Semantic Information in Medical Information Systems:
Utilization of Text Mining Techniques to
Analyze Medical Diagnoses


Andreas Holzinger
(Research Unit HCI4MED, Institute of Medical Informatics, Statistics & Documentation
Medical University Graz, Austria
andreas.holzinger@meduni-graz.at)

Regina Geierhofer
(Research Unit HCI4MED, Institute of Medical Informatics, Statistics & Documentation
Medical University Graz, Austria
regina.geierhofer@meduni-graz.at)

Felix Mödritscher
(Institute for Information Systems and New Media
Vienna University of Economics and Business Administration, Austria
felix.moedritscher@wu-wien.ac.at)

Roland Tatzl
(IT and IT Marketing, Campus 02
Graz University of Applied Sciences, Austria
roland.tatzl@campus02.at)



Abstract: Most information in Hospitals is still only available in text format and the amount of
this data is immensely increasing. Consequently, text mining is an essential area of medical
informatics. With the aid of statistic and linguistic procedures, text mining software attempts to
dig out (mine) information from plain text. The aim is to transform data into information.
However, for the efficient support of end users, facets of computer science alone are
insufficient; the next step consists of making the information both usable and useful.
Consequently, aspects of cognitive psychology must be taken into account in order to enable
the transformation of information into knowledge of the end users. In this paper we describe the
design and development of an application for analyzing expert comments on magnetic
resonance images (MRI) diagnoses by applying a text mining method in order to scan them for
regional correlations. Consequently, we propose a calculation of significant co-occurrences of
diseases and defined regions of the human body, in order to identify possible risks for health.

Keywords: Information Retrieval, Text Mining, Performance, Medical Documentation
Categories: H.3.1, H.3.3, I.2.7, I.7, J.3
1 Introduction and Motivation
Significant progress has been made in the last years in the application of text mining
techniques, in order to cope with the rapidly increasing information overload in the
area of medical literature [Sullivan et al., 1999], [Hall and Walton, 2004]. However,
Journal of Universal Computer Science, vol. 14, no. 22 (2008), 3781-3795
submitted: 20/9/08, accepted: 18/11/08, appeared: 28/12/08 © J.UCS

developments for text mining techniques in the area of clinical information systems
and medical documentation are rare [Noone et al., 1998], [Holzinger et al., 2007b].
The broad application of sophisticated medical information systems amasses large
amounts of medical documents, which must be reviewed, observed and analyzed by
human experts [Holzinger et al., 2007a]. All essential documents of the patient
records contain at least a certain portion of data which has been entered in free-text
fields and has long been in the focus of research [Gell et al., 1976], [Gell, 1983],
[Zingmond and Lenert, 1993].
Although text can be created simple by the end-users, the support of automatic
analysis is extremely difficult [Gregory et al., 1995], [Holzinger et al., 2000], [Lovis
et al., 2000]. It is likely that some interesting and relevant relationships remain
completely undiscovered, due to the fact that relevant data are scattered and no
investigator has linked them together manually, an example from medical literature
can be found in [Smalheiser and Swanson, 1998]. With regard to the fact that textual
information is definitely an extremely important part of medical documents and that
the amount of textual information is rather increasing then decreasing in future, the
Institute for Medical Informatics, Statistics and Documentation (IMI) of the Medical
University Graz (www.meduni-graz.at/imi) has been carrying out a variety of projects
to analyze and process medical documents by application of computer-based
techniques and to present this information in an end-user centered manner.
2 Theoretical Background
2.1 Text Mining: Definition
Text mining, sometimes called textual data mining is generally defined as a
knowledge-intensive process in which end users interact with a collection of textual
information by using analytic tools; typical text mining tasks include categorization,
clustering, concept extraction, production of taxonomies, sentiment analysis,
document summarization and entity-relation modeling [Feldman and Sanger, 2007].
The major challenge of biomedical text mining over the next years is to make such
systems useful to biomedical researchers. Certainly, this will require enhanced access
to full text, better understanding of the feature space of biomedical text, enhanced
methods for measuring the usefulness of systems to end users and continued
cooperation with the biomedical research community to ensure that their needs and
requirements are appropriately addressed [Cohen and Hersh, 2005].
Contrary to structured information, textual information is characterized by its
inherently unstructured and fuzzy nature, being language and domain dependent as
well as consisting of sentences and sub-sentences [Sistrom and Honeyman-Buck,
2005]. Text represents factual information in a complex, rich, and opaque manner –
which makes it difficult to be analyzed by standard statistical data mining methods:
relying on human analysis results in either huge workloads or the analysis of only a
tiny fraction of the database [Nasukawa and Nagano, 2001].
Consequently, text mining stands for a multidisciplinary field involving
“information retrieval, text analysis, information extraction, clustering,
categorization, visualization, database technology, machine learning and data
mining” [Tan, 1999].
3782 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

