الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 155 |  |  |
| | | from | 155 | | |
Abstract
The electroencephalogram (EEG) is a common method to detect the brain electrical activity. One of the most common neurological disorders that affect the brain activity is epilepsy, which can be detected using the recorded EEG signals. The EEG signals inspired several searches to detect or predict epileptic seizures for ameliorating the life of the epilepsy patients. This thesis introduces a classification system to detect and also predict epileptic seizures from samples of the EEG signals without the need to deal with the complete recorded signals. This system overcomes the complexity problems found during the utilization of long-term signals and saves the time. Three approaches are proposed to classify the EEG signals and detect the existence of seizures or to predict the incoming ones. The first approach depends on a pre-trained deep learning networks, namely Alex-Net, visual geometry group (16,19) VGG-16, and VGG-19. The results using this approach show an accuracy of 92.17% with a test time of 17 min and 51 sec by using 70% of the used data for training and 30% for testing with Alex-Net. The second approach depends on machine learning after transforming the signal into an appropriate domain. Discrete cosine transform (DCT), discrete sine transform (DST), fast Fourier transform (FFT), discrete wavelet transform (DWT) and S-transform have been adopted and compared for efficient classification. In addition, a hybrid structure comprising both DWT and S-transform has been considered. A moving time window without overlapping has been adopted for EEG signals framing. The different signal frames have been classified by a support vector machine (SVM) with radial basis function (RBF). The sensitivity (Se), specificity (Sp), and the computational time (Ct) are used to evaluate this approach. The results of this approach reveal an average sensitivity, an average specificity, and a computational time of 95.156 %, iv 82.163%, and 30 min and 58 sec, respectively in epileptic seizure detection, and 96.17%, 83.17%, and 36 min and 79 sec, respectively in epileptic seizure prediction. This approach is more accurate than the first approach, but the computational time and complexity needs to be improved. The third approach depends on empirical mode decomposition (EMD) to decompose the EEG signal into intrinsic mode functions (IMFs). The S-transform is utilized to extract significant features such as mean, maximum, variance, standard deviation, and skewness. The different features are classified by the SVM with different kernel functions. The accuracy (Ac), Se, Sp, area under the characteristic curve (AUC) and CT are used to evaluate this approach. Simulation results reveal classification accuracy, sensitivity, specificity, area under the characteristic curve, and a computational time of about 97.2%, 96.3%, 98.11%, 97%, and 45.88 sec, respectively, using the cubic Gaussian the SVM classifier with 90% of the dataset for training and 10% for testing. It is the best ratio that reveals best results.