DEVELOPMENT OF FUNDAMENTAL TECHNOLOGIES FOR
BETTER UNDERSTANDING OF CLINICAL MEDICAL
ONTOLOGIES
Hiroko Kou, Mamoru Ohta, Jun Zhou, Kouji Kozaki, Riichiro Mizoguchi
The Institute of Scientific and Industrial Research, Osaka University, 8-1 Mihogaoka, Ibaraki, Osaka, 567-0047 Japan
Takeshi Imai, Kazuhiko Ohe
Department of Medical Informatics, Graduate School of Medicine, Tokyo University
73-1m Hongo, Bunkyo-ku, Tokyo, 113-8655, Japan
Keywords: Medical ontology, Dynamic generation of is-a hierarchy, Natural language explanation.
Abstract: We have been building a Japanese medical ontology that provides an intelligent infrastructure for
systematization and synthetic understanding of medical knowledge on a large scale. The objectives of our
research include building a medical ontology and developing application systems based on it. We identified
a few common fundamental technologies for understanding the medical ontology and implemented them.
The main features of these technologies are summarized as the following two functions: dynamically
generating is-a hierarchy according to the user’s interest, and providing natural language explanations. We
built a prototype medical information service system using these fundamental technologies. We conducted
an informal evaluation in a workshop and received favorable comments from medical experts.
1 INTRODUCTION
A lot of medical data have been computerized to
improve the quality of medical services. However,
most of them are stored in different formats
depending on their domains and database
management systems. In addition, the government
has set IT-based structural reform of the healthcare
system as the top IT strategic focus to develop more
advanced medical information systems. Here we pay
special attention to ontology studies, because a
medical ontology would be a core technology for
various applications, such as electronic medical
records, and diagnostic support systems. Medical
ontology research is a key to the successful
development of various knowledge processing
applications that are interoperable with one another.
Although some medical ontologies and standard
vocabularies such as MeSH (MeSH), ICD-10 (ICD-
10), SNOMED-CT (SNOMED-CT), Galen
(GALEN), and FMA (FMA, Rosse, C. 2003) have
been developed, most of them are based on legacy
system terminologies, and some have quite a few
ontological problems. For example, Stefan Schulz et
al. (Stefan, S. 2007) point out some ontological
problems in SNOMED-CT, such as confusion
between the subclass-of (is-a) relation and instance-
of relation, multiple inheritance without considering
the inheritance of an attribute, and others. In
addition, these ontologies are not suitable for
Japanese medical practice due to many culture-
specific differences between Japan and Western
countries.
We have developed a medical ontology suitable
for Japanese medical practice (Mizoguchi, R. 2009).
This ontology is to be translated into English after
enough data is gathered, and will be mapped with
other standard ontologies for interoperability. This
mapping will evaluate the standard terminologies in
line with fundamental ontology engineering and
uncover culture-specific differences between Japan
and Western countries. In addition to building a
medical ontology, the objectives of our research
include constructing and developing application
systems based on it. The underlying philosophy is to
“Learn from the ontology” before using it for
building another application because ontology itself
is a rich source of knowledge about the domain. We
235
Kou H., Ohta M., Zhou J., Kozaki K., Mizoguchi R., Imai T. and Ohe K..
DEVELOPMENT OF FUNDAMENTAL TECHNOLOGIES FOR BETTER UNDERSTANDING OF CLINICAL MEDICAL ONTOLOGIES.
DOI: 10.5220/0003089102350240
In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2010), pages 235-240
ISBN: 978-989-8425-29-4
Copyright
c
2010 SCITEPRESS (Science and Technology Publications, Lda.)
identified a few common fundamental technologies
for understanding the medical ontology, such as a
method for navigation for ontology content
exploration and explanation generation, and we
implemented them. Also, we built a prototype
medical information service system using these
fundamental technologies.
This paper is organized into four sections,
including this Introduction. In the next section,
Section 2, we summarize our major principles for
building our medical ontology. In Section 3, we
discuss fundamental technologies for understanding
clinical medical ontologies and introduce the main
functional modules for dynamically generating is-a
hierarchies. Finally, we present concluding remarks
and discuss future work.