Techniques in the field of text mining aim at the extraction of coherences within
such unstructured information. Basically, text mining approaches apply statistical or
pattern-based algorithms [Biemann et al., 2004], in order to extract significant key-
word associations or to mine for prototypical documents (e.g. for part-of-speech
tagging or term extraction) [Rajman and Besancon, 1998].
2.2 Text Mining: Knowledge Discovery from Data (KDD)
Text mining is considered a sub-specialty of Knowledge Discovery from Data (KDD)
and has been headed strongly in the direction of Natural Language Processing (NLP)
in the last decade [Liddy, 2000]. According to [Granitzer, 2006], the following phases
of the text mining process can be identified:

(1) Pre-processing: As textual data is primarily encoded within documents of different
formats (XML, HTML, Word, etc.), lexical analysis methods can be applied to
prepare the texts, e.g. by lexical analysis and removal of layout information. Further,
this phase also allows the improvement of the text quality by utilizing methods for
feature analysis, such as stemming or part-of-speech tagging.

(2) Information extraction: This stage aims at transforming the unstructured text
entities into structured elements, e.g. through named entity recognition, co-reference
resolution, template element filling, scenario templates and others. Therefore,
techniques of machine learning or linguistic analyses are applied, whereby the quality
of the outcome is highly dependent on the domain and the correctness of the phrasing.

(3) Feature generation and statistical analyses: By applying statistical methods,
features of an information space can be extracted from the structured elements of a
text. Hereby, methods such as frequency analyses, collocations or co-occurrences (of
words) are utilized. [Zipf, 1949] examined the connection between frequency and
significance of words within a text corpora, which led to Zipf’s law as a widely used
basis for analyzing word frequencies in texts. Zipf's law states that the frequency of
words in a large corpus is inversely proportional to their rank [Le Quan et al., 2002].
Moreover, also word pairs (co-occurrences) and word groups (collocations) are of
interest for generating features from a text.

(4) Operations on feature spaces: As a result, each text object is described with
several features, which, for instance, spans an n-dimensional vector space with n
equals to the number of features for a text object. This feature space can be utilized
for further operations, e.g. comparing texts, visualizing relations in the text corpus,
calculating similarities or rankings, clustering, etc.
2.3 Text Mining in Medical Documentation
According to [Hotho et al., 2005], text mining approaches can be useful for many
different application scenarios, such as patent search, text classification for news
agencies, anti-spam filtering of emails, bioinformatics, etc.
Text mining in medical documentation is an extremely important area [Buenaga
et al., 2006], however, it encompasses a lot of various problems [Gell, 1983]. In this
3783Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

