A generalizable model for seizure prediction based on deep learning using CNN-LSTM architecture

This work proposes a novel deep learning-based model for prediction of epileptic seizures using multichannel EEG signals. Multichannel images are first constructed by applying short-time Fourier transform (STFT) to Electroencephalography (EEG) signals. After a preprocessing step, a CNN-LSTM neural network is trained on the STFTs in order to capture the spectral, spatial and temporal features within and between the EEG segments and classify them as preictal or interictal stage. The proposed method achieves a sensitivity of 98.21%, a false prediction rate (FPR) of 0.13/h and a mean prediction time of 44.74 minutes on the CHB-MIT dataset. As the main contribution of this work, by using a CNN-LSTM, in addition to capturing the time-frequency features of each segment using the convolutional network, the proposed model is able to capture the temporal patterns and transitions between sequential segments and hence improve the prediction performance in comparison to previous deep learning-based models. The method needs no complex feature extraction or channel and feature selection