REPRESENTING THE INTERNATIONAL CLASSIFICATION

OF DISEASES VERSION 10 IN OWL

Manuel M

oller, Michael Sintek, Ralf Biedert, Patrick Ernst, Andreas Dengel

German Research Center for Artiﬁcial Intelligence (DFKI) and University of Kaiserslautern, Kaiserslautern, Germany

Daniel Sonntag

German Research Center for Artiﬁcial Intelligence (DFKI), Saarbrcken, Germany

Keywords:

Formal knowledge representation, Automatic ontology generation, Medical ontologies, International classiﬁ-

cation of diseases.

Abstract:

Current efforts in the biomedical ontology community focus on establishing interoperability and data inte-

gration. In covering human diseases, one of the major international standards in clinical practice is the In-

ternational Classiﬁcation for Diseases (ICD), maintained by the World Health Organization (WHO). Several

country- and language-speciﬁc adaptations exist which share the general structure of the WHO version but

differ in certain details. This complicates the exchange of patient records and hampers data integration across

language borders. We present our approach for modeling the hierarchy of the ICD-10 using the Web Ontology

Language (OWL). Our model captures the hierarchical information of the ICD-10 as well as comprehensive

class labels for English and German. Specialties such as “Exclusion” statements, which make statements about

the disjointness of certain ICD-10 categories, are modeled in a formal way. For properties which exceed the

expressivity of OWL-DL, we provide a separate OWL-Full component which allows us to use the hierarchi-

cal knowledge and class labels with existing OWL-DL reasoners and capture the additional information in a

machine-interpretable way.

1 INTRODUCTION

Over the last decades healthcare has changed from

isolated treatments towards a distributed treatment

process. This process depends greatly on the coop-

eration of specialized medical disciplines. Moreover,

medicine questions require to take into account an

enormous amount of expert knowledge before deci-

sions are made. To facilitate information exchange

and sharing of knowledge in medical domains, stan-

dardization is playing an important role. The goal is

to increase the interoperability within all domains of

the healthcare industry so that the interchange of doc-

uments can be simpliﬁed and work ﬂows can be im-

proved.

The difﬁculty in the area of medical knowledge

management is the high diversity of knowledge about

single entities. Let us consider a patient in a clini-

cal environment. Even if he has a simple disease, he

would have to undergo a high number of examinations

in different clinical departments. In each step of his

treatment, huge amounts of metadata are created and

stored, based on single speciﬁc models every time.

The challenge is to integrate these islands of informa-

tion (Lenz, 2005), so that an overall knowledge base

can emerge.

Another problem of this integration process is se-

mantic heterogeneity, which means, that there are dis-

agreements about the semantics or interpretation of

concepts between the terminologies. Additionally,

medical knowledge is very complex and evolves con-

tinuously over time. Therefore, new architectures and

standards are needed that deal with these problems

(Sonntag et al., 2009), (Sonntag, 2010).

Standardized terminologies have a long history in

medicine. For human diseases, the ﬁrst approaches

date back to the 18th century. The roots of the modern

International Classiﬁcation of Diseases (ICD) can be

traced back to the Bertillon Classiﬁcation of Causes

of Death. The ICD was introduced in 1893 at the

International Statistical Institute in Chicago. Five

years later, the American Public Health Association

Möller M., Sintek M., Biedert R., Ernst P., Dengel A. and Sonntag D..

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL.

DOI: 10.5220/0003082400500059

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2010), pages 50-59

ISBN: 978-989-8425-29-4

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

(APHA) recommended that Canada, Mexico, and the

United States should also adopt it. Many other coun-

tries joined subsequently. Over the last 100 years it

was revised several times. The sixth revision included

morbidity and mortality conditions and was renamed

the “Manual of International Statistical Classiﬁcation

of Diseases, Injuries and Causes of Death (ICD).”

Since 1948 the World Health Organization assumed

the responsibility for maintaining and publishing re-

vised versions of the ICD.

The currently internation-

ally effective revision is ICD-10 from 2006.

Although the overall structure of the ICD-10 was

accepted by numerous countries, different versions

exist which are maintained by national institutions.

For instance, the German version of the ICD-10 is

maintained by the DIMDI

which is under the author-

ity of the German Federal Ministry of Health. While

major parts of the ICD-10 hierarchy are equal both in

the DIMDI version and the WHO version, we found

out that the structure and content of certain parts of

the ICD-10 varies. Section 4 provides details of these

differences.

The aim of the work presented here is to gener-

ate an ontology covering the domain of human dis-

eases based on the classiﬁcations of the two country

speciﬁc ICD-10 versions described above. The ulti-

mate goal is to leverage technologies from the Se-

mantic Web to ease the work of medical experts by

supporting them in (1) making medical image data as

well as patient records available for semantic search,

and (2) by providing intelligent annotation sugges-

tions based on rich formal models for medical do-

main knowledge. This research was triggered by the

broader effort within the research project MEDICO.