paper, we report about our experiences on utilizing text mining techniques for medical
documentation.
3 Related Work
In the last years, mining in medical digital libraries has come up with new findings
and hypotheses. [Srinivasan, 2004] reports on the development of text mining
methods on the basis of medical subject headings (MeSH). These algorithms generate
hypotheses by identifying potentially interesting terms related to the specific input.
Additionally, new protein associations have been found by clustering learning and
vector classification techniques [Fu et al., 2003]. Another interesting approach utilizes
a rule-based parser and co-occurrence for extracting and combining relations [Leroy
et al., 2003]. Further, a new way to use thalidomide has been discovered by mapping
phrases to concepts of the Unified Medical Language System [Weeber et al., 2001a].
Finally, co-occurrences are useful to build up gene networks [Jenssen et al., 2001] and
to discover gene interactions [Stephens et al., 2001].
Derived from these experiences, three kinds of application areas for text mining
can be outlined in the field of medical documentation:
- Firstly, such methods are applied in order to build up an infrastructure or models
for biomedicine, i.e. by finding patterns or relations in texts and generating a feature
space. Inspecting the projects on the basis of the well-known text mining framework
GATE (http://gate.ac.uk/projects.html), MultiFlora or myGRID can be identified as
examples for this approach.
- Secondly, text mining is used to observe and retrieve documents with innovative
ideas in the scope of a restricted domain (cf. projects such as BioRAT or InESBi).
- Thirdly, text mining techniques are utilized to extract information or features
out of a medical text corpus, which is generated as product of clinical documentation
for further operations such as information retrieval. The MedDictate software
comprises one solution in this scope. Although this last category of applications areas
overlaps partially with the first one, there are only a few reports about the mining of
medical diagnoses. In addition to these experiences, we want to report on our
approach towards aiming at the detection of diseases in MRT diagnoses. In the
following two sections, this project and its outcome are described in detail.
4 Methods and Materials
The success of text mining methods in medical research was the origin of our idea to
apply statistical techniques in order to find hints for possible locations of diseases in
MRT diagnoses. Therefore, we aimed at the topological proximity between anatomic
structures and pathologic expressions and implemented a tool which calculates the
significant co-occurrences of anatomic and pathologic terms within the diagnoses.
4.1 Text Corpus and Pre-Processing
Accompanying measures during the evaluation of an information extraction tool
revealed the need for additional assistance in finding topological relations between
3784 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

anatomic structures including regional indicators and diseases such as tumors. Given
these requirements, there was a demand for domain specific databases for anatomic
structures and pathologic expressions.
At the beginning of this project, we used a text corpus of about 6.000 diagnoses,
which comprises comments of medical experts on magnetic resonance imaging (MRI)
material. Table 1 shows the distribution of diagnoses per year.

Year Number of diagnoses
1987 1456
1988 2143
1989 1714
1990 52
2003 331
2004 313
Table 1: Diagnoses per year
These findings derived from diverse radiologists and are completely written in capital
letters and packed with medical terms. Hence, we were confronted with three serious
problems: synonyms, medical dialects and abbreviations [Weeber et al., 2001b].
Further, the diagnoses are spread over a period of 17 years. As a consequence, time-
dependent changes of terminology might be possible. These textual diagnoses were
made anonymous and imported into a database. A distinction between non-important
free text and biomedical entities as proposed by [Chen et al., 2005], p.20, does not
seem to be necessary because of the pure provenience of the diagnostic corpus.
4.2 Information extraction
Considering the systemic performance and the sentence-based statistical calculation,
the texts were also split up into sentences. Pre-processing of the free text is realized in
a way, that the occurrences of expression pairs are counted and stored. Sentences have
been identified by break-iterators of the Jakarta Group. In addition a length smaller
then 15 characters has not been accepted as a complete sentence. Following this
methodology, we identified 15731 sentences. Consequently, the calculation of the co-
occurrences can now be executed by means of one of the three formulas which are
shown in figure 1.
Performance issues demanded a full-text indexing of the diagnoses as well as a
reduction of the anatomic terms. Unfortunately, two or three character words are
found in many other words, consequently they must be excluded from the reference
database.
Additionally, we were in need of anatomic and pathologic terms for our approach.
A corpus of approximately 6.800 anatomic structures was generated from an anatomic
dictionary [Dauber, 2005]. This dictionary (in German) offers a rough allocation of
anatomic structures to anatomic regions. More precision in finding such structures can
be reached by using synonyms, however, the gathering of which is extremely time
consuming. Efforts have been made to start with a synonym enhancement for the
anatomic data at the IMI, which has been used for the calculation. On the other hand,
3785Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

a Pathology database has been set up manually due to a lack of accessible resources.
These corpuses represent the domain specific database sources for the statistic
calculation and can be maintained in special application modules.
4.3 Feature extraction
Our methodology for the identification of proximity is based on the calculation of
significant co-occurrence. This term describes the cumulative joint appearance of
pairs of words in a defined environment which, in our case, is determined to be a
sentence.
4.3.1 Basic Algorithm and Methodology
The calculation of significant co-occurrences is based on the Poisson distribution. In
accordance with [Biemann et al., 2004], [Heyer et al., 2006], the original formula can
be simplified for two different ranges of the input parameters (see also figure 1).
Hereby, a stands for the number of sentences containing term A, b for the number
of sentences containing term B, n for the number of all sentences and k for the number
of sentences containing both terms.

