Linear features, principal component analysis, and support vector machine for epileptic seizure prediction progress

One of the main issues in seizure prediction is to provide a workable approach to apply in implantable devices. For this purpose, power consumption and computational resources should be taken into account. Hence, our motivation for pursuing this work was to propose an algorithm in which not only implementation requirements could be adopted but also sufficient sensitivity and specificity could be obtained. Low computational burden of linear features make them as a proper choice for seizure prediction. With Selection of optimal features using Principal components analysis (PCA) technique, the speed of algorithm can be increased. Support Vector Machines (SVMs) have robust performance in high dimensional and imbalanced data. Therefore the proposed solutions are concentrated on power spectrum over different frequency bands and PCA for dimensionality reduction of features. Finally, SVM is applied for distinguishing brain states. In this study, seizure prediction method has been applied to EEG of 9 patients in the Freiburg database and has been achieving high sensitivity of 88.9 % and low false alarm rate of 0.21 per hour.