From our discussions with clinicians we learned that

a representation of the ICD-10 is an absolute neces-

sity for efﬁcient semantic radiological image annota-

tion in the everyday practice of the university hospital

participating in MEDICO (M

oller et al., 2008).

2 RELATED WORK

The initial idea for generating an OWL version of

ICD-10 from data available on the web dates back to

a similar approach for generating an OWL model for

ICD-9 as presented in (M

oller and Mukherjee, 2009).

“History of the development of the ICD,” available on

the WHO website at http://www.who.int/entity/classiﬁca-

tions/icd/en/HistoryOfICD.pdf

http://www.who.int/classiﬁcations/icd/en/

Deutsches Institut f

ur Medizinische Dokumentation

und Information

Biomedical ontologies and terminologies received

high attention in the last decade and provide promis-

ing technologies for data integration. Bodenreider

et al. evaluated popular large scale ontologies such

as SNOMED, FMA, and Gene Ontology and stated

that “ontologies play an important role in biomedi-

cal research through a variety of applications” (Bo-

denreider, 2004). In this context, a number of semi-

structured medical terminologies and classiﬁcation

systems have been converted to formally structured

formats recently.

For the Systematized Nomenclature of Human

and Veterinary Medicine (SNOMED) (Cote et al.,

1993), an OWL ontology was created and used to

detect weaknesses in the original modeling (Schulz

et al., 2007; Schulz et al., 2009).

Noy and Rubin have presented an approach for

translating the Foundational Model of Anatomy on-

tology (FMA) to OWL (Noy and Rubin, 2008). From

their approach we adopted the idea to split the gen-

erated ontology into an OWL-DL and an OWL-Full

component.

Cardillo et al. presented an approach for a formal

representation of mappings between ICD-10 and the

International Classiﬁcation of Primary Care version 2

(ICPC-2) (Cardillo et al., 2008). However, their fo-

cus was on the formal representation of mappings be-

tween ICD-10 and ICPC-2. The work presented in

this paper tries to complement their efforts by provid-

ing a formal model of additional relations within the

ICD-10.

3 APPROACH

This section describes our general approach for the

generation of the ICD-10 in OWL. Figure 1 shows

the data ﬂow during the ontology generation process.

Following the elements in this diagram, the subse-

quent sections will discuss the different processing

steps and give details about the applied techniques

and algorithms.

3.1 Data Sources

The OWL ontology which we generated is based on

data available via the websites of the organizations

responsible for maintaining the respective ICD ver-

sions. As we will show, the data which is publicly

available on the Internet is well suited to generate a

rich formal model of the classiﬁcation of human dis-

eases. The websites are highly structured and contain

enough information to ﬁt the use case in the MEDICO

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL

GermanICDGermanICD‐‐1010

websitewebsite,,

maintainedmaintained byby

InternationalICDInternationalICD‐‐1010

websitewebsite,,maintainedmaintained

WHOWHO

GermanICDGermanICD‐‐1010

(XML(XMLfilefile))

DIMDIDIMDI

CrawlerCrawler

/XML Parser/XML Parser

CrawlerCrawler

/XML



Parser/XML



Parser

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

OntologyOntology MergerMerger

OWLOWL

differencesdifferences

filfil

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

OWLOWL‐‐DLDL

componentcomponent

OWLOWL‐‐FullFull

componentcomponent

fil

Figure 1: Data ﬂow of the ontology generation process.

project. Another advantage is that we can reﬂect up-

dates of the ICD published on the websites by re-

running our crawlers.

For the English version of the ICD we used the

ofﬁcial WHO website.

The website only partly re-

ﬂects the original hierarchical structure of the ICD-

10. As an additional source we used the ICD-10

manual (WHO, 2004). Figure 2 (a) shows a screen-

shot covering the ﬁrst of “Nutritional anaemias (D50-

D53).”

From the different German sources available we

chose to use the current ICD-10-GM, “GM” being the

“German Modiﬁcation” (see Section 4). Figure 2 (b)

shows the same fragment of the ICD-10 as the previ-

ous screenshot, but this time in German. Our start-

ing points for the German ICD-10 is the respective

website and a publicly available XML ﬁle which is

structured using the Classiﬁcation Markup Language

(ClaML). As the name suggests, ClaML is special

language designed to represent classiﬁcation hierar-

chies (Hoelzer et al., 2002). It provides special no-

tations to state super- and subclass relations, declare

attributes, and to specify metadata elements, among

other things. To interpret the notation correctly,

we use the WHO manual (WHO, 2004) and a sup-

plementary documentation for attributes exclusively

stated in the DIMDI version (Deutsche Krankenhaus-

gesellschaft, 2009) using the the DIMDI website.