2 DEVELOPMENT OF MEDICAL
ONTOLOGY
2.1 Outline of our Medical Ontology
Our major principles for building a medical ontology
are that we clearly distinguish between context-
dependent concepts from context-independent
concepts, and that we describe all types of diseases
in a common framework. In the framework, diseases
are described as combination of abnormal states, and
all types of abnormal states of the human body
described in the same framework, which is
represented in terms of a triple <object, attribute,
attribute’s value> (Figure 1(a)). For example, “a
high level of HbA1c in the blood” is described as
<blood, HbA1c, high>. Also, we explicated the
specification of causal chains, which are expressed
clearly by sets of causal relationships between
abnormal states and their causes (Mizoguchi, R.
2009). This ontology is described using Hozo which
can describe roles correctly based on the role theory
and can export ontology in OWL (Kozaki, K. 2002).
For example, diabetes is described as combinations
of abnormal states, such as diabetical
hyperglycemia, dry mouth, excessive urination, and
diabetic glycosuria (Figure 1(b)). These abnormal
states play roles such as symptoms and main
pathology of the disease. A role is defined as a
dependent entity played by another entity in a
context. An entity playing a role in a specific context
is called a role holder. By a class constraint, we
mean a constraint on the class to which an instance
playing the role belongs. (Mizoguchi, R. 2007)
To evaluate whether it is possible to describe
most clinical observations in clinical practice by our
conceptual description framework using abnormal
states, we verified a number of physical findings.
We used 3465 data of physical findings (which is a
standard vocabulary used when symptoms and
opinions were described in a care card) that have
been opened to the public by The Medical
Information System Development Center in Japan
(MEDIS-DC). We confirmed that our conceptual
description framework is suitable for most clinical
observations. Similarly, we verified the adequacy of
the conceptual description frame of diseases by
describing 50 disease data that were extracted to
cover each area from 12,000 disease data of ICD-10
that WHO has made public. After validating the
description framework, we created a template for
describing knowledge about each disease and
developed dedicated input software to record data,
and data on 6051 diseases are currently being
collected from 12 hospital departments. Our medical
ontology imported the collected data, and as the
result it consists about 15,000 concepts and 60,000
slots at present. The person in charge of each
medical area plans to continue the description and
data collection in the future.
2.2 Comparison with Other Medical
Ontologies
In SNOMED-CT, the disease is one of the 19 top-
level categories. Therefore, there is no super-concept
of the disease. SNOMED-CT’s 19 top-level
categories preserve the legacy of the former
SNOMED axes, which do not easily agree with any
formal upper level ontology. This is in fact
acknowledged as one of SNOMED-CT’s problems.
(SNOMED-CT, Stefan, S. 2007, SNOMED Clinical
Terms User Guide). On the other hand, we used
Figure 1: Conceptual description and diabetes concept
description as example on ontology.
KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development
236
top-level ontology of YAMATO (YAMATO),
which is compatible with DOLOCE (DOLOCE),
and can be viewed as a high-level ontology based on
philosophical considerations.
In general, concepts are related to one or more
other concepts. In fact, SNOMED-CT allows
multiple “is-a” relationships for describing that
complicate classification without providing any
information regarding a viewpoint for systematizing
hierarchical classification. We adopt the approach of
accepting an essential hierarchical classification and
switching between the hierarchical classifications. It
is then necessary to develop fundamental
technologies that gather related information from the
ontology and dynamically generate an is-a hierarchy
using the information. This function of dynamic
generation of is-a hierarchy has been realized first in
the world.
3 FUNDAMENTAL TECHNOLOG
IES FOR UNDERSTANDING ME
DICAL ONTOLOGIES
3.1 Applications based on Medical
Ontologies
A medical ontology would be a core technology for
various applications, such as electronic medical
records, diagnostic support systems, medical
electronic dictionary and electronic textbook,
medical knowledge navigation, portal site of medical
information, and so on. The underlying philosophy
on application systems in our research is that “Learn
from the ontology” before using it for building
applications since ontology is a rich knowledge
source about the domain. We identified a few
common fundamental technologies for
understanding the medical ontology, such as a
navigation method for ontology content exploration
and content management for explanation generation,
and we implemented them. The main features of
these technologies are summarized as the following
two functions: dynamically generating an is-a
hierarchy according to user’s viewpoints, and
providing natural language explanation. This paper
concentrates on the dynamic is-a hierarchy
generation function for navigation of ontology
content.
