Neural Networks versus Genetic Algorithms as Medical Classifiers - Foundations on Natural and Artificial Computation - page 393

systems (CDSS) emerge to support the experts' decision making tasks [2]. Using a
diagnostic support system could reduce the number of needed tests having as con-
sequences the improvement of the global patient care and the reduction of the costs
associated with urodynamic testings. For these reasons and many others medical
decision support systems are used more and more.
A key question to build up these systems is to choose the suitable classifier to
perform the diagnostic process. In this study we want to test if genetic algorithms
are a good alternative to other strategies commonly used in CDSSs. To do this,
a comparison between the diagnostic results of two implemented ANNs and a
genetic algorithm will be done.
2 Automatic Classification Techniques and Diagnosis
Since its rising, the field of machine learning theory has had as one of its main
objectives the use of its techniques in the field of health [3] [4].
This area of knowledge describes algorithms and techniques that have the
aim of solving problems related to the automatic acquisition of knowledge, in
order to simulate the trial of an expert in solving problems of classification and
discrimination. This is the aspect of machine learning that is most interesting
for the healthcare field: the possibility of using its techniques for classifying
symptoms, diseases,..., to automatically get a diagnosis as if it were a human
expert.
Within machine learning, artificial neural networks (ANNs) are not the only,
but an extensively used tool to perform automatic classification tasks. Many dif-
ferent kinds of ANNs have been widely used in the diagnose of several diseases,for
instance, multi-layer perceptrons and radial basis function neural networks [5]
[6], self-organized maps [7], and others [8].
Besides ANNs, there are other more recent soft computing strategies like
genetic algorithms (GA), but their application in medical enviroments is much
lower than in ANNs case. Even, in most cases GAs are not used for diagnostic
purposes, and only to optimize the feature selection process. This lead us to
test the feasibility of using GAs as a basis of a diagnostic algorithm comparing
its performance with other widely proofed classifiers in medical diagnosis, like
multi-layer perceptrons.
2.1 Genetic Algorithms for Diagnosis
Genetic algorithms are included in the area of artificial intelligence, specifically
inside evolutionary computing. They are so named because they are inspired by
biological evolution and molecular-genetic basis [9]. To use a genetic algorithm,
first we must define a population to work on. Each individual of this population
would be what we call chromosome and each chromosome consists of a string
of features each of which we call gene. When the chains of chromosomes are
represented in binary form, we call it genotype. After defining the population,
it evolves by means of interactions among chromosomes. This will produce new
generations of chromosomes, and we call each new generation, descendants.