3.2 OWL Model Generation

The general structure of the ICD-10 is as follows. It

consists of “Chapters” using Roman numerals from I

http://apps.who.int/classiﬁcations/apps/icd/

icd10online/

http://www.dimdi.de/static/de/klassi/diagnosen/icd10/

htmlgm2009/index.htm

(a) Example for an English entry from the WHO ICD-10

website

(b) Respective entry from the German DIMDI ICD-10 web-

site

Figure 2: Language-speciﬁc ICD-10 online versions.

to XXI. The chapters again contain “Blocks of cat-

egories” (e. g., Chapter III: “Diseases of the blood

and blood-forming organs and certain disorders in-

volving the immune mechanism”) which specify a

range of categories of a particular aspect (e. g., D50-

D89). These blocks then contain “Categories,” de-

noted by an ICD-10 code, a capital letter, and Ara-

bic numbers (e. g., D50-D53: Nutritional anaemias;

D55-D59: Haemolytic anaemias; etc.). They are

further subdivided into “Subcategories.” The sub-

KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development

categories are coded by attaching an additional digit

after the decimal point (e. g., D50.1: Sideropenic

dysphagia) . These codes differentiate from the

speciﬁc language or writing system of the differ-

ent ICD-10 versions. Contrary to the WHO web-

site, the DIMDI XML ﬁle constrains the subcate-

gories by using an additional decimal number, which

is appended to the codes of their parent categories.

These subcategories are also deﬁned using “Modi-

ﬁers” which specify a set of “ModiﬁerClasses” as

subclasses. Each of these classes possesses a number,

a label, and additional information such as “Exclu-

sions” or “Inclusions.” If a category contains a “Mod-

iﬁer,” it will be specialized by generating new cate-

gories with each particular “ModiﬁerClass.” These

relations are represented in OWL by creating a new

OWL class for each “Modiﬁer,” which is a subclass of

icd10:Modifier, and deﬁning the appropriate Mod-

iﬁerClasses as subclasses. The combination is de-

noted using an owl:unionOf of the speciﬁc category

and each ModiﬁerClass. This form of speciﬁcation is

used extensively. Our analysis has shown that 4488 of

the 16214 classes are speciﬁed in this way by DIMDI.

ICD-10 is not only a classiﬁcation of diseases, but

the terminology also includes links to other related

aspects, such as symptoms, signs and consequences

of other external causes. Therefore, the manual de-

scribes an additional level of order which groups cer-

tain chapters according to their particular aspects. For

example, “Chapters I to XVII relate to diseases and

other morbid conditions.” It is worth mentioning that

this systematic level is not available from the website

but only from the manual.

Using this information about chapters and groups

of chapters we modeled the ﬁrst two hierarchy levels

by hand. The OWL class icd10:Entry is the super

class of the bilingual ICD-10 hierarchy. As mentioned

before, differences between the German and English

ICD-10 exist. Our analysis shows that there are ICD-

10 categories which are present in the German ICD-

10 but not in the English ICD-10 and vice versa. Sec-

tion 4 gives details about these differences.

In addition, the origins of the concepts are also

encoded in our OWL model. The ﬁrst approach was

to add a super class for each class to denote its ori-

gin. But this proved to be incorrect. Let us con-

sider the block R00-R99, which is present in both

terminologies and thus gets both super classes. The

symptom R65, which has the super class R00-R99, is

only stated by the DIMDI. Consequently, a reasoner

would infer that R65 is in the DIMDI and WHO ver-

sion, because it would build the transitive closure and

R65 would get both super classes. For that reason,

we decided to denote the provenance using the two

Figure 3: General class hierarchy of the OWL model.

boolean OWL-Full properties icd10:isDIMDIEntry

and icd10:isWHOEntry.

Figure 3 shows an (abbreviated) example of

the generated class hierarchy. We use two HTTP

crawlers, implemented in Java, to generate OWL

models for each of the two input sources. OWL

classes and axioms are generated using the Jena On-

tology API.

Other libraries—such as the OWL API

(Bechhofer et al., 2003)—were not able to handle the

OWL-Full expressivity of our modeling.

The generated OWL model consists of two com-

ponents. The OWL-DL part contains the hierarchy

of the ICD-10 according to the hierarchical structure

described above. All ICD-10 categories and subcate-

gories are reﬂected by OWL classes. The hierarchical

information is reﬂected by owl:subClassOf axioms.

For a discussion of the contents of the OWL-Full part

see further below.

We will explain the next steps by giving an exam-

ple. Figure 2 shows the ﬁrst part of the WHO ICD-

10 website about “Nutritional anaemias (D50-D53).”

