الفهرس 
Chapter 1: IntroductionOnly 14 pages are availabe for public view
 |  | from | 109 |  |  |
| | | from | 109 | | |
Abstract
Electroencephalography (EEG) is a monitoring tool to record the electrical
activity of the human brain. EEG signals are acquired from human brain
in order to classify the activities of the patient. One of the most important
applications of the EEG signals is the epileptic seizure detection and prediction.
EEG signals have a multi-channel nature. For EEG signal processing,
not all channels are effective. So, channel selection is required to reduce the
computation time and complexity.
In this thesis, we present an efficient frequancy-domain approach for EEG
seizure detection. It is based on the segmentation of EEG signals into 1-
sec segments, then EEG signals are reformated into two dimensions (2-D).
Fast Fourier Transform (FFT) is applied to these segments, and frequency
magnitudes are taken without phase information. Scale Invariant Feature
Transform (SIFT) is used to generate a map for the discriminative points in
the frequancy domain magnitude. Statistical distributions for the number of
discriminative points in seizure and non-seizure segments are estimated. Based
on the PDF of the number of discriminative features, a certain threshold is
set for discrimination. The proposed method achieves an average accuracy
of 99.16%, an average sensitivity of 99.97% and 98.09% average specificity.
The results indicate that the proposed algorithm achieves high sensitivity,
specificity and accuracy compared to other existing algorithms.
Moreover, the thesis presents a frequency-domain analysis of EEG signals
for epileptic seizure prediction. First of all, EEG signals are segmented
for pre-ictal, and ictal activities for the training purpose. The segments used
are of length 10-second of different channels. Five attributes of EEG signals
in the frequency domain are considered. These attributes are frequency
domain amplitude, local mean, local variance, derivative, and local median.
A composite feature vector is composed from these attributes, and classification
of states for seizure prediction is performed with an Artificial Neural
Network (ANN). Performance of the proposed algorithm is studied using the
publicly available CHB-MIT dataset, achieving an average accuracy of 90.9%.
A comparison study with state-of-theart algorithms prove superiority of the
proposed algorithm from the accuracy perspective.
Furthermore, this thesis introduces a patient-specific approach for seizure
prediction applied to scalp Electroencephalography (sEEG) signals. The proposed
approach depends on computing the instantaneous amplitude of the
analytic signal by applying Hilbert Transform (HT) on EEG signals. Then,
the Probability Density Functions (PDFs) are estimated for amplitude, local
mean, local variance, derivative and median as major features. This is followed
by a threshold-based classifier which discriminates between pre-ictal and interictal
periods. The proposed approach utilizes a patient-specific algorithm for
channel selection to identify the optimum number of needed channels which
is useful for real-time applications. It is applied to all patients from the CHBMIT
dataset, achieveing an average prediction rate of 96.46%, an average false
alarm rate of 0.028077/h and an average prediction time of 60.1595 minutes
using a 90-minute prediction horizon. Experimental results prove that HT is
more efficient for prediction than other existing approaches.