-Knowledge of abnormalities is important for understanding diseases. A number of resources provide terms related to abnormalities in biomedicine. In this paper, we investigate the differences in the hierarchical structure of these biomedical resources and discuss issues of reuse and integration with regard to abnormal states based on ontological theory. Then, we show a solution for integrating them by linking abnormal states in our abnormality ontology to other resources according to the meaning of the concepts at each level.
INTRODUCTION
In order to understand diseases, one must first capture the
abnormal states in diseases adequately. They are observed as
symptoms by patients or as signs by clinicians in clinical
findings. Clinical test data can also provide evidence for the
existence of abnormal states. In addition, in basic research, through
analysis of disease models of animals, many researchers make
efforts to understand how causative agents are related in the
etiological process. Moreover, abundant knowledge about
abnormalities is available in scientific articles. The above
observation shows that abnormality is a key factor for capturing
diseases, assisting in the interoperability between basic
research and clinical medicine for integrating a wide variety of
knowledge across domains. We have been involved in the
development of a disease ontology [
A number of terminologies and standard vocabularies have been developed for many years, and recently, ontologies have also been constructed in the biomedical domain. They offer useful data, and some of these terms include abnormality concepts. In order to make efficient use of these resources, we need a solution that ensures interoperability between abnormality knowledge across domains. To achieve this, first, one has to elucidate one's own perspectives of the resources before using them and to make the meanings of concepts explicit. In this paper, we discuss differences of existing biomedical resources based on ontological engineering theory with respect to abnormalities. Next, we investigate the relationships between abnormal states in our abnormality ontology and corresponding terms in biomedical resources. By mapping concepts from our abnormality ontology to other biomedical concepts, we can establish mutually complementary relationships between biomedical concepts in different levels, which will contribute to ensuring interoperability across biomedical resources. Our goal is to provide not only a theory for a better understanding of abnormal states but also useful information for clinical practice. We are planning to integrate abnormalities in the definition of diseases in the Department of Cardiovascular Medicine and other medical departments at the University of Tokyo Hospital in our ontology with external biomedical resources at each level of meaning as a concrete example.
This paper is organized as follows. In Section 2, we introduce our ontology of abnormalities. In Section 3, we examine the characteristics of biomedical resources, discuss the problem with them based on ontological theory and propose a solution for integration with respect to abnormal states. Finally, in Section 4, we present concluding remarks and a give an outline of future work.
II.
In order to develop an is-a hierarchical tree of
abnormalities, it is important to conceptualize them from a consistent
viewpoint. To this end, we have been developing an
abnormality ontology [
The top-level categories define very basic and generic concepts, for example, "small in area," "hypofunction," etc., which are commonly used not only in clinical medicine but also in other domains. Level 2 concepts are dependent on objects. In the lower level of the tree, concepts are designed to represent abnormalities at specific human organ / tissue / cell levels. For example, by specializing "small in area" at Level 1, " narrowed cross-sectional area of tube ", where the cross-sectional area of a tubular structure has become narrowed, is defined at Level 2, and this is further specialized in the definitions "vascular stenosis" (blood vessel-dependent), "arterial stenosis", "coronary artery stenosis" (coronary artery-dependent), and so on. Level 3 concepts are captured as specific disease-dependent (contextdependent) abnormal states. For example, "coronary artery stenosis" at level 2 is defined as a constituent of ischemic heart disease at Level 3. In our ontological approach, common concepts can be kept distinct from specific ones and can be appropriately defined according to their context.
As of 11 May 2013, our ontology has 21,669 abnormal states constituted of 6,302 diseases across 13 medical departments, and among them, clinicians have currently refined concepts of 9,985 abnormal states constituted of 1,602 diseases across five major departments.
