Hybrid CNN-TCN Modeling of EEG Temporal Signals for Epileptic Seizure Prediction

Artificial Intelligence (AI) is rapidly advancing in health investigation, particularly in the detection of neurological diseases such as epileptic seizures. Despite this, detection of epileptic seizures from time-series EEG patient records remains a significant challenge. This research paper proposed a new hybrid CNN-TCN (Convolutional Neural Network- Temporal Convolutional Network) modelling approach to address this gap by identifying active seizure patterns using CHB-MIT Scalp EEG temporal data. CHB-MIT data from pediatric patients with insoluble epilepsy are composed of multiple channels that generate EEG-time-dependent signals, segmented into seizures and non-seizure phases such as ictal and pre-ictal that highlight important temporal patterns. Therefore, the proposed model's CNN layer extracts spatial features from EEG channels, while the TCN layer extracts long-term time dependencies and enables an accurate identification of the seizure patterns. The hybrid architecture integrates the dilated convolutions of the TCN layer and the 1D convolutions of the CNN layer, and the batch normalization and the regularization of the elastic net to improve generalization and reduce over adjustment. The model was evaluated by four experiments, comparing the performance of CNN, LSTM-TCN, and LSTM on the same datasets to test the validation of the proposed deep learning approach against the other deep learning approaches. The results show significant improvements, with a 10-fold cross validation accuracy of 99.49%, while the ROC curve reached a perfect score of 1.00, strengthening the reliability and effectiveness of the model.