Epileptic Seizure Prediction and the Dimensionality Reduction Problem

Abstract

Seizures prediction may substantially improve the quality of life of epileptic patients. Processing EEG signals, by extracting a convenient set of features, is the most promising way to classify the brain state and to predict with some antecedence its evolution to a seizure condition. In this work neural networks are proposed as effective classifiers of brain state among 4 classes: interictal, preictal, ictal and postictal. A two channels set of 26 features is extracted. By correlation analysis and by extracting the principal components, a reduced features space is obtained where, by an appropriate neural network, over 90% successful classifications are achieved, for dataset with several patients from the Freiburg database.