Epileptic Seizure Prediction Using Stacked CNN-BiLSTM: A Novel Approach

Abstract

In this work, we propose a novel hybrid architecture for epileptic seizure prediction, utilizing a deep learning approach by stacking the convolutional neural network (CNN) and bidirectional long short-term memory (Bi-LSTM) layers. The proposed approach employs a series of 1-D convolution layers, each with several filters with lengths varying exponentially. The deep Bi-LSTM layers are subsequently integrated to the design to create a densely connected feed-forward structure. The model effectively prioritizes spatiotemporal information, thus extracting key insights for identification of interictal and preictal features. The Boston Children’s Hospital–MIT datasets (Children’s Hospital Boston-Massachusetts Institute of Technology (CHB-MIT)) are utilized and fivefold cross validation is applied for training the model. The proposed model has undergone comprehensive evaluations, with sensitivity of 97.63%, precision of 98.30%, F1-Score of 98.25%, and an area under curve (AUC)-receiver operating characteristic (ROC) of 0.9 across six patients. It can predict seizures 30 min before their onset, allowing individuals ample time to take preventive measures. Compared to the state-of-the-art approach, our model achieves a higher accuracy by 3.44% and demonstrating improved prediction times.