We will focus on the entry “D50.0 Iron deﬁciency

anaemia secondary to blood loss (chronic)” as a guid-

ing example throughout this paper.

Each class is identiﬁed by an URL, which con-

sists of a speciﬁc ICD-10 name space and the special

term as the URL anchor. The terms for the categories

are simply the particular ICD-10 codes. For blocks

http://jena.sourceforge.net/ontology/

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL

and chapters, a range pattern is used which covers

their content, e. g., D50-D53. From this we create an

OWL class with the local name “D50.0.” The “.0”

indicates that this is a sub-category of “D50 Iron deﬁ-

ciency anaemia.” Thus, we add an owl:subClassOf

axiom which represents this relationship. The bold-

faced name of the sub-category becomes the En-

glish rdfs:label of this class. Later, by merging

with the OWL model of the German ICD-10, we

can also add the German labels “Eisenmangelan

amie

nach Blutverlust (chronisch).” For some concepts,

the DIMDI speciﬁes up to three labels, which dif-

fer in their length and detail. The smaller labels are,

thereby, necessary because some print formats require

them. In our case we can neglect this limitation and

use always the most detailed label available. We use

the standard XML language tags to differentiate be-

tween these languages.

3.3 ICD-10 Characteristics

Despite a specialization hierarchy, multiple character-

istics are stated in the WHO manual (WHO, 2004)

and the DIMDI supplement (Deutsche Krankenhaus-

gesellschaft, 2009), and they can be shared by differ-

ent classes. These are:

Dagger and Asterisk Categories. Statements con-

taining information about an underlying disease

with a particular additional manifestation can be

expressed thanks to asterisk and dagger codes.

Underlying diseases are marked with a dagger and

are the primary criterion. Therefore, they have

to appear in the diagnostic statement, whereas

the manifestation marked with an asterisk is

only additional. These circumstances are rep-

resented in OWL using two properties, namely

icd10:hasAdditionalManifestation with its

inverse icd10:hasUnderlyingDisease. The

ﬁrst one’s domain is all classes which represent

a dagger category and have a range of all asterisk

categories. The restriction that an additional man-

ifestation needs at least one underlying disease is

expressed by the property’s cardinality, which is

at least one.

Optional Concepts. The DIMDI deﬁned a supple-

mental characteristic and this is only applied in

their version of the terminology. Optional con-

cepts are similar to dagger and asterisk categories.

If marked as optional, a concept will be manda-

tory for some diagnoses but only supplemental for

other ones.

Categories Limited to one Gender. The ICD-10

contains several categories which are only ap-

plicable to either males or females. Consider,

for instance, diseases of the genitals, like “D40

Neoplasm of uncertain or unknown behavior of

male genital organs” or conditions which occur

during the pregnancy of women, e. g., “O00

Ectopic pregnancy.” The facts are represented by

one super class for each gender.

Sequelae Categories. Sequelae categories are used

for mortality cause encoding. They indicate that

the death is not caused by the main effect of a

given disease. Instead it is caused by residual ef-

fects.

Postprocedural Disorders. Categories which fall

under this characteristic point out conditions and

complications which occur after treatment, e. g.,

surgical wound infections or shock.

Contrary to dagger and asterisk categories, each char-

acteristic is represented using a speciﬁc super class.

All classes which share the characteristic are sub-

classes of this class.

3.4 Handling ICD-10 Exclusions

For some ICD-10 categories so-called “Exclusions”

also exist. According to the ICD-10 manual (WHO,

2004), they exclude certain conditions that, “although

the rubric title might suggest that they were to be clas-

siﬁed there, are in fact classiﬁed elsewhere.” The

example in Figure 2 (a) lists two such excludes for

D50.0: “acute posthaemorrhagic anaemia” with a link

to ICD-10 category D62 and “congenital anaemia

from fetal blood loss” with a link to category P61.3.

We capture this information by adding owl:dis-

jointWith axioms between D50.0 and D62 as well

as between D50.0 and P61.3. This can be expressed

using the expressivity of OWL-DL (see Figure 4).

However, by relying exclusively on owl:dis-

jointWith axioms, we would lose important infor-

mation. As the ICD-10 manual states, exclusions can

be extended using additional strings and constructions

of braces. They indicate that neither the words that

precede them nor the words after them are proper

terms. Thus, a more precise qualiﬁcation has to be

applied (WHO, 2004). If we compare the brace con-

structs with the encoding in the XML ﬁle, it becomes

clear that they are used to provide a more compre-

hensive structuring of the data. The XML ﬁle reﬂects

them by splitting up the “Exclusions” and adding a

new fragment for each brace element to the “Ex-

clusions.” Figure 5 depicts “Inclusions” for concept

“O71.6: Obstetric damage to pelvic joints and lig-

aments” using braces and ﬁgure 6 shows them us-

ing the German XML encoding. As we see, without

proper post-processing we are not able to relate that

KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development

OWL-FullOWL-DL

icd10:ICDDescription

rdf:type

icd10:hasExcludes

icd10:concernsClass

congenital anaemia

from fetal blood loss

rdfs:label

icd10:P61.3

icd10:D62

icd10:D50.0

icd10:

ICDDescription

icd10:ICDDescription

acute

posthaemorrhagic

anaemia

rdfs:label

icd10:hasExcludes

rdf:type

Figure 4: Structure and relationship of OWL-DL and OWL-Full component by example.

the “Exclusion” of concept M54.1 is shared by multi-

ple Exclusions.