3.2 Navigation of Ontology Content
By navigation of ontology content, we mean to
guide users to explore ontology content as a rich
source of knowledge about the domain. Navigation
technology supports efficient access to concepts
defined in medical ontologies. It is very important
because locating targeted medical knowledge is
otherwise difficult in the vast amount of medical
information available. Navigation includes
technologies such as providing indexes for access,
linking to related information, and retrieval
functions. We focus on indexing based on a medical
ontology because it provides the basic infrastructure
for navigation. For this purpose, ontology content
exploration should gain much attention.
Ontology content is usually explored following
subclass-of (is-a) hierarchy. However, we have
obvious problems because many types of
hierarchical classification could exist, and disease
characterization is susceptible to various
interpretations. A disease is interpreted from various
viewpoints. Consider diabetes as an example.
Clinical technologists may pay attention to the body
part that has the abnormality and classify diabetes as
an abnormal blood sugar level. On the other hand, a
certain specialist may pay attention to the main
condition and may classify diabetes as an
abnormality in metabolism, and another specialist
may classify diabetes as a lifestyle disease. Staff
administering the medical care implicitly
understands which is-a hierarchy should be used for
disease interpretation in correlation with their
respective interpretations. This suggests that one is-a
hierarchy of diseases cannot cope with such a
diversity of viewpoints, since a single-inheritance
hierarchy necessarily represents one viewpoint.
Some researchers would say “this is why we use
multiple-inheritance is-a hierarchy. Why don’t you
use it for disease organization?” The answer to such
a question is as follows. In ontological theory, is-a
hierarchy must represent essential property of things
and hence it should be single-inheritance, since
essential property of things cannot be multiple.
Imagine objects, processes, attributes, all of them
have their own unique and essential properties. The
use of multiple-inheritance for organizing things
necessarily blurs what is the essential property of
things. This observation is strongly supported by the
fact that both of the well-known upper ontologies:
DOLCE and BFO use single-inheritance hierarchy.
If we add a practical difficulty, we can indicate the
instance management issue. Instances of a class
must have their own appearance/disappearance
policy according to their essential properties.
Multiple-inheritance is-a hierarchy hides essential
property of things and hence you cannot identify
DEVELOPMENT OF FUNDAMENTAL TECHNOLOGIES FOR BETTER UNDERSTANDING OF CLINICAL
MEDICAL ONTOLOGIES
237
Figure 2: Gathering knowledge described in ontology to generate the classification hierarchy with paying attention to
particular thing that caused the abnormal finding.
what policy to use for their
appearance/disappearance.
So, the problem is how to reconcile the
conflicting requirements of multiple-views and
single-inheritance in a good ontology. Each of these
is-a hierarchies is significant and we try to keep both
requirements. Furthermore, in order to obtain a deep
understanding of a disease, it is important to use
more than one disease is-a hierarchy which
represents the essential structure underlying the
target world. In order to tackle this important issue,
we adopt an approach of dynamically generating is-a
hierarchies of diseases according to the viewpoint of
users from an ontology using single-inheritance.
3.3 Dynamic Generation of is-a
Hierarchies
To develop a function of dynamic generation of is-a
hierarchy, we need to classify plausible viewpoints
of the hierarchical classifications and relate these
viewpoints to the conceptual structures. This
function uses the aspect to determine is-a hierarchies.
This aspect is used to traverse the ontology and
collect related information. An is-a hierarchy is then
generated using this information.
For instance, to generate an is-a hierarchy from
the viewpoint of the pathological condition, the main
pathological condition of metabolic disorders is an
abnormal state consisting of a metabolic
abnormality, and the subclassification of a metabolic
disorder is generated using is-a hierarchy
information about the metabolic abnormality, which
is the disorder’s main pathological condition. The
subclassification of a metabolic disorder is similar to
an is-a hierarchy of metabolic abnormalities.
The diabetic main pathological condition is
carbohydrate metabolism abnormality, which is a
particular type of metabolic abnormality. Therefore,
the disease of carbohydrate metabolism is
subordinate to metabolic disease, and diabetes is one
of the diseases of carbohydrate metabolism.
