Epileptik EG İşaretlerin Aşırı Öğrenme Makineleri ile Sınıflandırılması
Abstract
Bu çalışmada Epilepsi tanısı konulmuş hastalardan alınan EEG işaretleri, nöbet öncesi, nöbet anı ve nöbet
sonrası olarak sınıflandırılmıştır. EEG işaretleri lineer ve durağan olmayan işaretler olup beynin elektriksel
aktivitelerini gösterirler. Nörolojik anormallerde EEG işaretlerin alt bantlarında normal durumdan farklı
olarak belirgin değişimler gözlemlenmekte ve bu değişimler nörolojik hastalıkların belirtisi olmaktadır.
Epilepsi gibi nörolojik hastalıklarda EEG işaretleri içerisindeki bantlarda normal durumdan farklı olarak
bir faz senkronizasyonu ortaya çıkmaktadır. Bu faz eşleşmelerini yüksek dereceden spektral analizi
tekniklerinden olan ikiz spektrum analizi ile ortaya çıkararak EEG işareti içerisinden özelikler elde
edilebilmektedir. Elde edilen bu özelliklerin bir sınıflandırıcının girişine verilmesi ile epileptik EEG
işaretleri sınıflandırılmaktadır. Çalışmada hızlı ve yüksek doğruluk sağlaması açısından sınıflandırıcı olarak
aşırı öğrenme makineleri kullanılmıştır. Kullanılan bu yöntem ile %98,60 gibi yüksek bir doğrulukla
sınıflandırma gerçekleştirilmiştir. Bu çalışmanın nörologlara epilepsi tanısında yardımcı olacağı
düşünülmektedir.
Keywords
Classification of Epileptic EEG Signals by Extreme Learning Machines
Abstract
In this study, the EEG signals obtained from patients
that diagnosed with epilepsy seizure, were classified
as before, during and after seizures. EEG signals
are the non-linear and non-stationary signals that
indicate the electrical activity of the brain. Different
from normal situation of the brain, in the abnormal
neurological, changes are significantly different in
the sub-band of EEG signals, and these changes are
signs of neurological disease. Since epilepsy starts
the dynamic in the brain changes while the
nonlinearity and non-Gaussanity increases in the
EEG signal. So, the phase synchronization arises
during seizure. During this phase match the features
of the EEG signals can be obtained by using
bispectrum analysis which is one of the higher order
spectral analysis techniques. Bicoherence, as the
normalized version of the bispectrum, of EEG
signals obtained from eight patients were
determined, and quadratic phase coupling (QPC)
identified. These features, which is obtained by
epileptic EEG signals were fed to the input of the
classifier. In terms of providing fast and high
accuracy for classification of the EEG signal, the
extreme learning machine (ELM) was used. The
ELM is a single hidden layer feed-forward neural
network. For comparison the artificial neural
network (ANN) and support vector machine (SVM)
classifiers were also used.
In the study, it was shown that the QPCs in the EEG
signal increased during epilepsy compared with
before epilepsy. This result shows that the
complexity and non-Gaussianity increase during
epilepsy seizure. By considering the sub-bands of
EEG separately, during epilepsy, the ratio of QPC
has increased in the low frequency compared to high
frequency. In the study it was also shown that the
QPC after epilepsy is higher than before epilepsy,
however, the QPC after epilepsy is lower than
during epilepsy. This suggests that the brain
dynamic after epilepsy seizure is more synchronous
than before epilepsy seizure. This situation is going
on until brain activities became normal.
In the study 8 patient’s EEG that were diagnosed
with seizure were used. 400 episodes of each pre
epilepsy, during epilepsy and after epilepsy were
obtained from whole data. A train/test data rate of
50%-50% was used in the classifiers. The test results
show that the ELM has higher accuracy than ANN
and SVM as shown in the Table 2. By using ELM a
high classification accuracy of 98.60% was
obtained. For ANN and SVM the test results of
%95.33 and %91.25 obtained respectively.
Furthermore, it was also shown that the ELM is
much faster than ANN and SVM classifiers. This
study is thought to help neurologists in the diagnosis
of epilepsy.
Keywords
