A Signal-Based One-Dimensional Convolutional Neural Network (SB 1D CNN) Model for Seizure Prediction

Convolutional Neural Networks (CNNs) have become increasingly popular in seizure detection and prediction research. While traditional CNNs are effective in image classification tasks, applying them to seizure signal analysis requires specific architectures. In this study, we propose a Signal-Based One-Dimensional Convolutional Neural Network (SB 1D CNN) model that is customized for seizure signals. The SB 1D CNN model replaces traditional ReLU and Pooling layers with counterparts that are better suited to negative signal fluctuations and adjusts the training procedure accordingly. Additionally, the model introduces time/frequency-sensitive kernels in the initial convolution layer to capture significant features across time and frequency domains. To evaluate the proposed SB 1D CNN model, we conducted experiments using epileptic EEG signals from the CHB-MIT database. We carried out two sets of experiments: the first to identify optimal EEG channels through single-channel evaluations, and the second to train a robust SB 1D CNN model for seizure prediction. Comparative analysis with a traditional 1D CNN with a similar structure revealed that the SB 1D CNN model excels in feature extraction and classification of epileptic EEGs. Notably, training 1D CNNs exclusively with relevant data significantly enhances their performance. Overall, this study highlights the importance of tailored architectures in improving the effectiveness of 1D CNNs in seizure prediction tasks. The proposed SB 1D CNN model offers a promising avenue for enhancing the accuracy and reliability of seizure prediction systems, with potential implications for improving patient care and management in epilepsy.