. Since healthcare systems differ between countries, the problem becomes even harder for research carried out internationally. All this makes robust, reproducible and automated research a significant challenge.

The solution is the standardization of the data model (syntax) and a standardization of the representation (coding). This allows methods and tools to operate on data of disparate origin, freeing the analyst from having to dissect the idiosyncrasies of a particular dataset and manipulating the data to make it fit for research. It also allows to develop analytical methods on one dataset, and applying it an any other dataset. Which means, patient-level data available in OMOP CDM explicitly require the representation of all facts, events and information using concepts from the Standardized Vocabularies. With few exceptions, there are no verbatim pieces of information in the CDM tables.

Initially, the Standardized Vocabularies were constructed together with the Common Data Model during the OMOP Partnership for the conduct of the OMOP experiments. That use case called for a very simple design, facilitating the minimum functionality necessary to conduct observational outcome studies. After OMOP ended in November 2013, the OMOP Standardized Vocabularies are now supported by OHDSI. The simplicity and pragmatism applied allows to maintain this resource within an Open Source Project and its limited resources.

Principles

The Standardized Vocabularies are constructed with a few principles in mind. Not every principle has been executed to perfection, but it represents a general motivation and direction of the ongoing improvement and development process:

1. OHDSI Use Cases Focused:
   OHDSI Vocabularies serve OMOP CDM and are driven by the use cases of storing and analyzing patient-level data for evidence generation. Storing data for other purposes such as keeping track of clinical transactions is outside of scope.
2. Persistency:
   Nothing gets deleted from the Vocabularies tables.
3. Openness:
   OHDSI Vocabularies are maintained as open-source and support transparency in generation processes.
4. Unique Standard Concepts:
   For each Clinical Entity there is only one concept representing it, called the Standard Concept. Other equivalent or similar concepts are designated non-Standard and mapped to the Standard ones.
5. Pre-defined Domains:
   Each concept is assigned one of a list of predefined domains (defined by OHDSI). “Dirty”, i.e., not well-defined concepts, which are mostly non-Standard, can also belong to more than one domain. The assigned domain also defines in which CDM table a clinical entity should be placed into or looked up in at query time.
6. Comprehensive Coverage:
   Every event in a patient's healthcare experience (e.g. Conditions, Procedures, Exposures to Drug, etc.) and some of the administrative artifacts of the healthcare system (e.g. Visits, Care Sites, etc.) are covered within the domain. This principle includes concept coverage and mapping coverage (mappings from non-standard concepts to standard concepts).
7. Polyhierarchy:
   Within a domain all S and C concepts are organized in a hierarchical structure. This allows to query for all concepts (e.g., drug products) that are hierarchically subsumed under a higher level concept (e.g. a drug class). This entails addressing two separate problems: o Each concept should have one or more classifications (bottom up). o Each classification should contain all the relevant concepts (top down).
8. No Negative Information:
   No concept should represent absence of evidence. Instead, the users should come up with data-level ways to model the information (exception is measurement values that explicitly model negative information).
9. No Timing:
   Temporality should be enforced from the data and not rely on the Vocabularies.
10. No Flavors of NULL:
    No standard concept should indicate flavors of null (unknown, no reported).
11. Pre-coordinated Concepts:
    Vocabularies do not support post-coordination. Disease attributes cannot be split.
12. Efficiency:
    Vocabularies should support efficient computations and use.

Work still needs to be done to achieve all the criteria in all of the domains. Currently, for the most important domains we can achieve the following compliance:

Domain	Unique Concepts	Pre-defined Reliable Domains	Comprehensive Coverage	Hierarchy
Drug	yes	yes	yes	yes
Condition	yes	mostly	yes	yes
Procedure	partial	yes	yes	partial
Measurement	partial	mostly	yes	partial
Device		mostly
Unit	yes	yes	yes

The life cycle is implemented for all concepts, and its rules are described in the CONCEPT table and in the discussion of the individual vocabularies.

It is important to note that these criteria have the purpose to serve observational research. In that regard the Standardized Vocabularies differ from large collections with equivalence mappings of concepts such as the UMLS, which supports indexing and searching of the biomedical literature. UMLS resources have been used heavily as a basis for constructing many of the Standardized Vocabulary components, but significant additional efforts have been made to adjust the framework:

• Additional Vocabularies, mostly for metadata purposes, are established.
• Mappings and relationships are being added to achieve comprehensive coverage. If equivalence cannot be achieved, “uphill” relationships from more granular non-standard to higher level Standard Concepts are created.