Comparison of Different Deep Learning Networks to Classify Epilepsy Seizure Based on EEG Signals

Epilepsy, a widespread neurological condition impacting approximately 50 million people worldwide, is often identified and studied through electroencephalography (EEG) due to its rich temporal information. Convolutional neural networks and recurrent neural networks have demonstrated potential in automatically detecting seizures from EEG data. Despite this, there's a shortage of comprehensive studies comparing the detection capabilities of different designs. This research assesses and contrasts the effectiveness of five renowned deep neural network models - CNN, LSTM, BiLSTM, GRU, and BiGRU - in binary categorization of EEG segments into seizure and non-seizure groups. Training and evaluation utilize the publicly accessible Bonn dataset, specifically crafted for epileptic seizure detection. Performance metrics like accuracy, precision, recall, and F1-score are employed in the analysis. The principal findings highlight BiGRU's impressive accuracy at 95.80% and F1-score at 92.96% in individualized seizure identification. Additionally, the study introduces a groundbreaking deep learning architecture, the self-normalizing neural network (SNN), to enhance seizure detection accuracy. The recommended SNN exhibits outstanding performance, boasting an accuracy of 98.50% and an impressive F1-score of 98.50%.