RWD are defined as routinely collected data relating to a patient's health status or the delivery of healthcare as opposed to data retrieved from traditional clinical trials. The current landscape of RWD comprises a large variety of types: clinical observations, medication, claims, molecular profiling, family history, mobile health, environmental, patient reported and social media [1,2].
RWD can provide a holistic view of people's health by capturing information from sources that reflect the experience, behavior, and social and environmental context of patients collected by health records, claims data, registries, patient-reported outcomes, consumer credit-card spendings, geospatial data, and medical devices-harvested data [3,4].
RWD can make drug discovery more efficient by better identifying unmet medical needs, optimizing the design of clinical studies, informing benefit-risk assessments, and enabling largescale analytics. It can expand the on-label indications for existing products by showing their effectiveness in different populations and/or conditions [4,5].
RWD can be used to make the healthcare systems more sustainable by supporting evidencebased decision making, comparative effectiveness research, health technology assessment, and pharmacovigilance. It can support the approval and market access of new drugs by demonstrating their safety, efficacy, and cost-effectiveness in real-world settings. [6].
However, RWD have specific characteristics, which poses challenges for its integration and findability [4,7]. These challenges arise from the heterogeneity of RWD since it is retrieved from sources that differ by nature and quality, their dynamism and voluminosity as more data are generated and collected over time. Integration of RWD requires harmonization of data formats, standards, and terminologies across sources and domains. Findability of RWD requires effective metadata (i.e., descriptive data that characterize other data creating a clearer understanding of their meaning and suitability for a specific purpose).
While some initiatives have addressed the data integration issue by providing common data models (CDMs) and analytical tools for observational data such as the Observational Medical Outcomes Partnership (OMOP) CDM [8], the Heads of Medicines Agencies (HMA) and the European Medicines Agency (EMA) advance on the findability topic by developing a framework to better integrate RWD into the regulatory decision-making processes [9]. This framework includes a list of metadata for RWD catalogues, a guide of good practices for the use metadata [10,11]. HMA/EMA have also published Data Quality Framework for EU medicines regulation that provides actions, metrics, and a maturity model to support data-driven regulatory decision making [12].
However, it is not enough to have a list of abstract metadata to promote data discoverability and fit-for-purpose assessment. This should be accompanied by an RWD-specific metadata vocabulary. A machine-actionable representation of metadata able to be queried and processed in a standardized fashion would be ideal but requires a model for the metadata vocabulary. The Data Catalog Vocabulary (DCAT) and schema.org can be used as the foundation of such a model [13,14]. For instance, the Google Dataset Search service [15] relies on the structured description of datasets published on the web using both DCAT and schema.org for dataset discovery [16]. These models are domain-agnostic i.e.: they describe cross-domain entities in a high-level fashion and provide mechanisms to extend them for applications in domain-specific settings.
An extension of these models would enable the definition of the semantics and syntactic aspects of the domain-specific metadata elements, as well as the relationships and constraints among them. It would also facilitate the validation and quality assessment of the metadata and allows the mapping and alignment with other metadata standards. A RWD domain-specific extensions is still missing. It would lead to a better description of the RWD assets through standardized metadata models and vocabularies, which would overcome the findability issues.
Schema.org is a structured markup vocabulary to describe resources published on the web by defining types of information and their corresponding properties. It is increasingly being applied to web pages as it improves its findability and enhances the search experience for end users. Schema.org also provides the mechanisms to extend the vocabulary by defining additional types, properties, and constraints [17]. Bioschemas is one of those extensions aimed at improving the exposure of structured data in life sciences by adding data types related to biology and defines a set of minimum metadata properties [18,19].
DCAT is a structured vocabulary developed by W3C and designed to enable a publisher to describe datasets and data services in a catalog using a standard model. It comprises a minimal set of classes, properties, and constraints specified through the Resource Description Framework (RDF) data model [20]. DCAT also reuses terms from other vocabularies as proposed by the Dublin Core Metadata Terms Recommendation [21]. DCAT can be extended by adding constraints as cardinality restrictions, classes and properties, controlled vocabularies, and requirements for access mechanisms. A DCAT extension is known as a DCAT profile. There are profiles for data portals in Europe and geospatial, statistical, and national variations of it [22].
