Epileptic seizure prediction using EEG signals with deep learning

Epilepsy is a chronic disease that dates back to ancient times and affects individuals only during seizures. Due to the unknown onset and duration of seizures, it significantly impairs the quality of life for patients. If the onset of seizures can be predicted with sufficient advance notice, patients who respond to medication can prevent their seizures using appropriate drugs, while those whose seizures cannot be controlled with medication can be provided an opportunity to move to a safe zone until the seizure subsides. In this thesis, the main objective is to create a time window of approximately 30-60 minutes before the occurrence of epileptic seizures, allowing patients to receive necessary alerts. In this context, an open-access dataset consisting of electroencephalography (EEG) recordings from 24 pediatric patients was utilized. Frequency-based and time-based feature extractions were performed on EEG data. The frequency-based and time-based methods utilized wavelet transformation and non-linear energy calculations respectively. These feature extraction methods yield to 3-class EEG features that are: ictal (seizure), preictal (pre-seizure), and interictal (between-seizure). Deep learning and machine learning classifiers were applied on these 3-class outputs. A 1-dimensional Convolutional Neural Network (1D-CNN) model, constructed from deep learning algorithms, and commonly used machine learning algorithms, namely k-Nearest Neighbors (k-NN), Random Forest (RF), Support Vector Machine (SVM), and C4.5 algorithm (J48), were employed to assess the classification performance of 3-class outputs. In the study, besides comparing different features with each other, machine learning and deep learning classifiers have also been compared, and the best results have been observed. According to the classification results, the highest average accuracy of 95.30% was achieved with a 1D-CNN classifier based on wavelet features, while the highest average accuracy of 96.60% was obtained with an SVM classifier using non-linear energy features.