| This thesis presents a comprehensive approach to enhancing epileptic seizure detection
using advanced feature extraction strategies and hybrid deep learning models, specifically
focusing on EEG and MEG data. The motivation for this study stems from the increasing
prevalence of epilepsy and the need for accurate and timely diagnosis, which can significantly
improve patient outcomes. By leveraging the capabilities of both EEG and MEG signals,
this research aims to provide a robust framework for identifying epileptic seizures with high
accuracy.
The methodology involves two primary feature extraction strategies: the Simple
Continuous Morlet Wavelet Transform (CMWT) and the Complex Continuous Morlet Wavelet
Transform (CCMWT). These techniques are used to transform the time-series EEG and MEG
data into scalograms, which are then fed into various deep-learning models for classification.
The simple CMWT provides a basic level of feature extraction, while the complex CMWT
captures more intricate details of the signal, potentially leading to better performance in
seizure detection.
Three types of different deep learning models are employed in this s tudy: a 3-layered
simple 2D Convolutional Neural Network (CNN), two transfer learning approaches using
the VGG16 and ResNet50 models, and two hybrid cascaded models combining CNN with
Long Short-Term Memory (LSTM) networks. These models are trained and tested on both
EEG and MEG datasets to evaluate their effectiveness in identifying epileptic s eizures. The
comparative analysis of these models provides insights into their strengths and limitations,
contributing to the development of more accurate and efficient seizure detection systems.
The results indicate that the complex CMWT outperforms the simple CMWT in terms
of feature extraction capabilities, leading to higher classification accuracy across all models.
The hybrid CNN-LSTM model demonstrates the best performance, effectively capturing both
spatial and temporal features of the EEG and MEG signals. The transfer learning models
also show promising results, highlighting the potential of pre-trained models in medical signal
processing tasks.
A detailed performance analysis is conducted, comparing the simple and complex
CMWT strategies. The findings reveal that while the simple CMWT is computationally
less intensive, the complex CMWT provides significantly better results in seizure detection
accuracy. This trade-off between computational efficiency and detection performance is
crucial for practical implementations of seizure detection systems in clinical settings.
In addition to the technical contributions, this thesis also explores the clinical impli-
cations of the proposed methods. The ability to accurately detect epileptic seizures using
non-invasive techniques like EEG and MEG can lead to better patient monitoring and man-
agement. This research underscores the importance of integrating advanced signal processing
and machine learning techniques in medical diagnostics.
In summary, this thesis demonstrates the effectiveness of advanced feature extraction
strategies and hybrid deep learning models in detecting epileptic seizures using EEG and
MEG signals. The comparative analysis of simple and complex CMWT strategies provides
valuable insights into their respective strengths and limitations, paving the way for future
research and development in this critical area of medical diagnostics.
Keywords: epileptic seizure detection, wavelet transforms, deep learning, EEG, MEG, signal
processing, hybrid models |