In medicine, abnormal states are interpreted from the
diverse perspectives of specialists such as clinicians, pathologists,
biologists, geneticists, and so on, and correspondingly a variety
of representations of abnormal states are used. Therefore, we
classified the abnormal states into three categories: a property1
(e.g., hypertension), a qualitative representation (e.g., blood
pressure is high), and a quantitative representation (e.g., blood
pressure 180 mmHg). In previous work, we proposed a
Property-Attribute interoperable representation framework for
abnormal states [
We captured all abnormal states as properties2 represented by a tuple: <Property (P), Property Value (Vp)>, e.g., <stenosis, true>. We specified the property by decomposing it into a tuple: <Attribute (A), Attribute Value (V)>. The Attribute Value can be either a Qualitative Value (Vql) or a Quantitative Value (Vqt). For example, "arterial stenosis" is decomposed into <cross-sectional area (A), small (Vql)> as a qualitative representation, or <cross-sectional area (A), 5mm2 (Vql)> as a quantitative representation. Then, we introduce "Object" to identify the target object, and we represent an abnormal state as a triple: <Object (O), Attribute (A), Attribute Value (V)>. This is the basic form in our representation model of abnormalities. In addition, we introduce "Sub-Object (SO)" as an advanced representation for what will be focused on. For example, in the case of "hyperglycemia", since the glucose concentration (A) means the ratio of the focused object (SO) relative to the whole mixture (O), the representation of "hyperglycemia" is a quadruple, <blood (O), glucose (SO), concentration (A), high (V)>. In another case of an advanced representation, "Colonic polyposis" is described as <colon (O), polyp (SO), number (A), many (V)>. Our model can deal with both clinical test data and abnormal states in the definition of diseases. The clinical test data can be represented in the form <Object (O), Attribute (A), Quantitative Value (Vqt)> (OAVqt), which can be converted into a property representation form <Object (O), Property (A), Property Value (Vp)> (OPVp) via a qualitative representation 1 The property discussed here is based on ontological theory, not Web Ontology Language (OWL) property which means a link (relation) between two nodes. 2 A state derived from or associated with a property (P) is defined as: "a temporal entity derived by a time-indexed property (P) in which the bearer participates". A property has a bearer, while a state has the bearer as a participant. form. For example, in terms of the state of hypertension3, our model ensures interoperability among the forms <blood (O), pressure (A), 180mmHg (Vqt)>, <blood pressure, high>, and <hypertension, true>. Therefore, our model realizes interoperability between test data and abnormal states in the definition of diseases. In practice, we found that more complicated terms, such as adding modifier words (e.g., transient hypertension), and three types of compound words (e.g., “blood pressure” as an OA type: < blood (O), pressure (A)>). We have developed a guideline and deal with them as a variation of data representation.
III.
In this section, we investigate the hierarchy of "coronary artery stenosis" in existing biomedical resources and perform a comparison between our ontology and existing biomedical resources to make them interoperable with each other with respect to abnormalities.
a) SNOMED-CT: Systematized Nomenclature of
Medicine-Clinical Terms (SNOMED-CT) is a clinical terminology
first developed by the College of American Pathologists
(CAPs) and is currently maintained by the International Health
Terminology Standards Development Organization (IHTSDO)
[
b) MeSH: Medical Subject Headings (MeSH) is a
thesaurus of medical terms developed by the National Library of
Medicine and is used for indexing biomedical articles [
2) PATO and HPO Phenotype Ontologies: In biomedical domains, scientists observe entities through experiments or clinical findings to capture abnormal states. Therefore, repre3 Our model can describe the mechanism “hypertension” more specificity by using other factors; e.g., accumulation of atheroma in arterial wall <arterial wall (O), atheroma (SO), quantity (A), much (V) >→decreased elasticity of arterial wall <arterial wall (O), elasticity (A), low (V) >→hypertension. sentations of knowledge about abnormalities are usually given as phenotypes.
