Epidemiology is a domain of knowledge interconnected with many other domains, thus making it a good candidate for reusing existing ontologies that, despite having been created for different purposes, characterize information frequently manipulated by epidemiologists and public health scientists. This paper presents an evaluation of existing ontologies for the semantic annotation of epidemiological resources. We selected a set of ontologies and proposed a Network of Epidemiology-Related Ontologies (NERO), which can form the core of semantic annotation for data-intensive epidemiology-related information systems, such as epidemic forecasting infrastructures. To support this selection, we defined a set of requirements for inclusion of ontologies in NERO, based on good ontology practice, the interdisciplinary nature of the epidemiological domain and support of semantic web technologies. Most of the selected NERO ontologies are current candidates or members of the Open Biological and Biomedical Ontologies initiative.
Epidemiology is a truly multidisciplinary subject in the sense that it relies on diverse areas of knowledge, such as medicine, statistics, social sciences and geography. As a result, epidemiological data is both one of the most diverse and also one whose characterization can benefit the most from semantic technologies. In fact, only a framework able to understand heterogeneous and multidisciplinary resources can ultimately deal with all aspects of epidemiology. Ontologies can help address this issue by improving the integration, categorization and sharing of epidemiological resources.
Consider the following example: a research team is building a model for influenza spread and they need to know the number of infections caused by this disease over time to fit the parameters of their model. They are interested in building a model that works in France. By means of an appropriate query, they can try to find data resources about “influenza in France.” To effectively find the resources in need using an automated process, these should be correctly annotated in advance. One of the best approaches for creating such annotations is to associate ontological concepts to each resource. The annotations could then be used by a service that would locate the resources satisfying the query and rank them by means of semantic similarity applied to the ontological concepts associated to the query and the annotated resources. Furthermore, by using ontologies, the process of annotating a resource can be facilitated through the identification of concept labels using text mining for analysis of the content of the resource.
To provide the service described in the above scenario, it is paramount to establish a comprehensive and powerful set or inter-related semantic concepts ensuring the consistency of the annotations. By doing so, we stimulate the creation of tools to serve epidemic modelers and epidemiologists in general, thereby contributing to the adoption of semantic technologies in epidemiology.
This paper proposes NERO, a Network of Epidemiology-Related Ontologies, a first step towards the generation of a useful source of epidemiological concepts and the relations among them supporting the characterization of data resources used in epidemiological studies. In this aspect, NERO can be considered an ontology per se, but the disparate set of domains in this area suggests that it should, in fact, be composed of a number of distinct inter-related ontologies. To ensure good interoperability between these ontologies, we propose a list of requirements for inclusion in NERO. A survey to the state-of-the-art in epidemiologically relevant ontologies yielded a set of ontologies that partially fulfill the proposed requirements. 2
The ontologies to be included in NERO were selected based on the
domains of knowledge that are generally present in epidemiological
ontologies (diseases, locations etc.; see Table 4.1 for a full list of
NERO domains). However, since its goal is to serve epidemiological
research, we also based this selection on the needs of one
particular epidemic research platform, the Epidemic Marketplace
(EM)
We started by discussing which requirements should be fulfilled by the ontologies to be included in NERO, in order to create a concise, comprehensive and good quality network of ontological concepts for the epidemiological domain. We then surveyed current ontologies in this domain and evaluated them according to their fulfillment of the requirements. Since it was found that these ontologies generally fail in satisfying many of the requirements, we then considered general-purpose ontologies, like UMLS, and ontologies with a more focused domain, like the Disease Ontology. 3
The ability to reuse concepts from one application in another is
one of the premises of the semantic web
In any comprehensive collection with multiple provenances,
there must be a set of requirements ensuring and enabling both
a good interoperability among those resources and an overall
cohesive structure. Besides the specific requirements derived from
the particular goals of NERO, some of the requirements presented
here include adaptations of (i) principles of the W3C semantic
web
This list should be considered as a set of guidelines, not as absolutely mandatory rules. They lead NERO towards an ideal scenario where ontologies have a high level of expressibility and good interoperability between each other, all the while enabling a simple, yet powerful, implementation of semantic web technologies.