Despite the similarities, schema.org and DCAT show differences in their technical design and the use cases they address. Schema.org is aimed at enhancing discovery and indexing of online resources via search engines while DCAT is meant to represent detailed information on datasets and data catalogues. Therefore, schema.org has a broader scope since it covers a variety of resources like web pages, articles, books, and events. On the technical side, schema.org has a simpler and flatter structure as it defines fewer classes and properties, does not use nested or complex objects, and allows multiple types for the same resource [23].
This paper introduces a DCAT-based metadata profile for an RWD-specific metadata vocabulary. It is aimed to describe the specific details of the variety of types of data and sources comprised in RWD catalogs. It enables the standardized definition of the essential aspects of the RWD datasets such as therapeutic area, patient data, healthcare setting, legal and regulatory, and governance details.
The proposed profile described below extends the core six DCAT classes. A dataset is a collection of published data available for access or download in one or more serializations. Each serialization is represented by a distribution that can be provided via data services. The description of the datasets, distributions, and data services are registered in a catalog. The catalog records are used to describe such a registration by providing, for instance, provenance details. The dataset series allows to represent a collection of datasets that are published separately but share some characteristics that group them.
A catalog compliant with the profile must fulfill the following additional constraints. The catalog must have at least one catalog record. A catalog record must (a) describe the registration of exactly one dataset or data service, (b) have the date of formal recording of the asset, and (c) have the most recent date the catalog entry was updated. A dataset must (a) have an identifier, (b) have a title, (c) have a textual description, and (d) be serialized through at least one distribution.
A set of previously proposed Linked Data modeling patterns were applied to design the profile [24]. The Custom Datatype pattern to further describe the structure of literal values, Label Everything to ensure that every resource has a human-readable name, Link Not Label to ensure the complex entities are modeled as resources instead of labels, Multi Lingual Literal to allow for internationalized text, Qualified Relation to improve the modeling of complex relationships, Repeated Property to allow for multi-valued properties, Topic Relation to states the linkage between content-related resources, and Typed Literal to constrain the properties allowed values.
The profile defines a design pattern to assert the presence or absence of the features on the datasets [11]. This pattern comprises (1) the is described by property to link a dataset with the Qualified Existential class, (2) the Qualified Existential class to asserts the occurrence of a dataset feature through the is data available and is about dataset feature properties, and (3) the Dataset Feature vocabulary of features.
Usage of this pattern provides a mechanism to extend the profile by defining new dataset features on the Dataset Feature vocabulary. This allows us to avoid an explosion in the number of properties that would hamper its maintainability.
The metadata profile also defines additional classes and a variety of properties. The Patient Group class and its descendants -the Patient Gender Group and Patient Age Group classes -allow to describe the population aspect of the datasets.
An overview of the profile core classes, and the additional properties is shown in Figure 1. The diagram should be interpreted under the RDF open-world assumptions about the sub-class-of semantics in contrast to the object-oriented generalization meaning.
This paper introduced a DCAT-based metadata profile for an RWD-specific metadata vocabulary. The model is aimed at tackling the findability-related challenges by enabling querying and processing of metadata in a standardized fashion. It extends the core DCAT classes through the definition of additional constraints, classes, and properties. A design pattern to specify the data elements applicable to the description of the RWD sources is also proposed.
As a proposal in its early stages, it opens several lines of further development. This involves a comprehensive analysis of the metadata vocabulary defined by the HMA/EMA and the assessment of a potential alignment with the proposed model. This article describes work-in-progress for which the authors are looking for collaborations, and kindly ask interested parties to get in touch for further details. Given this scenario, the licensing issue is under discussion. This makes it unfeasible to openly release a fully-fledged RDF serialization of the proposed DCAT
Special thanks go to Michèle Arnoe and Marie Sophie Schwalm (IQVIA) for the useful comments on the overall process of cataloguing and publishing real-world data, and for the feedback on this manuscript.