Clinical MetaData ontology: a simple classification scheme for data elements of clinical data based on semantics

Clinical data should be collected in a consistent manner by applying a standardized format so as to facilitate unified data collection, sharing, and integration. Clinical data from multiple sites need to be effectively integrated and compared in order to improve patient care and clinical research. There have been numerous efforts to standardize clinical data. One approach is to construct a common data model, which we call a top-down approach since a top-level knowledge model agreement is applied to the underlying data models of the interoperating parties to ensure successful data exchange [1]. The HL7 Reference Information Model (RIM) and EN 13606 standards are representative data models in the healthcare domain, which include generic reference models of concepts and relationships (e.g., CEN/ISO 13606, openEHR Reference Model, and HL7 RIM) and more-detailed models (e.g., openEHR Archetypes/Templates and HL7 Detailed Clinical Model) [2, 3]. To achieve semantic interoperability, these models utilize connected terms from various standard terminologies such as Unified Medical Language System (UMLS), SNOMED-CT, and LOINC. However, a major problem of this top-down ontology-construction approach is that it takes an unacceptably long time for users to adopt and use them under all possible situations that are encountered in healthcare [4].

Another approach is to simply encourage clinical users to employ standard vocabularies to encode complex real-world concepts. For instance, the patient may say ‘During the accident I sustained an injury to the back of my head and neck.’ No concept in UMLS corresponds to the problem ‘injury to the back of the head and neck.’ Instead, the problem concept Injury (C3263722), the direction Back (C0205095), and anatomical locations Head (C0018670) and Neck (C0027530) must be postcoordinated into an on-the-fly concept [5]. We call this a bottom-up approach because complex real-world concepts are constructed from elementary vocabularies. However, a complex clinical concept is often too sophisticated to be comprehensively encoded in a uniform manner, resulting in postcoordination ambiguities with many different encodings that in turn prohibit semantic interoperability.

The ISO/IEC 11179 international standard for a metadata description and registry (MDR) specifies a metadata model for representing the common data elements (CDEs) that are a logical data unit that provides for data definitions (including an identifier), response values to indicate the value type, and detailed information to represent data concepts and their semantics [6, 7]. The CDE consisted of two parts: a data element concept (DEC) for the meaning of data, and the value domain (VD). For instance, the DEC for a person’s sex is established using the object concepts of Person (C0027361) and Sex (C1522384), while the VD for sex covers the permissible values of ‘male’ and ‘female.’ The complete CDE is defined by combining the DEC for a person’s sex with VD < male|female> [8].

Well-defined CDEs can be collected and reused as a content standard. Here we propose ‘middle-out’ approach that contrasts with the top-down ontology-construction and bottom-up vocabulary-encoding approaches for achieving semantic interoperability at the level of clinical data elements contained in clinical documents. We first create and use CDEs based on the ISO/IEC 11179 standard for an MDR and then incrementally improve their quality. We call this a ‘middle-out’ approach because CDEs are pragmatically defined at a highly practical level for immediate real-world use, and then we systematically link them ‘up’ to the ontology classifications and ‘down’ to the standard controlled vocabularies. CDEs are designed at a pragmatic or ‘middle’ level and then ‘out’ to the higher ontologies and the lower vocabularies in a systematic way.

There are multiple benefits in using CDEs based on the ISO/IEC 11179 standard, including (1) effective and rapid data collection that reduces the burden on investigators and thereby facilitates their participation in clinical research, (2) improved data sharing and data aggregation due to employing common forms and standard definitions, and (3) higher data quality by providing unified data and their descriptions [9, 10]. Numerous large-scale clinical studies have developed standardized CDEs based on ISO/IEC 11179, such as that from the National Cancer Institute [11, 12], the National Institute of Neurological Disorders and Stroke CDE project [13‐15], and other clinical projects with various aims [16, 17]. The National Institutes of Health encourage the use of CDEs [17], which has led to them being deployed in case-report forms (CRFs) and clinical documents [6, 7, 10], and subsequently demonstrating the high effectiveness and usability of employing CDEs. We also implemented ISO/IEC 11179 based on a metadata registry called the Clinico-Histopathological Metadata Registry (CHMR), which contains more than 20,000 highly curated CDEs [18, 19].

One important characteristic of a successful MDR is being able to efficiently search for an appropriate CDE stored in it. For this purpose ISO/IEC 11179 provides a classification scheme (CS) structure for the conceptual classification and identification of data elements. Thus, when constructing an MDR or registering designed CDEs into an MDR, it is also necessary to select or design the contents of the CS using controlled vocabularies [20, 21]. However, most MDRs do not fully utilize or register a CS, and some MDRs support only two or three concept items in each CS for classifying their own metadata. Moreover, most of the CDE browsers that are developed in projects do not apply formal CSs, relying instead on simple keyword-based search engines. Keyword searching suffers from imprecision and ambiguity; for example, documents containing synonyms of the query keywords will not be retrieved, and homonyms cannot be properly managed. An ontology-based search approach can be considered as one example of a semantically enhanced information retrieval method.