Be
n
ba ⋅
=λ then:
n
i
e
BAsig
k
i
i
log
!
1
1log
),(
1
0
⎟⎠
⎞⎜⎝
⎛
⋅−−
=
∑−
=
− λλ


If ( ) 5.21 >+λ
k then:
n
kk
BAsig
log
!loglog
),(
+⋅−
≈
λλ

If 10>k then:
n
kk
BAsig
log
)1log(log
),(
−−⋅
≈
λ
Figure 1: The three formulas to calculate significant co-occurrences of two terms
Due to performance reasons, we decided to implement the calculation of the
significant co-occurrences independently, instead of re-using existing text mining
modules. The different ways to calculate the co-occurrence allow the usage of the
fastest algorithm for each diagnose, as the simplified formulas (the last two in the
figure) require less time and processing power.
Listing 1 shows a JAVA code example of the calculation class which handles all
of the required steps for the statistical analysis.

1 /**
2 * fact calculates a mathematical function in the formula.
3 */
4 static double fact(double n) {
5 if (n <= 0)
6 return 0;
7 if (n == 1)
3786 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

8 return 1;
9 else
10 return (n * fact(n - 1));
11 }
12
13 /**
14 * sigsum sums up a function in the formula.
15 */
16 static double sigsum(double k, double lambda) {
17 double result =0;
18 System.out.println("k: "+k);
19 System.out.println("lambda: "+lambda);
20 for(double i=1; i<=(k-1);i++){
23 result += Math.pow(lambda,i)/fact(i);
24 }
25 return result;
26 }
27
28 /**
29 * sig calculates the significance for the given parameters.
30 */
31 static double sig(double a, double b, double k, double n){
32 double lambda = (double)(a*b)/n;
33 double decision = (k+1)/lambda;
34 double sig =0;
35 System.out.println("func lambda: "+lambda);
36 System.out.println("func decision: "+decision);
37 System.out.println("sigsum: "+sigsum(k,lambda));
38 if (decision > (double)2.5){
39 if (k > (double)10){
40 System.out.println("k>10");
41 sig = k*(Math.log(k)-Math.log(lambda)-1)/Math.log(n);
42 } else{
43 System.out.println("k<10");
44 sig = (lambda-k*Math.log(lambda)+Math.log(fact(k)))/Math.log(n);
45 }
46 } else{
47 System.out.println("dec<2.5");
48 if (k < 200){
49 sig = -Math.log(1-Math.pow(Math.E,-lambda)*sigsum(k,lambda))/
Math.log((double)n);
50 } else{
51 System.out.println("Math.log(k): "+Math.log(k));
52 System.out.println("Math.log(lambda): "+Math.log(lambda));
53 //System.out.println("Math.log(n): "+Math.log(n));
54 sig = k*(Math.log(k)-Math.log(lambda)-1)/Math.log(n);
55 }
56 //sig = k*(Math.log(k)-Math.log(lambda)-1)/Math.log(n);
57 }
58 return sig;
59 }
Listing 1: Code example from class DBMANAGER.JAVA
3787Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