Phenotypic Quality Ontology (PATO) is an ontology of
phenotypic qualities for annotating biological phenotypes
across species [
The Human Phenotype Ontology (HPO) is an ontology for
human specific phenotypes [
3) LOINC: Logical Observation Identifiers,
Name and Codes (LOINC) was developed by the
Regenstrief Institute as a universal code for
exchanging clinical test data [
1) PATO and HPO: Both PATO and HPO representations
are based on property representations in accordance with the
upper ontology Basic Formal Ontology (BFO) [
The next issue is that PATO does not differentiate Property (P) from Attribute (A) in the same way as BFO does, which results in multiple inheritance from two parents in the PATO hierarchical tree: one is abnormal states, and the other is property (quality). For example, "decreased area" in PATO has two super-classes, "decreased size" and "area" (Fig. 1). Such multiple inheritance makes things complicated, and it is difficult to understand the underlying meaning of concepts. If computers integrate PATO concepts and other resources in a naïve way, inappropriate consequences might be derived unexpectedly. For example, HPO provides the concept of "coronary artery stenosis," and an incorrect relation such as "coronary stenosis is-a area" might be derived by a naïve combination of PATO and HPO. "Area" is not an abnormal state. In our ontology, parameters such as pressure, area, concentration, and so on, are properly dealt with as an attribute (A), so that there is no possibility of deriving inappropriate consequences such as the one shown above.
With respect to the intrinsic nature of abnormal states, states can change, and their essential characteristic is understood with respect to how the corresponding attribute (A) has the value (V), which can change as time goes. Then, one can find that the authentic "is-a" relation should be "decreased area is-a decreased size" rather than "decreased area is-a area", and our model adopts the former.
Next, we discuss ontological issues in HPO. The HPO hierarchy is based primarily on where the abnormal states occur (e.g., arterial stenosis is-a abnormalities of peripheral arteries). Therefore, it is difficult to know the is-a relation of "stenosis" comprehensively. In the HPO tree, ad hoc creation of correct is-a relationships is found, such as "renal artery stenosis is-a arterial stenosis." To make matters worse, HPO has no generic concepts at the upper level nor any information about what attribute takes what value, and we cannot capture the essential property of the state as "stenosis." Furthermore, specialization using the part-whole relation appears in the is-a hierarchy, e.g., <abnormality of the vasculature is-a abnormality of the cardiovascular system>, which is misleading.
With respect to abnormal states, it is important to capture the identity of the state by revealing "where and how its intrinsic attribute takes a particular abnormal value." On the basis of such understanding, our ontology adopts a single is-a relationship to inherit the intrinsic nature of abnormal states. For example, all lower classes of "small in area" inherit the property of "the area (A) is small (V)," such as" narrowed crosssectional area of tube," "vascular stenosis," and so on.
In practice, however, only a single hierarchical structure
seems incompatible with various perspectives in biomedicine.
In order to support different perspectives, we developed a
technology for dynamically generating an on-demand
classification hierarchy [
Fig. 2 Reorganization of is-a relationships of abnormal states in terms of anatomical structure.
Another critical issue with HPO is that HPO does not differentiate abnormal states from diseases and organize them in one hierarchy, which may be misleading and confusing in clinical practice. In HPO, the parent of "coronary stenosis" is considered to be not an abnormal state but a disease, namely, "coronary disease." However, as shown in Fig. 1, "coronary disease" has three parents: "arterial stenosis," "coronary artery abnormality", and "atherosclerosis," which seem to be abnormal states. Therefore, there is no guarantee that a computer can derive the correct answer to the question of whether "coronary artery stenosis" is a disease or an abnormal state.
Our ontology enables us to make a distinction between generic abnormal states, object-dependent ones, and diseasespecific ones with a unified representation and allows us to specialize concepts to the required granularity with consistency. We conceptualize a disease as an entity represented in terms of abnormal states, and we deal with abnormal states and diseases as different entities. Therefore, we can focus on the intrinsic nature of the states themselves from one viewpoint and can develop an ontology from the viewpoint of state, without mixing up the viewpoint of disease.
