Predicting epileptic seizures from raw EEG signals using advanced deep learning techniques

Epilepsy is one of the most common neurological diseases globally. Patients diagnosed with epilepsy suffer from sudden epileptic seizures. Several solutions exist for epilepsy patients, such as anti-epileptic drugs and deep brain stimulation however, these solutions can cause irreversible brain damage. Hence, the demand for seizure prediction devices is rising. A solid predictive algorithm that can take as an input raw EEG signals and predict the signal type to be seizure or non-seizure is in demand. In this paper, we aim to explore building a patient-specific model using raw EEG signals preprocessed by different techniques such as representing signals using grayscale pixel intensity and applying short-time Fourier transforms and spectrograms. The uniqueness of the implemented models lies in their ability they work on raw EEG data. This promotes the algorithms to work on other datasets too. Box plots are used in this work to visualize the difference between training and testing data. Four models are created and trained in this project; a simple convolutional neural network (CNN), a normalized CNN, a time-distributed CNN, and VGG. During training and validation, selected models using grayscale pixel intensity, resulted in great AUC scores that exceeded 0.85, during testing, the normalized CNN model with spectrograms showed the highest area under the curve (AUC) for the test files of patient 1 which was 0.58.