Deep Learning Based Method for Output Regularization of the Seizure Prediction Classifier

Seizure prediction has been a great challenge for neuroscientists in the last decade. Forecasting an epileptic episode has a substantial role in preventing or mitigating the harm that comes along, especially in some medically intractable cases. Despite the remarkable breakthrough in Artificial Intelligence and automated reasoning, the prediction of the epileptic seizure is still challenging due to the lack of dataset. Additionally, Electroencephalogram (EEG) signals are patient-specific, thus making it difficult to benchmark findings. In this study, instead of classifying short EEG segment, we detected the preictal phase as one long sequence of a latent space representation. We have compared different deep learning architectures to elect the best model. Using Long Short-Term Memory (LSTM) neural network or Gated Recurrent Units (GRUs) helps to obtain promising results. In fact, we proved that LSTM/GRUs could be applied successfully as a post-processing regularizer for the classifier output.