SNOMED-CT and MeSH started before ontological
engineering had become mature, and thus they have some problems
ontologically [
SNOMED-CT terms have been used for electronic health record (EHR) systems. Because SNOMED-CT does not differentiate abnormal states from diseases, serious clinical problems may occur. For example, imagine a case where a clinician examines an angiographic image and, by using the SNOMED-CT term "coronary artery stenosis," records it for evidence of "abnormal states" in the EHR system. Due to inheriting properties from one of the upper concepts in SNOMED-CT, namely, "disease," there is a possibility of deriving an erroneous consequence that the "abnormal state" is a "disease." Our ontology makes a distinction between diseases, disease-dependent abnormal states, and diseaseindependent abnormal states. Therefore, there is no possibility of deriving inappropriate consequences, which is important for computer processing. In order to manage high-level clinical knowledge in EHR, we need a reliable method for representation, and our ontological model provides us with high reliability and sophisticated technology for realizing interoperability between heterogeneous pieces of clinical information.
3) LOINC: LOINC was excluded from the abovementioned comparative research, because LOINC does not deal with concepts related to abnormal states.
LOINC provides the form O (So) A and is useful for interoperability among various clinical test data. However, LOINC does not have Value (V). To realize interoperability between clinical test data and abnormal states, a Quantitative Value (Vqt) is needed in the representation form. Our model can deal with quantitative data in the OAV form, and, therefore, we can transform it into the OP form of abnormal states. As a result, our model has the ability to maintain interoperability between clinical test data and abnormal states in diseases.
Some readers might think, "Why not reuse other resources such as clinical terminology or existing ontologies?", because it seems that by combining existing resources, many abnormal states would be easily covered. Such good candidates for reuse concerning abnormal states are PATO, HPO, LOINC, and SNOMED-CT. As we illustrated in the problem above, just reutilizing existing resources cannot integrate all knowledge about abnormal states. It is too difficult to find all inappropriate usages of is-a and resulting misclassifications from the huge and heterogeneous system of concepts and to make an effort to modify them. Furthermore, since each resource has its own viewpoint, integrating them into a unified perspective must be a hard task because it necessarily requires comparative analysis and validation of accuracy in integrated concepts.
Unified theoretical considerations have resulted in interoperability between various representation forms, which will enable us to establish a computer-understandable model for abnormal states. We need more sophisticated organization of related representations, including quantitative and qualitative data and knowledge at higher levels of abstraction about abnormal states in the definition of diseases to exploit all of them in a consistent manner. Our model is the first one to make such exploitation possible, and will be of great assistance in medical practicee.
The differences of resources are summarized in Table 1.
Mapping our ontology to other resources for integrating
various data related to abnormalities will bring benefits to the
users of other resources, too. First, one can find concepts from
generic to specialized terms easily by referring to the single
isa tree in our abnormality ontology. For example, although
HPO does not care about consistent is-a relationships in terms
of "stenosis", by referring to "arterial stenosis" at Level 2 in our
ontology through mapping, HPO users can get the is-a
relationships: "arterial stenosis is-a vascular stenosis is-a narrowed
cross-sectional area of tube is-a small in area." Since "small
area" is linked to a PATO concept, via our ontology, users
might find orthologous concepts of other species. Specifically,
human phenotypes can be linked to the phenotypes of model
organisms, e.g., mouse, rat, etc., if the set composed of
Attribute (A) and Value (V) are identical, and the Object (O) has
structural similarity. PATO2YAMATO aims to integrate
phenotype descriptions residing in differently structured
comparison contexts [
As illustrated in Section 2, our ontology provides three levels of abnormal states from generic to disease-specific.
Level 1 in our ontology defines generic concepts
corresponding to the PATO concepts, and our Level 1 concepts can
be mapped to related PATO concepts (Fig. 3). The lower Level
2 concepts are human anatomical structure-dependent
abnormal states, which correspond to the HPO concepts. By creating
links between Level 2 concepts and HPO concepts, it will be
possible to navigate from the HPO concepts to the upper
generic concepts of PATO. Level 3 provides disease-specific
abnormal states, such as "myocardial ischemia in ischemic heart
disease," "chest pain in angina pectoris", and so on. In the
revised version 11 of the International Classification of Diseases
(ICD), diseases contain information of "causal properties" [
We are also planning to link the components of the <Object
(O), Attribute (A), Value (V)> representation of abnormal
states to other resources. We will try to connect the Object (O)
to concepts in FMA (The Foundational Model of Anatomy
ontology) [
By mapping MeSH terms, it will be possible to retrieve biomedical articles related to abnormal states or diseases. If we obtain new findings of the constituents of diseases, we can add new relationships to the causal chains of the diseases, which might contribute to the elucidation of the etiological mechanism. Our mapping is also useful for MeSH users to understand how their research subjects are involved in various abnormal states in the human body in diseases. This will contribute to the progress of biomedical research.