These qualiﬁcations are covered in additional

OWL individuals of class icd10:Description. The

individuals can have several properties of type

icd10:concernsClasses. Because this property

concerns other ICD-10 categories, it needs to have

a class-valued range. Thus, the individuals require

OWL-Full expressivity. To encode the string infor-

mation of an “Exclusion,” we use rdfs:label. For

each excluded statement, we generate one individ-

ual. We also encode the information contained in the

brace constructs, which appear in the WHO version.

Therefore, we create an individual for the information

which occurs after the braces. They are related to each

Exclusion using an OWL-Full property icd10:qua-

lifiedBy. Extracting this information for the DIMDI

data is among our next steps.

Additionally, a closer look at the ICD-10 revealed

that for numerous categories, the “Exclusions” do

not point to other ICD-10 categories but to arbi-

trary descriptions of certain medical symptoms. Fig-

ure 5 gives an example showing ICD-10 subcategory

“O71.6: Obstetric damage to pelvic joints and liga-

ments.” As we cannot generate proper disjointness

axioms for these exclude expressions we decided to

store them using the exclude individuals described

above without pointing to another ICD-10 category.

3.5 Handling ICD-10 Inclusions and

Notes

Similar to the “Exclusions” two other properties for

categories are part of the ICD-10. “Inclusions” are ad-

ditions to the rubric in which they occur. The ICD-10

manual describes them as a guide and provides exam-

ples to formulate diagnostic statements. “Inclusions”

are represented using the individuals in OWL-Full in

the same way as “Exclusions.”

In addition, it is possible that a note is provided

for an ICD-10 element. These notes give hints how

to use the particular category, block, or chapter. For

example, a physician who is writing a medical report

sees from these hints that the category “G09 Seque-

lae of inﬂammatory diseases of central nervous sys-

tem” is to be used to indicate conditions whose pri-

mary classiﬁcation is G00-G08 (i. e., excluding those

marked with an asterisk) as the cause of sequelae,

themselves classiﬁable elsewhere. The only purpose

of the notes is to support human beings in interpret-

ing the ICD-10. Also, they are not interpretable for

reasoners because they only contain continuous text.

For that reason, we do not relate the notes individuals

to classes with OWL-Full properties, instead we use

owl:AnnotationProperties.

3.6 Merging the English and German

OWL Models

To merge the two ICD-10 variants, we have to distin-

guish between the OWL-DL and -Full parts. The two

OWL-Full parts are merged by just importing them

into a new ontology. This is possible because they

only contain properties of the classes deﬁned in the

DL versions and the class deﬁnitions were not altered

during the merging process.

The merging of the DL ontologies can be divided

into two phases. First, an automatic integration is per-

formed, which is then reﬁned by a manual step. The

automatic merging process starts with the WHO on-

tology. As stated in the last paragraph of section 3.2,

all classes are identiﬁed using their particular ICD-10

code. In most of the cases, these codes are the same

in the DIMDI and WHO version. Therefore, we be-

gin the integration by checking if each DIMDI class

is present in the WHO version. If so, we add the label

and owl:subClassOf properties of the class. Adding

owl:subClassOf axioms is necessary because only

some of them exist in the DIMDI version. For ex-

ample, if a block only appears in the DIMDI version,

all owl:subClassOf relations concerning this block

only occur in the DIMDI ontology. If the class is not

contained, it will be created and all properties will be

copied.

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL

Figure 5: Example for ICD-10 category with “Exclusions” that do not point to other ICD-10 categories.

Figure 6: Example for ICD-10-GM braces constructs within “Exclusions”.

Figure 7: Screenshot of the OWL-DL version of the ICD-10

in the Ontology Editor Protg.