4.3.2 Performance issues
Of course we had to consider pre-processing steps of the MRI diagnoses in order to
complete calculations on the initial text corpus within a reasonable period of time.
These steps include a pre-selection of anatomic or pathologic expression which occur
in the diagnostic corpus to speed up calculation later on.
4.4 The Web Application
A basic end user access control system is used for logging purposes. Maintenance
operations must be executed as administrator. End user actions include registration,
editing of the registration information, login, logout and observing the login history.
The application itself offers modules for the following functionality:
• Diagnoses can be listed and filtered according to two terms;
• Anatomic terms can also be filtered;
• The location for each anatomic term is indicated;
• The synonyms module shows all available synonyms for each anatomic
term. These synonyms cover small parts of terms, but will increase in future;
• The menu option “ADD PATHOLOGY” provides a dialogue to add a term;
• Significant co-occurrents are listed in module COOCCURRENTS.
Additionally new calculations can be started;
• A maintenance module provides splitting of the diagnoses as well as
calculation of the occurrence of the single terms;
4.5 Core Functions and Advantages
For performance purposes, the administrator can initiate a pre-calculation of the
occurrence of each single term. Thus the number of sentences and anatomic terms for
the statistic calculation can be reduced. The splitting of the diagnoses is the second
method of improving performance during the calculation. Additionally the split
sentences are reduced by excluding sentences with a character length < 15. Thereby
abbreviation sentences are most likely eliminated.

3788 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

Figure 2: Filtering diagnoses according by the terms “TUMOR” and “KLEINHIRN”
Search for topological relations is based on MRI diagnoses in a heterogeneous
context (e.g. cranial and spinal MRI diagnoses), as visualized in figure 2. The table of
diagnoses can be filtered manually and retrieved in a list. Simple IR-techniques, such
as query term highlighting, are used to visualize the results. These functions support
medical experts on comparing the results for the calculated pairs of terms.
Anatomic and pathologic terms can be edited in separated dialogues. Because of a
lack of a Pathology reference corpus, authoring functionality has been implemented
for the application in order to manually add, modify or remove expressions. The
processing of the medical free text is based on the formulas described in subsection
3.1. In order to improve the overall performance during calculation, the query
considers only sentences which contain any of the anatomic or pathologic terms and
only terms which were previously identified in the diagnose corpus.

3789Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

Figure 3: Tough graph visualization for the pathological term “GLIOBLASTOMA”
The resulting pair-list is shown in a table in descending order of significance for
each expression. For further investigations, the most promising pairs can be used to
filter the diagnoses in order to evaluate the results. For visualization purposes, a
Touchgraph applet was implemented which enables the medical expert to estimate the
proximity at a glance (see figure 3). In addition to the basic relations between the
pathologic expression and the anatomic structures, the interconnection significance
among the anatomic structures is calculated, to show their potential proximity within
the corpus of diagnoses. This visualization clearly shows the topologic relationship.
5 Results and Experiences
The results of the calculation must be analyzed in their special context. Each pair of
terms implicates specific anatomic-pathologic issues, and, in addition, the meaning of
the co-occurrence must be interpreted individually.
5.1 Results for an Exemplary Scenario
The results are discussed on the basis of the malign brain cancer glioblastoma, which
generally occurs in the group of middle-aged men and is located most likely in the
corpus callosum and the temporal lobe. According to [Schlegel et al., 2003] and
[Hopf et al., 1999] the percentage of glioblastoma in all intracranial tumors is 15 to
20%. In addition to the location information the age of the patients could easily be
retrieved inside the clinical information system in order to check the occurrence of the
tumor in male or female age history. The peak incidence of glioblastoma is reported
to be in middle adult life (56 to 60 years) but no age group is known to be exempt.
The incidence is higher in men (ratio of approximately 1.6:1) [Ropper & Brown,
05].
3790 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

An analysis in a linguistic database (http://wortschatz.uni-leipzig.de) showed a
very low occurrence in common language sources, such as newspaper articles and
books, and emphasizes the importance of a well maintained domain specific database.
The result set for the calculation (refer again to figure 3) showed 10 pairs of terms
for glioblastoma including synonyms. Finally, figure 4 shows the detailed result set
for glioblastoma.


Figure 4: Calculation details for the Glioblastoma result set
We emphasize, that a highly significant co-occurrence does not prove the affliction of
an anatomic structure with the paired disease. Also, a co-occurrence suggests a
relation for further investigation. The most significant result suggests the co-
occurrence with TEMPORA and SPLENIUM, whereby both locations are well
known and located in the most significant decile of the result set. The distribution of
the result set is shown in figure 5. The threshold for the upper decile is 0.564.