Moreover, one might want to see an is-a
hierarchy of diseases analogous to the location
where their main pathological condition appears. In
such a case, the part-of relationships of the human
body converted into an is-a hierarchy and is used to
generate such an is-a hierarchy. This is-a hierarchy
is similar to the part–whole relationship of the
human body. For instance, mitral valve disease and
tricuspid valve disease are classified as
subcategories of cardiac disease, because the
Part-of hierarchy and class
constraint information of
heart and circulatory system
disease’s abnormally
object information
KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development
238
Table 1: Some hierarchical classifications.
(a) Viewpoints for
classification of diseases
(b) Aspects for dynamic generation of is-a
hierarchies of diseases from ontology
finding site
(organs)
(m)
organs of the human body (part-of), main
pathological condition information (slot)
(s)
organs of the human body (part-of),
symptoms information(slot)
finding site
(organ systems)
(m)
organ systems(part-of), main pathological
condition information (slot)
(s)
organ systems, symptoms information(slot)
pathogenic abnormality information (is-a),
main pathological condition information
(slot)
symptoms of abnormality hierarchy
information(is-a),
diagnosis and treatment department
information, symptoms information(slot)
Mapping
ICD10
Mapping
* (m): main pathological condition occurred, (s): symptoms
tricuspid valve and the mitral valve have a part–
whole relationship with the heart. Mitral stenosis
and mitral prolapse are classified as diseases of the
mitral valve, and tricuspid stenosis is classified as a
disease of the tricuspid valve. Furthermore, many
different classifications are possible for human
organs. For example, the heart might be classified
occasionally as just a human organ, and on other
occasions as a circulatory organ. Converting
between these is-a hierarchies can be achieved by
viewpoint switching (Figure 2).
Thus, the dynamic generation of an is-a
hierarchy enables us to switch some is-a hierarchies
using collected information. To collect information,
we should trace not only is-a and part-of
relationships but also the relationships based on the
role played by a particular concept. In addition, as in
ICD-10 and ICPC2 (ICPC2), there are is-a
hierarchies that were established as useful to those
particular purposes. We listed generally needed is-a
hierarchies as determined by a general medical
textbook, some medical information web services,
and the opinion of clinicians. For comparison, the
matrix below lists generally needed is-a hierarchies,
the kind of information in the ontology used to
generate is-a hierarchy (see. Table 1).
There are many advantages in choosing not to
realize all the classifications within an ontology, but
to dynamically generate is-a hierarchy from the
ontology. For instance, the inconvenience of
multiple inheritance might be averted as discussed
above. A longstanding problem is the
incompatibility of theory and practice in the use of
is-a relation: ontological theory tells us that is-a
relation must be used only between concepts which
the lower concepts are genuine subconcepts
inheriting essential property from their upper
concept, and hence is-a hierarchy Is unique, while
Figure 3: Switching the index using the dynamic
generation of is-a hierarchy function, and switching the
content with the natural language explanation.
concepts are often classified into multiple classes in
practice. Our method proposed in this paper is a
good compromise of this conflicting situation. It
allows people to consider their particular viewpoint
is “the” essential aspect to them without any harm to
others. This function could readily apprehend
systematic knowledge, for example, about diseases
that have many facets, from various perspectives.
This function would also support medical students in
systematically understanding that knowledge
effectively. Furthermore, this function is important
on a clinical site in which many specialists with
varied backgrounds cooperate.
3.4 Implementation of Medical
Information Service System
We implemented fundamental technology discussed
above, and built a prototype of medical information
service system using them. They are developed
using HozoCore, which is java API for ontology
building by Hozo, and Java Servlet. Figure 3 shows
the prototype of medical information service system.
The system generates is-a hierarchies according to
the specified user’s intentions. When a concept in
the index is clicked, detailed explanation is
displayed. It changes the medium used for
expressing content, such as a table or natural
Switching contents
Table view
Index
Natural language
Dynamic switching of
classification hierarchy
Conceptual map
DEVELOPMENT OF FUNDAMENTAL TECHNOLOGIES FOR BETTER UNDERSTANDING OF CLINICAL
MEDICAL ONTOLOGIES
239
language. For the natural language explanations, the
users can choose out of the three types according to
their understanding about diseases. When the system
displays a network of complex relationships such as
causal chains, the user can use the conceptual map
generation tool developed in our laboratory. It can
generate conceptual maps based on any viewpoint
and help users understand the knowledge extracted
from ontologies (Kozaki, K. 2008).