In addition, a few classes exist which are seman-

tically very similar, but differ by their ICD code. For

example, the DIMDI chapter D50-D90 only varies in

the range and the existence of the class D90 from the

WHO chapter D50-D89, but both concern the same

diseases. To merge these classes, we ﬁrst manually

determine all possible pairs (classA, classB) which

differ in this sense. After that, we deﬁne a new super

class unionAB for each pair. This is the owl:unionOf

of the pair’s classes and gets the a concatenation of

the local names of both. To determine the location of

the new class, we search the ﬁrst super class which

classA and classB have in common. unionAB is then

added as a subclass of this class. This traversing is

necessary, because there can be super classes not cov-

ering the entire range of unionAB. For example, we

merge the blocks D80-D89 and D80-D90 and the di-

rect super classes are D50-89 and D50-D90 respec-

tively, which are only present either in the WHO or

DIMDI version. Therefore, we have to traverse the hi-

erarchy one step further and ﬁnd the appropriate super

class, which is owl:unionOf of the classes D50-D89

and D50-D90.

Besides the new integrated ontology, a differ-

ence ontology is generated during the merging pro-

cess, which distinguishes between WHO and DIMDI

classes only occurring in one version. We know

that this knowledge is already contained in the on-

tology by the label and icd10:isDIMDIEntry or

icd10:isWHOEntry. However, we decided to pro-

duce an explicit representation of the differences, be-

cause it makes the merging process more transparent

and the differences are easier to examine. For these

reasons, we denote the differences in both ontologies

using OWL-Full properties in the difference ontology.

We are using one property for every ICD source to de-

KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development

note the exclusiveness and one property to denote the

classes later manually added by the merging process.

Figure 7 shows a screenshot of the generated OWL

class for ICD-10 category D50.0 in Protg

4 RESULTS AND DISCUSSION

By the deﬁnition given in the ICD-10 manual, the ICD

is a classiﬁcation system with “a hierarchical structure

with subdivisions.” And further: “A statistical classi-

ﬁcation of diseases should retain the ability both to

identify speciﬁc disease entities and to allow statis-

tical presentation of data for broader groups, to en-

able useful and understandable information to be ob-

tained.” From this we concluded that the ICD is based

on a hierarchical system of classes. The relations be-

tween these classes are proper subset relations in the

sense of set theory. Thus, we decided to represent the

relations of the ICD using OWL and its subClassOf

relations.

Table 1 lists some general metrics for the gener-

ated ontology. It also lists differences between the

German and the English ICD-10 versions in terms of

number of classes. The majority of all ICD-10 cate-

gories, i. e., about 60%, share the same ICD-10 code

(for details see Section 3.2) and thus could be mapped

using this as an identiﬁer. However, there were some

discrepancies between the WHO and German ver-

sions. One reason were different modeling granu-

larities producing more categories in some branches.

These differences are discussed subsequently.

We decided to split our OWL model into two

components similar to the approach of Noy and Ru-

bin in (Noy and Rubin, 2008) for translating the

Foundational Model of Anatomy ontology to OWL.

The OWL-DL component allows to perform DL-

reasoning using standard OWL-DL reasoners like

Pellet (Sirin et al., 2007). Information from the ICD

which requires modeling in OWL-Full is still avail-

able in the OWL-Full component. To use the com-

plete model, the OWL-Full component can be loaded.

This variant imports the OWL-DL model.

Differences between WHO and DIMDI

Versions of ICD-10

In both terminologies we located classes that either

only occur in the DIMDI or WHO version. These

differences can be classiﬁed into two categories:

• classes which only appear in one ICD-10 variant

and have no particular counterpart in the other and

http://protege.stanford.edu

• classes which have a slightly different identiﬁca-

tion, but can be merged manually.

The ﬁrst category is the most extensive. We iden-

tiﬁed 1,145 classes which are exclusively part of the

WHO version and 5,707 classes exclusively part of

the DIMDI version. It is important to note that we

are only regarding the classes of the actual ICD-10

entries and not the classes for constructs like mod-

iﬁers; these will be discussed later. Furthermore,

there are blocks which differ in both versions. This

means that one version speciﬁes some parts of its

terminology with more granularity than the other or

that some concepts were simply left out. For exam-

ple, the WHO subdivides the block “V01-X59 Ac-

cidents” into 27 sub-blocks using two hierarchy lev-

els. In contrast, the DIMDI version does not make

any further subdivisions here. Moreover, the DIMDI

describes the block “U60-U61 Stadieneinteilung der

HIV-Infektion” (“Staging of HIV-Infection”), which

is not present in the WHO version at all.

We derived the second category during a manual

examination of all differences. Hereby, we identiﬁed

ﬁve blocks, which vary in their ICD-10 code. Ta-

ble 2 lists them and opposes the WHO blocks with

the details which the DIMDI contains. In all cases,

there are differences in the range of the blocks. This

is interesting, as even though a block can specify a

broader range, it can be semantically more restricted.

