Neural Networks versus Genetic Algorithms as Medical Classifiers - Foundations on Natural and Artificial Computation - page 396

patient's personal information and values related to 15 tests made to each pa-
tient: age, gender, urine volume, maximun and average urine flow, vesical ca-
pacity, micturition time, micturition urine volume, detrusor muscle contraction
degree, residual post-micturition, first sensation, detrusor pressure at Qmax, de-
trusor filling pressure, lower abdominal pressure, anal tone, perianal and perineal
sensitivity.
To facilitate the processing of the experimental data, we have been forced to
homogenize the data in a preprocessing stage. After consulting with specialists
in urology, we chose the discretization technique to do this. Since we were dealing
with data from very wide ranges, as it is common in biomedicine and urology, this
preprocessing strategie could be carried out with a minimum loss of information
and behaves in a stable way. In the fews cases that there was loss of information,
as a result of being empty any of the samples' fields, the average value of that
property was assigned to those empty fields.
Moreover, given the small number of valid samples that were available for the
tasks of training, validation and testing of the classifiers, we used the technique
of cross validation [16], in order to have a greater number of samples to carry
out these tasks. So we have divided the whole sample set into 3 disjoint groups,
using one as training set and the other two as validation and test sets respec-
tively, in several iterations. This allows us to obtain better results and better
generalization rates for each classifier.
3.2 Neural Network and Genetic Algorithm Based Classifiers
Implemented
The two ANNs implemented are supervised ANNs, that is, they adquire the
needed knowledge to simulate real clinical experts performance, calculating the
error rate obtained in partial classifications during the training stage. Deppend-
ing on the value of this error rate, a bigger or lower adjust of net's connections
weights is done. As the training process goes on, the error rate is expected to be
lower reducing the needed adjustments.
A set of training, validation and testing tests were performed using the samples
obtained after applying the cross validation technique to the preprocessed input
samples. Different configurations of MLPs and RBFs were tested, varying the
parameter's values. For the MLP implementation we tested the use of three
different transfer functions (linear, logarithmic, and tangent-sigmoid), the use
of 1 or 2 hidden layers with different numbers of neurons inside them, various
number of training iterations (epochs) and five modes of the backpropagation
algorithm. The final values were chosen depending on the accuracy rates obtained
using them. The best rate was obtained with a hidden layer of 15 neurons and an
output layer of 2 neurons, where each neuron gives the probability of suffering
from one of the observed diseases. The chosen training algorithm was Levenberg-
Marquardt with 5000 iterations in the training stage, beyond this number of
repetitions the obtained results remained the same (Fig. 1).