We propose ten requirements: 1. Relevant Domain The most important requirement for incorporating an ontology in NERO is that it should encode a domain of knowledge that is interesting from the point of view of epidemiology, i.e. it must contain concepts that are relevant to annotate epidemiological resources. Likewise, the full network should cover almost all of the epidemiological domain (diseases, modes of transmission, geographical locations, etc.). 2. Appropriate granularity To achieve high domain coverage and improve the semantic characterization of epidemiological resources, an ontology must provide an adequately detailed representation of its domain. Biomedical and geospatial ontologies tend to comply to this requirement quite well, and in fact some contain many thousands or even tens of thousands of concepts spread over many levels of depth, allowing specific annotations such as the exact strain of a virus instead of its family. In contrast, the best ontology to describe a given domain can be too granular for the purpose of epidemiological annotation (e.g. photon in ChEBI). In these cases, NERO should disregard the unwanted branches of the ontology. 3. Expressiveness with tractability A large number of concepts is an important advantage, but being able to manage the ontologies is also a technical requirement, and as such the ontologies must be well structured and tractable from a computational point of view. Specifically, this means that the properties must be formally defined and that these definitions should be adjusted to the domain in question. For example, in an the anatomy domain, it does not make sense to have only an is-a relationship type, since properties like part-of and arterial-supply are equally relevant. 4. Cross-references Ontologies encoding different domains can, nevertheless, be related to one another. For example, symptoms are usually associated with diseases and vice-versa. External references that cross from one ontology to another link together concepts from different domains, enabling semantic web technologies to explore multiple-domain relations. 5. Textual definitions Since NERO will be employed primarily to annotate epidemiological resources, it is important that its users understand the meaning of each concept. Ontologies encode theirs concepts’ meaning in machine-readable code, but for users to correctly identify the concept they want, concepts should be complemented with textual definitions. 6. List of synonyms As synonyms are abundant in natural language, particularly in the biomedical field, it is important that NERO explicitly states these synonyms (e.g., “AIDS” is a synonym of “acquired immune deficiency syndrome” and users expect that both refer to the same concept). 7. Popularity NERO ontologies should be well known in the epidemiology community, since familiarity increases the chance that users more easily choose the correct concept. Furthermore, if an ontology is popular, there is a higher probability that its development does not stall in the foreseeable future, ensuring that NERO is kept updated with the most current knowledge. 8. Publicly available It is imperative to adopt an open-source ontology rather than one that needs licenses or other form of control over usage. This requirement stems from the fact that we intend NERO to be publicly available to anyone in order to further increase the spread of semantic web technologies into epidemiology. Also important, in this context, is that users have the opportunity to submit corrections, suggestions and other improvements to the ontologies. 9. Persistent identifiers Since ontologies constantly change in response to advances in the field, errors found, etc., some concepts may change their definition, ultimately resolving in some annotations becoming wrong. To mitigate this issue, ontologies in NERO should have semantic-free identifiers which are never removed from the ontology. 10. Distributed access to the ontology Several languages have been developed to encode ontologies (most notably OWL and OBO format); but other formats exist, from simple tree-like structures described in a text document to tables on a database. Instead of having to cope with all these differences, NERO ontologies should be easily accessible through “the cloud”, (web services or equivalent). Moreover, by not having a local copy of the ontology, there is no need to take special actions in order to keep it up-to-date.
The first three requirements (domain, granularity and expressiveness) are scope-related, since they refer specifically to the knowledge encoded in the ontology itself. The others are properties that simplify the tractability of an ontology and improve its usefulness as a scientifically sound source of concepts for annotating epidemiological resources, while ensuring a certain degree of user-friendliness, which is important given that one of the aims of NERO is to facilitate the annotation process. 4
This section presents a survey of the state-of-the-art in ontological representation of the epidemiological domain and an overview of currently existing ontologies that, despite having been created for other purposes, can be used to describe concepts relevant for this field of research, such as diseases, modes of transmission, demographics or geography. A summary of the considered ontologies and the domains they represent is given in Table 4.1.
As we are interested in covering the whole spectrum of epidemiological domains, we have not defined a minimum number of requirements to be fulfilled by NERO ontologies. In fact, not all requirements are equally important. Instead, NERO includes the most well-adjusted ontology for each domain, and for the topics where we could not find ontologies of good quality, it fills the gap with controlled vocabularies, such as taxonomies and dictionaries. It is expectable that relevant ontologies will be developed in the future, and given the modular nature of NERO, these could, when available, easily replace the lower quality terminologies. 4.1
Ontologies specific to the epidemiological domain
There have been two attempts at organizing epidemiological
terminologies in a hierarchical manner
There has also been a number of automatic systems designed
to monitor epidemic surges. One example is the BioCaster Global
Health Monitor
Another ontology built especially for epidemiological studies is
the Epidemiology Ontology, developed by HuGE NET
Other ontologies containing epidemiological concepts Given the low suitability of those resources for inclusion in NERO, our focus moved to ontologies not built for epidemiology but which, nonetheless, contain relevant epidemiological concepts. The relevant domains were assumed to be the ones in need for the EM metadata model, which agree with the domains of knowledge represented in BioCaster, the Epidemiology ontology and the Dictionary of Epidemiology, namely: demography, diagnostic methods and other clinical methods, diseases, drugs, environment, geographical location, socio-economic conditions, symptoms, taxonomy, transmission modes and vaccination.
Some research has been conducted in epidemiology based
on the use of existing ontologies containing epidemiologically
relevant concepts, but which were not designed with that
specific domain of knowledge in mind. This category includes
the Unified Medical Language System (UMLS), a “collection
of ontologies and terminologies that promote the creation of
more effective and interoperable biomedical information systems
and services”
As an example, the work of
Additionally, MeSH is relatively unstructured and makes use of a single relation (just like the Epidemiology Ontology presented above). For example, Axial length and Eyebrow are categorized under Eye, but one is a property and the other is a nearby structure. Likewise, Eye is both categorized under Sense Organs and Face, but while it is a sense organ, it is part of the face. MeSH makes no distinction between these semantic relations, which we consider one of the main drivers for the use of ontologies.