We will illustrate this by an example. The WHO ver-

sion has the block “U80-U89 Bacterial agents resis-

tant to antibiotics.” The German version “U80-U85

Infektionserreger mit Resistenzen gegen bestimmte

Antibiotika oder Chemotherapeutika” has almost the

same range (“U80-U89” vs. “U80-85”). It could be

assumed that the block with the smaller range is also

more speciﬁc. But in this example the opposite is true

since the German term “Infektionserreger” (“infec-

tious agent”) includes diseases caused by both bacte-

ria and viruses. In contrast, the WHO block only cov-

ers diseases caused by bacteria. Section 3.6 describes

our manual approach for merging such classes.

5 CONCLUSIONS AND FUTURE

WORK

In this paper we presented our approach for modeling

the hierarchy of the ICD-10 in OWL. Our model cap-

tures the hierarchical information of ICD-10 as well

as comprehensive class labels both for English and

German. Peculiarities such as “Exclusions,” state-

ments which make statements about the disjointness

of certain ICD-10 categories, are provided in a sepa-

rate OWL-Full component. This component allows

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL

Table 1: Metrics for the generated ICD-10 ontology.

WHO ICD-10 German ICD-10

OWL classes 11,308 16,214

disjointness axioms (see Section 3.4) 13,094 27,899

excludes pointing to another category 5,150 4,417

excludes without a proper link to other categories 35 73

Table 2: ICD-10 blocks, which can be merged together, although they exclusively appear in the DIMDI or WHO version.

WHO ICD-10 DIMDI ICD-10

D50-D89 Diseases of the blood and blood-forming

organs and certain disorders involving the immune

mechanism

D50-D90 Krankheiten des Blutes und der blutbilden-

den Organe sowie bestimmte Strungen mit Beteili-

gung des Immunsystems

D80-D89 Certain disorders involving the immune

mechanism

D80-D90 Bestimmte Strungen mit Beteiligung des

Immunsystems

V01-Y98 External causes of morbidity and mortality V01-Y84

Auere Ursachen von Morbiditt und Mor-

talitt

O80-O84 Delivery O80-O82 Entbindung

U80-U89 Bacterial agents resistant to antibiotics U80-U85 Infektionserreger mit Resistenzen gegen

bestimmte Antibiotika oder Chemotherapeutika

the use of hierarchical knowledge and class labels

with existing OWL-DL reasoners. Our automatic

generation and merging method also revealed system-

atic differences between the German DIMDI and the

English WHO version.

The goal of this approach was to partly reduce

Semantic Heterogeneity in health care, mentioned in

Section 1 by integrating two semi-formal terminolo-

gies. We plan to combine the results with addi-

tional conceptualizations next, so that ontology net-

works can be created and interconnected. For ex-

ample, Cardillo et al. describe an approach to map

the ICD with the International Classiﬁcation of Pri-

mary Care Version 2 (ICPC-2) (Cardillo et al., 2008).

This would foster the creation of expressive medical

knowledge bases which would improve the retrieval

and reuse of knowledge by reducing ambiguity.

The generated OWL ontology is used in the re-

search project MEDICO to allow for semantic anno-

tation and retrieval across medical documents and im-

ages annotated with ICD-10 terms both in English and

German.

The current ontology represents the main and

most important parts of both ICD-10 variants. How-

ever, the XML ﬁle states additional information to

be integrated in the future. For example, the dag-

ger and asterisks are distinguished depending on the

treatment. Different asterisk and dagger terms have

to be used to formulate diagnoses for clinical treat-

ments if diagnosis reports are created for accounting

documents. During our examination of the generated

OWL ﬁles and the respective ICD-10 websites, we

recognized that certain verbal structures occur very

often, e. g., “Injury of X,” where X is an anatomi-

cal designation, like arm or leg. Moreover, the man-

uals describe predeﬁned terms which are used very

often, for example, the two acronyms NOS, meaning

“not otherwise speciﬁed,” and NEC, standing for “not

elsewhere classiﬁed.” Linguistic analysis can exploit

this information to extract relations, e. g., to concepts

represented in anatomical ontologies. This approach

has already been applied successfully to other corpora

within the MEDICO project (Wennerberg et al., 2009;

Wennerberg, 2009) and we plan to extend it to the

ICD-10 as well.

Another possibility to enhance our OWL models

is to include more languages. This will generate an

international representation of the ICD-10. For exam-

ple, it would be possible to include the French ver-

sion of the ICD-10

. The website is structured like

the German DIMDI version. Consequently, either our

HTML crawler could be used to extract the necessary

information or–if a XML encoding is available–this

could be parsed directly. However, at ﬁrst appearance

one can see that the French version also differs from

the other two version, e. g., it covers only 21 chapters

http://www.dimdi.de/dynamic/en/klassi/diagnosen/

icd10/htmlfren/fr-icd.htm

KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development

and omits the chapter about “Codes for special pur-

poses.” Therefore, an examination of the differences

will be necessary.