Figure 5: Statistical distribution of the result set
3791Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

The second group of results does not make sense, because the associated anatomic
terms CORPUS or EPENDYMA (thin epithelial membrane lining the ventricular
system of the brain and the spinal cord canal) are too common and do not implicate
worthwhile location information. The third group showed a combination of terms
which are less well known and might be an interesting hint for further investigation.
5.2 Discussion of Experiences and Problematic Aspects
The resulting set of the calculations are listed in descending order according to the
significance. The overall statistical evaluation says that the upper decile contains the
most significant co-occurrences with a high probability. Some of the less significant
results include interesting combinations of terms which may point out valuable new
conclusions. Finally, syntax highlighting enables the observer to find the terms of the
query quite easily in the result set of diagnoses. Despite such experiences, we also
identified the following problematic aspects for applying statistical text mining
techniques on diagnoses: (1) The group of results which are not useful is an evidence
of the weakness of the anatomic reference corpus. Expressions with a too widespread
meaning should not be considered in order to reduce the wrong results. (2) The
amount of diagnoses must be increased. Examples from professional common
language research show that about 5 million sentences or more are required to
validate our approach. (3) The diagnoses are specific for responsible radiologists.
Therefore, they are not thoroughly comparable, as different experts tend to use other
terminologies. (4) Words with only few characters (like “OS” or “COR”) are not
suitable for searching purposes. Thus, they have to be eliminated. (5) Acronyms and
abbreviations (dotted) cause difficulties when splitting the diagnoses into sentences.
5.3 Remedies for the problems
As a result the future work will include a reduction of useless words in the anatomic
reference corpus by implementing a relevance ranking. The increase of the amount of
diagnoses is no problem at all except the protection of privacy. In order to identify the
differences between radiologists additional calculation strategies must be introduced.
5.4 Opportunities
Nevertheless, we find that our methodology for analyzing medical diagnoses
comprises a promising approach: The analyzing of medical text corpora is fully
computer driven and fully automated. The overall calculations require from a few
minutes up to some hours, depending of the computational hardware and the amount
of data. Thus, this method can be applied on text corpora at any time, e.g. to use other,
more accurate anatomic and pathologic expressions for old diagnoses. Secondly, our
tool is of interest for clinical professionals in order to support them at their daily
tasks, for instance during pre-analyzing diagnoses. We also implemented some
functions to support general medical experts in their daily work in order to reduce
their cognitive load (see subsection 3.3). Finally, this kind of text mining algorithm
could be also valuable for other application areas.
3792 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

6 Conclusion
We emphasize the importance of computer-based methods in medical documentation
and the automatization of clinical processes, including analyzing diagnoses. In this
context, we developed a methodology for text mining in medical text corpora and
implemented a tool to evaluate our idea. The outcome of the calculations showed
valuable results although based on a relatively low number of sentences. Observing
all the diagnoses, generated in a hospital daily, will definitely improve the diagnostic
value. However, we still have no proof that the anatomic structure is affected by the
related disease, however, our experiences encouraged us to carry out further research
efforts on these co-occurrences. Mining in large amounts of textual medical
information can reveal new patterns for various questions. There will be a continued
need for new mining assistants solving problems which are not even known today.
We identified a huge benefit for the administration in identifying trends and
developments in time to come to the appropriate decisions. The appropriate
information presentation to the end users is a central future challenge, in order to keep
their cognitive load in an optimum level, providing cognitive performance support.
References
[Biemann et al., 2004] Biemann, C., Bordag, S., Heyer, G., Quasthoff, U. and Wolff, C.:
"Language-independent methods for compiling monolingual lexical data", Computational
Linguistics and Intelligent Text Processing, Springer-Verlag Berlin, Berlin, 2004, 217-228.
[Buenaga et al., 2006] Buenaga, M., Maña, M., Gachet, D. and Mata, J.: "The SINAMED and
ISIS Projects: Applying Text Mining Techniques to Improve Access to a Medical Digital
Library", Research and Advanced Technology for Digital Libraries, 2006, 548-551.
[Chen et al., 2005] Chen, H., Fuller, S., Friedman, C. and Hersh, W.: "Medical Informatics:
Knowledge Management And Data Mining in Biomedicine", Springer, Berlin, Heidelberg,
New York, (2005).
[Cohen and Hersh, 2005] Cohen, A. M. and Hersh, W. R.: "A survey of current work in
biomedical text mining", Briefings in Bioinformatics, 6, (2005), 57-71.
[Dauber, 2005] Dauber, W.: "Feneis’ Bild-Lexikon der Anatomie, 9th Edition", Thieme,
Stuttgart, (2005).
[Feldman and Sanger, 2007] Feldman, R. and Sanger, J.: "The Text Mining Handbook:
Adavanced Approaches in Analyzing Unstructured Data ", Cambridge University Press,
Cambridge (2007).
[Fu et al., 2003] Fu, Y., Mostafa, J. and Seki, K.: "Protein association discovery in biomedical
literature", Proceedings of the 3rd ACM/IEEE-CS joint conference on Digital libraries, (2003),
113-115.
[Gell, 1983] Gell, G.: "AURA - routine documentation of medical text", Methods of
Information In Medicine, 22, (1983), 63-68.
[Gell et al., 1976] Gell, G., Oser, W. and Schwarz, G.: "Experiences with the AURA Free Text
System", Radiology, 119, (1976), 105-109.
3793Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

