Abstract:
Deep learning played a vital role in the seizure prediction challenge. Nevertheless, most studies used generic architectures
that fail to consider the distinct characteristics of multivariate time-series Electroencephalography signals. Additionally,
many methods depend on inadequate EEG segmentation techniques, resulting in unreliable results. This study presents
an in-depth architectural design of a Convolutional Neural Network specifically tailored to extract features from the
wavelet-transformed EEG signals using Wavelet packet decomposition (WPD). In addition, the chosen testing strategy
and data segmentation methodology ensures accurate and trustworthy performance results. This study introduces a data
segmentation method to generate distinct intervals and effectively capture the temporal dynamics of the time-series data.
The proposed model evaluation utilized 12 subjects’ EEG data from the CHB-MIT dataset, employing a Leave-One-Out
cross-validation technique. The proposed architecture outperformed five reproduced state-of-the-art models in the
segment-based accuracy, sensitivity, and specificity metrics. The proposed model achieved 78.00% accuracy, 65.17%
sensitivity, and a high 90.83% specificity rate. Evaluation using the more straightforward KFold cross-validation technique
demonstrated robust performance, achieving 96.68% accuracy, 97.41% sensitivity, and 95.95% specificity. The significant
improvement in the model’s specificity rates indicates a substantial reduction in false alarms, making the proposed model
a reliable tool for seizure prediction.
