A Convolutional Long Short-Term Memory-Based Neural Network for Epilepsy Detection From EEG

Epilepsy (EP) is a severe neurological disorder characterized by recurrent seizures, which increases the risk of death three times more than normal. Currently, electroencephalography (EEG) has emerged as a highly promising technique for the diagnosis of EP. The majority of current EEG-based EP detection research has employed a variety of deep-learning (DL)-based models, but most of the approaches suffer from poor generalizability, optimal design, and performance rates. To address these issues, this study aims to develop an efficient framework based on the deep spatiotemporal neural network called convolutional long short-term memory (ConvLSTM) for EP detection from EEG signals. In the proposed model, first standard 19-channel EEG data are selected and resampled at 256 Hz and then those signals are segmented into 3-s time frames. Afterward, the segmented data are fed as input to the ConvLSTM model for identifying epileptic patients from normal subjects. To generalize the proposed model, we have tested it on two different datasets with varying population sizes. We have used the five-fold cross-validation and leave-one-out cross-validation (LOOCV) schemes to eliminate the experiment’s biases. To further validate the proposed framework, we have carried out various ablation studies. The experimental results demonstrate that the proposed model outperforms the current state-of-the-art results for the studied datasets, making it suitable for use as an automated system for the diagnosis of EP.