(Projektbeschreibung nur in Englisch)
M. Schuerz1, K.-P. Adlassnig1, C. Lagor2, B. Schneider3, G. Grabner1
| 1 | Department of Medical Computer Sciences,
Section on Medical Expert and Knowledge-Based Systems,
University of Vienna Medical School,
Spitalgasse 23, A-1090 Vienna, Austria
e-mail: michael.schuerz@akh-wien.ac.at |
| 2 | Department of Medical Informatics,
The University of Utah, Salt Lake City, USA e-mail: Charles.Lagor@m.cc.utah.edu |
| 3 | Department of Medical Statistics and Documentation,
University of Vienna Medical School, Schwarzspanierstrasse 17, A-1090 Vienna,
Austria e-mail: Barbara.Schneider@univie.ac.at |
| 4 | Professor emeritus of the Second Department for
Gastroenterology and Hepatology and of the Department of Medical Computer Sciences,
University of Vienna Medical School e-mail: Georg.Grabner@teleweb.at |
Background
Knowledge acquisition and representation are one of the central challenges for the successful
construction and use of medical expert and knowledge-based systems in clinical
practice. Since medical knowledge is immanently vague over wide ranges, fuzzy sets are
used to deal with uncertain linguistic medical concepts such as reduced, normal, elevated,
and highly elevated. During the diagnostic and therapeutic process observed symptoms,
signs, and test results are converted into fuzzy compatibility values reaching from
zero to unity with the linguistic medical concepts under consideration in a data-to-symbol
conversion component of a medical expert system. Then these “fuzzified” findings are
applied to establish confirmed and excluded diagnostic hypotheses by means of a knowledge base,
which contains fuzzy relationships — generated by fuzzy relations — for the
frequency of occurrence of findings with diseases and the strength of confirmation of
findings for diseases.
Objective
The aim of this study was to construct fuzzy sets in the form of S- and pi-shaped fuzzy
membership functions semi-automatically from data of one reference group of healthy
individuals and six disease groups of patients suffering from liver diseases for the observed
laboratory parameters for the linguistic medical concepts reduced, normal, elevated,
and highly elevated. After this, fuzzy relations were formed between linguistic
medical concepts and diagnoses in the field of hepatitis diseases [1].
Material and Methods
Laboratory test results from sample patients from reference groups of healthy individuals
and from the following hepatitis groups were investigated: type A hepatitis, type B hepatitis,
type C hepatitis, chronic hepatitis, alcoholic hepatitis, and hepatitis B carriers with values for
the laboratory parameters bilirubin, alanine aminotransferase, aspartate aminotransferase
gamma-glutamyltranspeptidase, alkaline phosphatase, lactate dehydrogenase, and for the
electrophoresis laboratory parameters albumin, alpha 1 globulin, alpha 2 globulin, beta
globulin, and gamma globulin. Considering the reference groups of the previously mentioned
laboratory parameters fuzzy membership functions were constructed for the linguistic
medical concept normal with non-parametrical distribution-free statistical methods. By this the
minimum, the 5%-percentile, the 95%-percentile, and the maximum were calculated from
data and used as parameters for S- and pi-shaped fuzzy membership functions [2].
In the next step fuzzy membership functions for reduced and elevated were formed. Since
these concepts are complementary to the concept normal, the corresponding S-
membership functions of normal were reflected on their inflection points to obtain the fuzzy
membership functions of reduced and elevated. Finally a differentiation of a subset highly
elevated from elevated could be obtained. An iterative gradient method using ordered sets
of maxima and 95%-percentiles from the diseases and the reference groups was applied to
find a threshold between elevated and highly elevated. From this, fuzzy relations between
uncertain findings and diagnoses were generated for the frequency of occurrence of findings
with diseases and the strength of confirmation of symptoms for diseases using relative
sigma-counts [3]. In this case these relative sigma-counts are the conditional probabilities
P(Dj /Si) and P(Si /Dj) with uncertain
(fuzzy) events Si and Dj with Si representing the
fuzzy compatibility value of a finding of a specific patient and Dj representing the
corresponding fuzzy compatibility value of a diagnosis of a specific patient.
Results
Based on laboratory test results, 41 fuzzy sets for the representation of the linguistic
medical concepts reduced, normal, elevated, and highly elevated were obtained for the
test results of the eleven above-mentioned laboratory and electrophoresis parameters (Fig. 4).
| Fig. 4: | Fuzzy membership functions for reduced (µ¯ ), normal (µ^), elevated (µ), and highly elevated (µ ) for gamma globulin. |
From this, 574 fuzzy relations for the generation of the frequency of occurrence of the
linguistic medical concepts reduced, normal, elevated, and highly elevated for every
above-mentioned laboratory parameter with the hepatitis and reference groups as well as
the strength of confirmation of the previously mentioned findings for the hepatitis and
reference groups were calculated (to be applied during the inference process of medical
expert systems) (Tab. 1). Hereby fuzzy relations with values equal zero as well as those
equal one need to be checked by a physician according to their frequency of occurrence
respectively strength of confirmation in the parent population since these values are calculated
from sample data and hence may contain biases.
| type A hepatitis | type B hepatitis | reference group | ||||
|---|---|---|---|---|---|---|
| µO | µC | µO | µC | µO | µC | |
| gamma globulin | ||||||
| reduced | 0.00 | 0.00 | 0.01 | 0.00 | 0.02 | 0.48 |
| normal | 0.40 | 0.03 | 0.64 | 0.00 | 0.96 | 0.34 |
| elevated | 0.60 | 0.12 | 0.35 | 0.00 | 0.02 | 0.02 |
| highly elevated | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Tab. 1: | Fuzzy relations between the uncertain symptoms reduced (µ¯ ), normal (µ^), elevated (µ), and highly elevated (µ ) for gamma globulin with type A hepatitis, type B hepatitis, and the reference group. The fuzzy relation µO indicates the frequency of occurrence of findings with diseases and µC indicates the strength of confirmation of findings for diseases. |
Technical Specification
Laboratory test results and diagnoses were extracted from the data base of the hospital
information system WAMIS [2] that runs on a IBM 2003 under VSE/ESA with CICS and
transferred to Microsoft Excel 97 for calculation of the corresponding parameters for the
S- and pi-shaped fuzzy membership functions as well as to calculate fuzzy relations.
Turbo Pascal 6.0 was used to write programs for the optimization of fuzzy sets parameters.
Finally, the fuzzy membership functions were plotted in postscript with GNU-Plot
MS-Windows version 3.5.
Conclusion
The aim of this study was to develop a feasible method of semi-automatic knowledge
acquisition of fuzzy sets for linguistic medical concepts and fuzzy relations between medical
findings and diagnoses. This could be shown with the proposed method on the
applied material. This method will remain semi-automatic because the results have to be
checked for plausibility and — should the occasion arise — adapted by a physician for the
intended application. Future work will cover not only an evaluation of different approaches
concerning the semi-automatic construction of fuzzy sets and fuzzy relations but also an
application of this method to different medical areas.
References