ACKNOWLEDGEMENTS

This research has been supported in part by the THE-

SEUS Program in the MEDICO Project, which is

funded by the German Federal Ministry of Economics

and Technology under grant number 01MQ07016.

The responsibility for this publication lies with the au-

thors.

REFERENCES

Bechhofer, S., Volz, R., and Lord, P. W. (2003). Cooking

the semantic web with the OWL API. In International

Semantic Web Conference, pages 659–675.

Bodenreider, O. (2004). The Uniﬁed Medical Language

System (UMLS): Integrating biomedical terminology.

Nucleic Acids Research, 32 (Database Issue):D267–

D270.

Cardillo, E., Eccher, C., Seraﬁni, L., and Tamilin, A.

(2008). Logical analysis of mappings between med-

ical classiﬁcation systems. In Dochev, D., Pistore,

M., and Traverso, P., editors, AIMSA, volume 5253 of

Lecture Notes in Computer Science, pages 311–321.

Springer.

Cote, R., Rothwell, D., Palotay, J., Beckett, R., and Brochu,

L. (1993). The systematized nomenclature of human

and veterinary medicine. Technical report, SNOMED

International, Northﬁeld, IL: College of American

Pathologists.

Deutsche Krankenhausgesellschaft (2009). Deutsche

kodierrichtlinien - allgemeine und spezielle kodier-

richtlinien f

ur die verschl

usselung von krankheiten

und prozeduren. Technical report, Institut f

ur das Ent-

geltsystem im Krankenhaus (InEK GmbH).

Hoelzer, S., Schweiger, R. K., Liu, R., Rudolf, D., Rieger,

J., and Dudeck, J. (2002). Xml representation of hier-

archical classiﬁcation systems: from conceptual mod-

els to real applications. Proc AMIA Symp, pages 330–

Lenz, R. (2005). Information management in distributed

healthcare networks. Data Management In a Con-

nected World, 3551:315–334.

oller, M. and Mukherjee, S. (2009). Context-Driven On-

tological Annotations in DICOM Images – Towards

Semantic PACS. In Azevedo, L. and Londral, A. R.,

editors, Proceedings of the Second International Con-

ference on Health Informatics, HEALTHINF, pages

294–299. INSTICC Press.

oller, M., Sintek, M., Buitelaar, P., Mukherjee, S., Zhou,

X. S., and Freund, J. (2008). Medical image under-

standing through the integration of cross-modal object

recognition with formal domain knowledge. In Proc.

of HEALTHINF 2008, volume 1, pages 134–141, Fun-

chal, Madeira, Portugal.

Noy, N. F. and Rubin, D. L. (2008). Translating the Founda-

tional Model of Anatomy into OWL. Web Semantics:

Science, Services and Agents on the World Wide Web,

6(2):133–136.

Schulz, S., Suntisrivaraporn, B., and Baader, F. (2007).

SNOMED CT’s problem list: Ontologists’ and logi-

cians’ therapy suggestions. In Proc. of The Medinfo

2007 Congress, Studies in Health Technology and In-

formatics (SHTI-series). IOS Press.

Schulz, S., Suntisrivaraporn, B., Baader, F., and Boeker, M.

(2009). SNOMED reaching its adolescence: Ontolo-

gists’ and logicians’ health check. International Jour-

nal of Medical Informatics, 78(Supplement 1):S86–

S94.

Sirin, E., Parsia, B., Grau, B., Kalyanpur, A., and Katz, Y.

(2007). Pellet: A practical owl-dl reasoner. Web Se-

mantics: Science, Services and Agents on the World

Wide Web, 5(2):51–53.

Sonntag, D. (2010). Ontologies and Adaptivity in Dialogue

for Question Answering. AKA and IOS Press, Heidel-

berg.

Sonntag, D., Wennerberg, P., Buitelaar, P., and Zillner, S.

(2009). Pillars of ontology treatment in the medical

domain. Journal of Cases on Information Technology

(JCIT), 11(4):47–73.

Wennerberg, P. (2009). Aligning medical domain ontolo-

gies for clinical query extraction. In Proc. of the 12th

Conference of the European Chapter of the Associa-

tion for Computational Linguistics: Student Research

Workshop (EACL ’09), pages 79–87, Morristown, NJ,

USA. Association for Computational Linguistics.

Wennerberg, P., M

oller, M., and Zillner, S. (2009). A lin-

guistic approach to aligning representations of human

anatomy and radiology. In Proc. of the International

Conference on Biomedical Ontologies (ICBO 2009).

WHO (2004). International statistical classiﬁcation of dis-

eases and related health problems. Technical report,

World Health Organization.

REPRESENTING THE INTERNATIONAL CLASSIFICATION OF DISEASES VERSION 10 IN OWL