Epileptic Seizure Detection by Learning EEG as Images (Prof. Imtiaz Ahmad)

Epileptic seizure detection is vital for anticipating and managing seizure events. This study explores various machine learning techniques aimed at detecting seizures, focusing on capturing both spatial and temporal seizure characteristics. By converting seizure data into images, additional features can be extracted to enhance detection accuracy. Notably, the study emphasizes the significance of leveraging Vision Transformer (ViT) and residual connections within Convolutional Neural Networks (CNNs) for this task. Seven different image representations, including Spectrogram, Scalogram, Gramian Angular Summation Field (GASF), Gramian Angular Difference Field (GADF), Poincare plot (PP), Recurrence plot (RP), and Raw EEG images, were employed to evaluate their performance with recent models. Three state-of-the-art machine learning models, namely Mix-ViT, Swin Transformer, and a novel CNN with residual connections, were utilized and assessed using the publicly available Bonn University and CHB-MIT datasets. The results demonstrate exceptional performance, achieving 100% accuracy, sensitivity, specificity, F1 score, and recall in both binary and multi-class classifications. Particularly on the CHB-MIT dataset, where only 8 channels were utilized, the proposed models achieved noteworthy accuracy, sensitivity, and specificity rates of 95.1%, 93.6%, and 97.1%, respectively. To ensure the generalizability of the model, it was initially tested on the small Bonn University dataset, consisting of only one channel, before being generalized to the CHB-MIT dataset, which features multi-channel input, to assess the performance of the models. The uniqueness of this study lies in its comprehensive evaluation and comparison with existing recent state-of-the-art methodologies, ensuring a thorough assessment of the proposed models effectiveness. Furthermore, the study conducted 51 extensive experiments on both the Bonn University and CHB-MIT datasets, highlighting the significance of the research.

Supervisor: Prof. Imtiaz Ahmad
Convener: Dr. Sa’ed Abed 
Examination Committee: Dr. Abdullah Alshaibani