There are other limitations with UMLS and MeSH: since they
have a generic and broad domain, the addition of new concepts
is non-trivial, and there is a high risk of introducing errors and
inconsistencies. In fact, it is known that UMLS houses many
inconsistencies
10 In NERO? Y N N Y N Y Y Y Y Y Y Y N N Y Y N
Y N Y Y Y Y N Y Y Y Y Y N Y Y Y Y
Y Y Y Y Y Y Y Y Y Y Y Y N N Y Y N N
N N N Y Y Y Y Y N Y Y N Y Y Y Y Y N N
Y Y Y Y N Y Y Y Y Y Y Y Y Y Y Y Y 9 N N N Y Y N Y Y Y Y Y Y Y Y Y Y Y Y Y
N N N N N N Y Y Y Y Y Y Y Y Y Y Y Y Y
No Yes No No Yes No Yes No No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Terminology published in a standard semantic web format, meaning that they do not integrate well with semantic web technologies. Moreover, UMLS needs a license to be used.
SNOMED-CT R , a subset of UMLS, was also considered
individually, since it is a comprehensive terminology for diseases
and other clinical terms
In face of these issues, we turned to attempts to create and
organize more formal ontologies in the biomedical field. OBO
should be highlighted, since it aims at providing a suite of
interoperable reference ontologies in the biomedical domain
Non-biological concepts must be retrieved from other resources.
Yahoo! GeoPlanetTM
We have been unable to find ontologies that specifically represent
demography or social and economic conditions, and suspect that
none exist that are publicly available. As such, we will have to
rely on resources such as MeSH, the Epidemiology Ontology and
NCI Thesaurus for those domains. In this context, it is important to
mention that resources like the Dictionary of Epidemiology can be
curated in an ontological format (e.g., through triplification
A graphical summary of this survey is shown on Table 4.1, which details the requirements fulfilled by each of the ontologies found, and which of them were included in NERO. 5
This paper proposes the Network of Epidemiology-Related Ontologies (NERO), which aims at providing a core of relevant concepts to semantically characterize epidemiological resources and therefore enable the application of semantic technologies in the epidemiological domain. NERO is being developed as part of the Epidemic Marketplace platform (EM) to annotate its resources, but can be further used by any epidemiologist or public health scientist.
The choice of a network of ontologies was motivated by the fact that existing epidemiology ontologies (BioCaster and the Epidemiology Ontology) are not comprehensive enough to suitably model the epidemiology domain. Nevertheless, they offer an insight into what an epidemiological network of ontologies should cover and how it should be organized, since they contain the branches of knowledge required in epidemiology (diseases, modes of transmission, locations, social conditions, etc.). By crossing this information with the needs of the EM, we concluded that selected ontologies from the OBO project complemented with branches of MeSH and the Epidemiology Ontology are the most appropriate terminologies for the biomedical portion of epidemiology, since together they span over a large amount of the biomedical domain on this area. Other resources, such as UMLS or SNOMED-CT R , are not as suitable due to their licensing, complexity, and difficulty in integrating into a fully semantic web approach.
For geographical information, we included the geographical ontology of Yahoo! GeoPlanetTM, based on its higher quality when compared to the other candidates. For the domains of demography and social and economical conditions, the lack of any specific ontologies prompted us to consider branches of MeSH and NCI Thesaurus. Additionally, the Epidemiology Ontology has also concepts in these domains.
Once epidemiological resources are annotated with NERO, it will be possible to exploit these annotations to perform complex semantic analysis on diverse data processing tasks tasks, such as information retrieval, integration and extraction. These tasks will provide epidemiologists, particularly epidemiology modelers, with tools that enable an easy discovery of models and the parameters to use in them. The Epidemic Marketplace is currently being developed with the intention of serving as a starting point for this semantic analysis.
Finally, it is worth noticing that, at the moment, NERO is little more than a list of requirements and the identified set of ontologies. As future work, we plan to completely integrate these ontologies in a more tangible network. For instance, semi-automatic alignments between the ontologies should produce pairs of equivalent concepts that can be merged. The mapping of these ontologies into upperdomain ones (such as BFO) should also contribute to a better integration of all the ontologies. Eventually, we plan to introduce other semantic web technologies that will make use of NERO, such as RDF triple stores and SPARQL endpoints.
The authors want to thank the European Commission for the financial support of the EPIWORK project under the Seventh Framework Programme (Grant #231807), and the Portuguese Fundac¸a˜o para a Cieˆncia e Tecnologia through the financial support of the SOMER project (PTDC/EIA-EIA/119119/2010), the PhD grants SFRH/BD/42481/2007 and SFRH/BD/69345/2010, and the PIDDAC Program funds (INESCID multi annual funding) and through the LASIGE multi annual support.