We successfully established the CHMR in 2006, since when we have been using it in various clinical trials and research studies [22‐24]. The main limitation we have experienced during the long-term use of an MDR is the lack of an ontology that can be used to semantically group, search, and integrate metadata. The objective of the present study was to develop an ontology—called the Clinical Metadata Ontology (CMDO)—for managing, retrieving, classifying, and integrating CDEs with the rich metadata attributes provided by the ISO/IEC 11179 standard. To construct an ontology for use with clinical documents, we used metadata in clinical documents obtained from the real-world clinical setting of a tertiary hospital in South Korea. We evaluated the utility of the developed CMDO using 1 personal health record (PHR) specification [CCR Plus (CCR+)] [22, 24], 6 HL7 templates [25], and 25 common clinical documents from 5 teaching hospitals in South Korea [26].

CMDO was developed using one of the appropriate methodologies for conceptual modeling, the General Formal Ontology (GFO) method [27], which is a manual iterative process comprising five steps: (1) defining the scope of CMDO by conceptualizing its first-level terms (or classes), (2) identifying CMDO concepts, (3) assigning hierarchical relationships among CMDO concepts, (4) developing CMDO properties (e.g., synonyms, preferred terms, and definitions) for each CMDO concept, and (5) evaluating the utility of CMDO. All metadata used in our work registered in the CHMR (http://​chmr2.​snubi.​org:​8083/​chmr/​).

We have created CMDO as a CS for CDEs created by increasingly popular CDE-related projects with an emphasis on their application to clinical documentation. The most-popular headings extracted from the clinical documents were assigned to CMDO concepts. A relationship between subordinates and superiors among CMDO concepts was defined manually by clinical informatics experts, while the remaining ontology development process was conducted according to the GFO method.

Despite the ongoing and rapid advances in informatics technology, it is still impossible to fully automate the management of the full semantics of clinical documents and their data elements, which is due to the data elements and their values being semantically too diverse and unscalable. In other words, human resources are still needed to manage the full semantics of clinical documents. CMDO may serve as a suitable CS for facilitating interactions between human resources and machines. We can expect CMDO to be useful as (1) a CS for CDEs for clinical documents and CRFs, and (2) a tool for integrating CDEs from diverse clinical documents and CRFs.

The HL7/LOINC Document Ontology (DO) was developed to provide a standard representation of the attributes of clinical documents using a multiaxis structure and to standardize the names of clinical documents as an essential first step toward the optimal use, management, and exchange of documents both within and between institutions [33, 34]. A CS has been provided for clinical documents, and document names have been standardized. However, clinical documents within the same DO classification may be semantically diverse, and those that are in different DO classifications may have a multitude of CDEs in common. Moreover, the structures of real-world clinical documents are constantly evolving. For example, an admission note and discharge summary may share many CDEs such as the chief complaint, present illness, and past history, making them structurally similar but with very different administrative roles. It is therefore necessary to not only classify clinical documents themselves but also classify the CDEs contained within them in order to comprehensively understand and manage such documents. Clinical documents can be automatically classified and automatically managed based on the similarities between the sets of CDEs that they contain. We reviewed the existing upper level ontologies in an attempt to identify an appropriate one for classifying CDEs in clinical documents. However, this was unsuccessful since the available ontologies were either too complicated and had a high complexity (e.g., UMLS) or were not appropriate (e.g., DO); we therefore constructed CMDO.

CMDO provides precise and comprehensive semantic annotations in terms of UMLS, in that each CMDO concept has properties including synonyms and definitions that are mapped to the UMLS CUI, adopting UMLS preferred terms considering CMDO as a part of UMLS in the clinical domain to classify CDEs. Each CMDO concept has properties to show how it can cover the same or similar meanings for discovering semantic correspondences.

It should be noted that while CMDO appears to be a feasible ontology for annotating CDEs in clinical documents, but it has the limitation of an insufficient concept coverage due to its contents and granularity; our evaluation revealed that 5% of the CMDO items are either too broad or too specific. Our construction, implementation and use of the ontology also revealed two further limitations. First, when constructing CMDO we did not consider other methods and compare them with the GFO method, instead only considering the GFO since we considered it an appropriate method for conceptualizing ontologies. Second, when using CMDO we adopted an approach that we called ‘middle-out,’ meaning that we systematically linked CDEs up to the ontology and down to the vocabularies. It was implicitly assumed that there were associated appropriate CDEs with the correct CMDO concepts; if this assumption was not valid, it would take more time to match CMDO concepts to CDEs.

In future work we plan to extend the current version of CMDO to also cover the parts that were identified in the evaluation as being either too broad or too specific. We will also apply other ontology methods to CMDO to check it for semantic consistency, and consider adding more concept relationships, such as part-of relationships. We also plan to determine how to simplify the process of using CMDO in order to save time and ensure high usability.

The sharing, understanding, and integration of data from multiple different domains can be facilitated by standardization. An MDR-based CDE is considered a type of standardized data with specified concept and VDs. This study has demonstrated that a clinical-content-based ontology can be used to identify standardized CDEs. The rapid expansion of CDEs from many types of clinical documents in numerous studies makes CMDO a useful CS and integration tool.