Scalogram based deep learning approach for classification of epilepsy types from EEG signals

In this thesis, it is aimed to classify different types of epilepsy from EEG signals. A scalogram-based deep learning approach was proposed by increasing the size of the raw EEG signal in the classification process. To be able to study this thesis; basic oscillations in the brain in the face of the simplest paradigms, activity measurement techniques such as EEG and signal processing methods. During the supervisory activities of the brain, the electrical activities of the nerve cell groups produce oscillations. These complex biopotential oscillations are called electroencephalogram (EEG) signals. Certain diseases can be detected using these signals. One of these diseases is epilepsy. Epilepsy is a disease that manifests itself as seizures. These seizures manifest themselves in different characteristics. These different characteristics divide epilepsy seizure types into two main groups. These seizures are called generalized and partial epilepsy. For physicians, detection of these types of seizures is important for the treatment of the disease. Visual evaluation of long-term EEG recordings used in the detection and follow-up of epilepsy is costly in terms of time. It is seen that the methods to minimize this cost can be proposed in engineering fields. In the literature, EEG-based studies in the field of engineering have been evaluated in two main groups. These studies; a) Conventional methods based studies, b) Deep learning methods based studies. In this thesis, it is aimed to classify different types of epilepsy from EEG from EEG signals. For this purpose, a scalogram-based deep learning approach has been proposed. In order to evaluate the success of the proposed method, the EEG data set of the Dicle University Neurology Clinic and the Bonn EEG data set, which are frequently used in the literature, were used. In the proposed method, the size of the EEG signal was increased by applying Continuous Wavelet Transform (CWT). Two-dimensional (2D) time-frequency scalogram images with increased dimensions were used as input patterns to the convolutional neural network and classified by training. The Bonn EEG data set includes different paradigms and different time / position signals of healthy and epilepsy. These signals are labeled A, B, C, D and E. As a result of classification prcess; A-E and B-E datasets were 99.50%(± 1.50), A-D and B-D datasets were 100%(± 0.00), A-D-E datasets were 99.00%(± 1.33), A-C-D-E datasets were 90.50%(± 1.70) and B-C-D-E datasets were 91.50%( ± 2.29) and A-B-C-D-E data sets were obtained with an accuracy of 93.60%(± 3.07). EEG data set obtained from DU Neurology Clinic as labeled four classes; healthy, generalized pre-seizure, generalized seizure and partial epilepsy. In the classification study performed for normal, generalized pre-seizure and generalized seizure EEG recordings 90.16% (± 0.20); healthy, generalized pre-seizure, generalized seizure time and partial EEG records were classified with an average of 84.66% (± 0.48) accuracy. According to these results, in the confusion matrix obtained; normal EEG recordings were 91.29%, generalized epileptic seizures (seizure time) 96.50%, partial EEG records 89.63%, pre-seizure EEG records 90.44% accuracy was achieved. The results of the proposed method were compared with the results of both similar studies and conventional methods. As a result, the performance of the proposed method was found to be acceptable.