[Granitzer, 2006] Granitzer, M.: "KnowMiner: Konzeption und Entwicklung eines generischen
Wissenserschliessungsframeworks", PhD Thesis TU Graz, Graz, (2006).
[Gregory et al., 1995] Gregory, J., Mattison, J. E. and Linde, C.: "Naming Notes - Transitions
from Free-Text to Structured Entry", Methods of Information in Medicine, 34, (1995), 57-67.
[Hall and Walton, 2004] Hall, A. and Walton, G.: "Information overload within the health care
system: a literature review", Health Information and Libraries Journal, 21, (2004), 102-108.
[Heyer et al., 2006] Heyer, G., Quasthoff, U. and T., W.: "Text Mining: Wissensrohstoff Text:
Konzepte, Algorithmen, Ergebnisse", W3L, Bochum, (2006).
[Holzinger et al., 2007a] Holzinger, A., Geierhofer, R. and Errath, M.: "Semantic Information
in Medical Information Systems - from Data and Information to Knowledge: Facing
Information Overload", Procedings of I-MEDIA '07 and I-SEMANTICS '07, Graz, 2007a, 323-
330.
[Holzinger et al., 2007b] Holzinger, A., Geierhofer, R. and Errath, M.: "Semantische
Informationsextraktion in medizinischen Informationssystemen", Informatik Spektrum, 30,
(2007b), 69-78.
[Holzinger et al., 2000] Holzinger, A., Kainz, A., Gell, G., Brunold, M. and Maurer, H.:
"Interactive Computer Assisted Formulation of Retrieval Requests for a Medical Information
System using an Intelligent Tutoring System", ED-MEDIA 2000, Montreal, (2000), 431-436.
[Hopf et al., 1999] Hopf, H. C., Deuschl, G., Diener, H. C. and Reichmann, H.: "Neurologie in
Praxis und Klinik", Thieme, Stuttgart, (1999).
[Hotho et al., 2005] Hotho, A., Nürnberger, A. and Paaß, G.: "A Brief Survey of Text Mining",
GLDV-Journal for Computational Linguistics and Language Technology, 20, (2005), 19-62.
[Jenssen et al., 2001] Jenssen, T. K., Laegrid, A., Komorowski, J. and Hovig, E.: "A literature
network of human genes for highthroughput analysis of gene expression", Genetics, 28, (2001),
21-28.
[Le Quan et al., 2002] Le Quan, H., Sicilia-Garcia, E. I., Ji, M. and Smith, F. J.: "Extension of
Zipf's law to words and phrases", Proceedings of the 19th international conference on
Computational linguistics - Volume 1, (2002), 1-6.
[Leroy et al., 2003] Leroy, G., Chen, H., Martinez, J. D., Eggers, S., Flasey, R. R., Kislin, K.
L., Huang, Z., Li, J., Xu, J., McDonald, D. M. and Ng, G.: "Genescene: Biomedical text and
data mining, Proceedings of the ACM/IEEE-CS joint conference on Digital Libraries", (2003),
116-118.
[Liddy, 2000] Liddy, E. D.: "Interview with Gayle Curtis, Modem Media", (2000), online
available: http://www.asis.org/Bulletin/Oct-00/liddy.html, last access: 2008-05-15
[Lovis et al., 2000] Lovis, C., Baud, R. H. and Planche, P.: "Power of expression in the
electronic patient record: structured data or narrative text?", International Journal of Medical
Informatics, 58, (2000), 101-110.
[Nasukawa and Nagano, 2001] Nasukawa, T. and Nagano, T.: "Text analysis and knowledge
mining system", IBM SYSTEMS JOURNAL, 40, (2001), 967-984.
[Noone et al., 1998] Noone, J., Warren, J. and Brittain, M.: "Information overload:
opportunities and challenges for the GP's desktop", Medinfo, 9, (1998), 1287-1291.
[Rajman and Besancon, 1998] Rajman, M. and Besancon, R.: "Text Mining: Natural Language
Techniques and Text Mining Applications", In: Spaccapietra, S. and Maryansky, F., (eds.),
3794 Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...