We performed an informal evaluation of the
implemented system in a workshop and received
favorable comments from medical experts. They
especially liked the dynamic is-a hierarchy
reorganization, which is the first solution to the
multi-perspective issues of medical knowledge in
the world. We also conducted a feasibility study of
the fundamental technologies for building
application systems using medical ontologies and
found that the technologies are independent of the
size of the ontology and are a common foundation
for the development of various systems.
4 CONCLUSIONS
After the conceptual framework of human anatomy
and disease were validated, we developed data input
software and templates to scale up the content of the
ontology. Our next tasks are to check data
consistency, provide feedback regarding the checked
data to the experts who input the data, and improve
and adjust the ontology. Some problems still need to
be resolved. We need to deepen investigation of
fundamental technology regarding each of the three
main parts (navigation method, content processing,
and media selection). Also, we should identify
functional enhancements of this fundamental
technology based on users’ needs and implement an
enhanced version. For the navigation method, the
added functionality of presenting ancillary
information of linked desired content pages would
help users decide whether to obtain information they
want, and this function might respond effectively to
user needs. So we have to consider additional
functions for better navigation based on users’ needs.
Also, the search system needs improvement of
information retrieval. Determining what kind of
searches are required in medicine, providing search
functions suitable to medicine, and functional
extension. Especially, in medicine, it is important
not only to search for diseases based on simple
symptoms, but also to find all diseases that may
cause the patient’s symptoms. In media selection, it
is necessary that the optimal media be chosen
according to timing and case. This selection method
is also a subject of ease of using.
ACKNOWLEDGEMENTS
This research was supported by the Ministry of
Health, Labour and Welfare, Japan, as Development
business and research of “Medical-knowledge-based
database for medical informatics system.”
REFERENCES
DOLCE: a Descriptive Ontology for Linguistic and
Cognitive Engineering, http://www.loa-cnr.it/
DOLCE.html
FMA: The Foundational Model of Anatomy ontology,UW
Medicine,http://sig.biostr.washington.edu/projects/fm/
GALEN: OpenGALEN, http://www.opengalen.org/
ICD-10: International Statistical Classification of Diseases
and Related Health Problems 10th Revision,
http://www.who.int/classifications/icd/en/
ICPC2: International Classification of Primary Care,
Second edition, http://www.who.int/classifications
/icd/adaptations/icpc2/en/
Kozaki, K., et.al, 2002. Hozo: An Environment for
Building/Using Ontologies Based on a Fundamental
Consideration of “Role” and “Relationship”. In Proc.
of EKAW2002, pp.213-218.
Kozaki, K., Hirota, T., Mizoguchi, R., 2008. Development
of a Conceptual Map Generation Tool for Exploring
Ontologies, In Poster Noets of ESWC 2008.
MEDIS-DC: Medical Information System Development
Center, http://www.medis.or.jp/
MeSH: Medical Subject Headings, http://www.nlm.nih.
gov/mesh/
Mizoguchi, R., et.al, 2007. A Model of Roles within an
Ontology Development Tool: Hozo, In Journal of
Applied Ontology, 2, pp.159-179.
Mizoguchi, R., et.al., 2009. An Advanced Clinical
Ontology, In Proc. of ICBO, pp.119-122
Rector, A., 2002. Analysis of propagation along transitive
roles: Formalisation of the GALEN experience with
Medical Ontologies, In International Workshop on
Description Logics, CEUR-Proceedings 53.
Rosse, C., et.al, 2003. A reference ontology for biomedical
informatics: the Foundational Model of Anatomy, In
Journal of Biomedical Informatics, 36, pp.478-500.
SNOMED-CT: Systematized Nomenclature of Medicine-
Clinical Terms, http://www.ihtsdo.org/snomed-ct/
SNOMED Clinical Terms User Guide, The International
Health Terminology Standards Development
Organisation, 2008
Stefan, S., et.al, 2007. SNOMED CT's Problem List:
Ontologists' and Logicians' Therapy Suggestions. In
Proc. of the Medinfo 2007 Congress, Studies in Health
Technology and Informatics (SHTI-series).
YAMATO: Yet Another More Advanced Top-level
Ontology, http://www.ei.sanken.osaka-u.ac.jp/hozo/
onto_library/upperOnto.htm
KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development
240