Bu çalışmada, sağlıklı ve nöbet esnasındaki EEG sinyallerini ayrıştıran bir sistem tasarımı amaçlanmıştır. Bunun için İlinti Boyutu, Dalgacık-entropisi ve Destek Vektör Makinesi(DVM) içeren kompozit bir sistem önerilmiştir. Çalışmada kullanılan EEG verileri, Bonn Üniversitesi Epileptoloji bölümü veritabanından alınmıştır. Bu veritabanından 50 adet sağlıklı ve 50 adet epileptik olmak üzere toplam 100 adet EEG bölütü kullanılmıştır. Bu bölütlere kaotik yöntemlerin uygulanabilmesi için öncelikle faz uzayları oluşturulmuştur. Faz uzayları üzerinden İlinti Boyutu değerleri hesaplanmıştır. Dalgacık analizi ile EEG bölütleri, literatürde standart olarak belirlenen alt-bantlara; delta=(0.5-4Hz), teta=(4-8Hz), alfa=(8-12Hz) ve beta=(12-32Hz) ayrıştırılmıştır. Bu bantlarda elde edilen EEG spektral bileşenlerin normalize enerjileri alınıp Shannon entropi’leri hesaplanmıştır. Sağlıklı ve epileptik EEG sinyallerinden özellik çıkarmak için ilinti boyutu analizinden elde edilen özgün veri ve dalgacık-entropi analizinden elde edilen özgün veriler (4 adet alt bant entropi’leri) DVM’nin girişine verilmek üzere her bir EEG bölütü için 5’li bir öznitelik vektörü oluşturulmuştur. Elde edilen tüm öznitelik vektörlerinin sınıflandırılması için DVM kullanılmıştır. DVM sağlıklı ve epileptik olmak üzere toplam 50 EEG bölütü ile eğitilmiş ve geriye kalan 50 bölütle de test yapılmıştır. Sağlıklı ve epileptik EEG bölütlerinin hesaplanan ilinti boyutları ve dalgacık entropilerinin sınıflandırmada ayırt edici olduğu görülmüştür. Başarım değerlendirme ölçütleri kullanılarak önerilen kompozit sistemin %98 gibi bir başarı ile sınıflandırma yapabildiği tespit edilmiştir.

Objective of this study was to design a system for classifying Epileptic and normal EEG signals. For this purpose, a system composed of correlation dimension, wavelet-entropy and Support Vector Machine was proposed. Epilepsy is a neurological disorder which can be seen all over the world. It can be diagnosed by brain’s electrical activity. The determination of epileptic attacks or seizures by Electroencephalogram (EEG) signals is quite common in both clinical and research fields. Because EEG signals are non-stationary signals, they must be examined with the nonlinear analysis methods. For the analysis of a chaotic signal or system, first of all, a trajectory of the attractor which represents the system, and depicts all states the systems acquire in the course of time must be created on the phase space. Provided that time series is the output of a chaotic system, the trajectory created on the phase space is anticipated to display a regular structure at times, and random at the other. It cannot be estimated in advance when and how long the trajectory is regular or random. However, some methods quantifying the degree of chaoticity of the system have been developed. With these methods, calculations are made using the trajectory created on the phase space by the system and the degree of chaoticity is quantitatively determined. Each method reflects chaoticity of the systems in different ways. In other words, each quantity obtained from the system defines a different feature vector. The excessive number of feature vectors means better recognition of the system. Since that means more parameters, the processing load also increases. In the Literature, the Lyapunov exponents, the correlation dimension and the entropy, are widely used for analysis of the chaotic time series or systems. In addition, time-frequency techniques can be used to analyze this kind of signals and systems. Support Vector Machines (SVM) is one of the methods commonly used in classification. SVM tries to find the most appropriate plane (hyperplane) separating the two classes. EEG data used in this study have been acquired from the database of University of Bonn, Department of Epileptology. From this database, 100 EEG segments (50 healthy and 50 epileptic segments) have been used. To apply chaotic methods to these segments, phase spaces have primarily been created, and then the Correlation Dimension values have been measured. In this study, the normal and the epileptic EEG signals were examined. First of all, the correlation dimension of both the normal and the epileptic EEG signals were measured. All of the EEG signals have been separated into the standard subbands which are: delta=(0.5-4Hz), theta=(4-8Hz), alpha=8-12Hz and beta=(12-32Hz). Then, the Shannon entropies of the EEG subbands are calculated; and then the feature vectors are formed by combining the values obtained with both methods. Finally, all the feature vectors are classified with SVM. SVM was trained with 50 EEG segments in total, composed of 25 healthy and 25 epileptic EEG segments, and a test was conducted with the remaining 50 segments. The measured correlation dimensions and wavelet entropies of EEG segments were detected to be distinctive in classification. The composite system that was proposed using performance evaluation criteria showed a 98 % success rate in classification.