Data mining and reverse engineering: Searching for semantics, Chapmann and Hall, London,
1998, 50-64.
[Schlegel et al., 2003] Schlegel, U., Weller, M. and Westphal, M.: "Neuroonkologie", Thieme,
Stuttgart, (2003).
[Sistrom and Honeyman-Buck, 2005] Sistrom, C. L. and Honeyman-Buck, J.: "Free text versus
structured format: Information transfer efficiency of radiology reports", American Journal of
Roentgenology, 185, (2005), 804-812.
[Smalheiser and Swanson, 1998] Smalheiser, N. R. and Swanson, D. R.: "Using
ARROWSMITH: a computer-assisted approach to formulating and assessing scientific
hypotheses", Computer Methods and Programs in Biomedicine, 57, (1998), 149-153.
[Srinivasan, 2004] Srinivasan, P.: "Text mining: Generating hypotheses from MEDLINE",
Journal of the American Society for Information Science and Technology, 55, (2004), 396-413.
[Stephens et al., 2001] Stephens, M., Palakal, M., Mukhopadhyay, S., Raje, R. and Mostafa, J.:
"Detecting gene relations from Medline abstracts", Pacific Symposium on Biocomputing,
(2001), 483-496.
[Sullivan et al., 1999] Sullivan, F., Gardner, M. and Van Rijsbergen, K.: "An information
retrieval service to support clinical decision-making at the point of care", British Journal of
General Practice, 49, (1999), 1003-1007.
[Tan, 1999] Tan, A.-H.: "Text Mining: The state of the art and the challenges", PAKDD 1999
Workshop on Knowledge Discovery from Advanced Databases, Beijing, (1999), 65-70.
[Weeber et al., 2001a] Weeber, M., Klein, H., Berg, L. and Vos, R.: "Using concepts in
literature-based discovery: Simulating Swanson’s Raynaud-Fish Oil and Migraine-Magnesium
discoveries", Journal of the American Society for Information Science and Technology, 52,
(2001a), 548-557.
[Weeber et al., 2001b] Weeber, M., Mork, J. G. and Aronson, A. R.: "Developing a test
collection for biomedical word sense disambiguation", Journal of the American Medical
Informatics Association, (2001b), 746-750.
[Zingmond and Lenert, 1993] Zingmond, D. and Lenert, L. A.: "Monitoring Free-Text data
using medical language processing", Computers and Biomedical Research, 26, (1993), 467-
481.
[Zipf, 1949] Zipf, G. K.: "Human Behaviour and the Principle of Least-Effort", Addison-
Wesley, Cambridge (MA), (1949).


3795Holzinger A., Geierhofer R., Moedritscher F., Tatzl R.: Semantic Information ...