You are here

Farklı zaman ölçekli EEG işaretlerinden epilepsi nöbetinin otomatik tespiti

Automatic detection of epileptic seizures for different time-scaled EEG signals

Journal Name:

Publication Year:

Author NameUniversity of AuthorFaculty of Author
Abstract (2. Language): 
Epilepsy is a common disease that recurs itself with constant seizures. This disease, which is seen in about 1% of the world population, is a clinical condition in which a sudden, unexpected and irregular electrical discharge occurs in a part of the brain or completely. The brain contains important information for the detection of electroencephalography (EEG) epilepsy, which is an electrical status analysis of the nerve cells. For this reason, EEG signals have become a research area that many experts are interested in. In this study, we presented an automatic pattern recognition system using only A and E point clusters of signs from healthy subjects and epileptic seizures in a sample length of 23.6 seconds taken from Bonn University database (A, B, C, D, E). The pattern recognition system that has been presented in three stages, pre-processing, feature extraction and classification. In the first stage, the EEG signals consisting of 23.6 seconds and 4096 samples are divided into segments of 128, 256, 512, 1024, 2048, 4096 lengths. With this segmentation, we tried to study the behavior of EEG signs at different lengths. In the second step, spectral information of EEG signals was obtained by using Peridogram and Welch methods of nonparametric Power Spectral Density (PSD) methods. When Welch PSD was performed, a Hamming window was used for each of the lengths of one EEG mark length, and the overlap ratio of the parts was selected as 50%. By using two different PSD estimation methods in the study, it was tried to investigate the behavior patterns of the data segments in different spectral methods. After the PSD estimation, an arithmetic mean is applied to all EEG signals to reduce the data size, and each segment is represented by the feature vector in 16 sample lengths. In the third and last phase, the feature vectors of 16 sample lengths obtained for each EEG segment are classified by k nearest neighbors (k-NN), support vector machine (SVM) extreme learning machine (ELM) using 5-fold cross-validation method. With this classification, we tried to investigate the performances of different classifiers of feature vectors obtained by different PSD estimates of different data segments. The accuracy performances for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by the Periodogram PSD estimation with k-NN were 99.30%, 99.66%, 99.81%, 99.75%, 100%, 100% respectively. The accuracy performances for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by the Welch GSY estimation with k-NN were 99.30%, 99.72%, 99.75%, 99.88%, 100%, 100% respectively. The accuracy performances for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by periodogram PSD estimation with SVM were 99.41%, 99.72%, 99.88%, 100%, 100%, 100% respectively. The accuracy performances for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by the Welch PSD estimation with the SVM were 99.24%, 99.75%, 99.88%, 100%, 100%, 100% respectively. The accuracy performance for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by the Periodogram PSD estimation by ELM was 99.33%, 99.72%, 99.75%, 99.88%, 100%, 100% respectively. The accuracy performances for the 128, 256, 512, 1024, 2048 and 4096 segments resulting from the classification of the feature vectors obtained by Welch PSD estimation by ELM were 99.18%, 99.72%, 99.75%, 99.88%, 100%, 100% respectively. It suggested that the pattern recognition system wherein the performance evaluated for the SVM classifier has been found that good performance is obtained. In terms of feature extraction methods used in the system Welch PSD estimation method of Periodogram PSD it was found to give better results according to the estimation. As accuracy performance is evaluated in terms of the different EEG data lengths used, in case of the data length redundancy, the accuracy performance is improved. A number of pattern recognition techniques have been proposed with different feature extraction and classification techniques than the EEG markers used in this study. The results obtained in these studies were observed to be close to the results when we performed the study.
Abstract (Original Language): 
Epilepsi halk arasındaki adı ile sara kendini sürekli nöbetler ile tekrarlayan yaygın bir hastalıktır. Dünya nüfusunun yaklaşık olarak % 1’de görülen bu hastalık beynin bir bölümünde yahut tamamında meydana gelen ani, beklenmedik ve düzensiz elektriksel boşalma sunucu ortaya çıkan klinik bir durumdur. Beyinde bulunan sinir hücrelerinin elektriksel durum analizi anlamına gelen elektroensefalografi (EEG) epilepsinin tespiti için önemli bilgiler içermektedir. Bu sebeple EEG işaretleri birçok uzmanın ilgilendiği bir araştırma alanı haline gelmiştir. Bu çalışmamamızda Bonn Üniversitesi veri tabanından (A,B,C,D,E) alınan 23,6 saniye 4096 örnek uzunluğunda sağlıklı ve epilepsi nöbeti geçiren deneklerden alınan işaretlerden sadece A ve E işaret kümeleri kullanılarak gerçekleştirilen bir otomatik örüntü tanıma sistemi sunulmuştur. Sunulan örüntü tanıma sistemi ön işlem, öznitelik çıkarım ve sınıflandırma olmak üzere üç aşamadan meydana gelmiştir. Birinci aşamada 23,6 saniye ve 4096 örnekten oluşan EEG işaretleri 128, 256, 512, 1024, 2048, 4096 uzunluğunda bölütlere ayrılmıştır. İkinci aşamada parametrik olmayan güç spektral yoğunluk (GSY) yöntemlerinden Peridogram ve Welch yöntemleri kullanılarak EEG işaretlerinin spektral bilgisi elde edilmiştir. Welch GSY kestirimi yapılırken her bir EEG işaret uzunluğunun dörtte biri uzunluğunda hamming penceresi kullanılmış ve parçaların örtüşme oranı %50 olarak seçilmiştir. GSY kestirimi yapıldıktan sonra veri boyutunu azalmak için tüm EEG işaretlerine aritmetik ortalama uygulanmış ve her bir bölüt 16 örnek uzunluğunda öznitelik vektörü ile temsil edilmiştir. Üçüncü ve son aşamada her bir EEG bölütü için elde edilen ve 16 örnek uzunluğunki öznitelik vektörleri 5-katlı çapraz doğrulama yöntemi kullanılarak k en yakın komşu algoritması (k-NN), destek vektör makinesi (SVM), aşırı öğrenme makinesi (ELM) ile sınıflandırılmıştır. Tüm sınıflandırıcılar ile yapılan çalışmalarda maksimum %100 sonuç elde edilmiştir.
745
757