A large volume of data and concepts related to abnormal states is currently available in existing resources. However, there are no resources that cover all levels of abnormal states from generic to disease-specific. We performed a comparison between our ontology and existing resources, and identified the issue that the heterogeneous levels of meanings in the different resources and multiple perspectives prevented us from reusing and integrating them. This motivated us to develop an abnormality ontology from the generic level to the disease-specific level.
Our medical ontology project started seven years ago. Since then, it has been refined and revised several times by discussion with both ontologists and clinicians. Our ontology will play a role in proper guidelines for giving an ontological point of view in various controlled vocabularies, and will lead to the development of a consistent hierarchical structure with a unified representation. Mapping our ontology with other resources at each level of meanings will contribute to ensuring interoperability across biomedical resources. Since our approach allows users to navigate from generic concepts to specific concepts in other domains by following links, it offers complementary information. We are currently applying the concepts of abnormal states in the definition of diseases in the Department of Cardiovascular Medicine and several other departments at The University of Tokyo Hospital in our ontology and mapping them to external biomedical resources, and this work is expected to be completed in the near future. Level 1 generic concepts have previously been developed by ontologists, and Level 3 disease-specific concepts were also described. Currently, we are developing and enriching Level 2 concepts to link each Level 3 concept to upper-level common concepts. By developing Level 2, we will be able to find more commonalities across diseases. For example, in cardiovascular medicine, "increased blood creatine kinase (CK) concentration in acute myocardial infarction" is defined by a clinician at Level 3. Next, "increased blood CK" at Level 2 is defined and mapped to "elevated serum creatine phosphokinase" in HPO. Then, the generic concept "increased concentration" is defined at Level 1 in our ontology, which is mapped to "increased concentration" in PATO. We can find commonalities with other concepts in the definition of diseases in other medical departments in our ontology. For example, "increased blood CK concentration in muscular dystrophy" used in the Neurology Department has commonality with "increased blood CK concentration," and moreover, "increased blood cholesterol concentration in hyperlipidemia" has commonality with the same generic concept, "increased concentration." We are planning to map the disease "acute myocardial infarction" in our disease ontology to "acute myocardial infarction" in the ICD. In our ontology, diseases are defined as causal relationships of abnormal states, and from this mapping, ICD users will be able to know the causal relationship, "decreased blood flow" causes "myocardial ischemia," which results in "myocardial necrosis" in acute myocardial infarction. This causal relationship is also available to users of HPO by following the linked concept of abnormal states. With respect to "myocardial ischemia," by mapping this to the MeSH term "myocardial ischemia," we can collect articles that are related to the concept.
In this way, our approach allows users to navigate various
abnormality knowledge across domains, which will create a
bridge between basic research and clinical medicine. We
published the causal chains in our disease ontology in the form of
Linked Data (LD)[
A part of this research is supported by the Japan Society for the Promotion of Science (JSPS) through its “FIRST Program”" and the Ministry of Health, Labour and Welfare, Japan. The authors are deeply grateful to Drs. Ryota Sakurai, Natsuko Ohtomo, Aki Hayashi, Takayoshi Matsumura, Satomi Terada, Kayo Waki, and other, at The University of Tokyo Hospital for describing disease ontology and assisting us with their broad clinical knowledge. We also would like to thank other team members, Drs. Yoshimasa Kawazoe, Masayuki Kajino, and Emiko Shinohara from The University of Tokyo for useful discussions related to biomedicine.