An efficient seizure prediction method using KNN-based undersampling and linear frequency measures

Seizure prediction based on analysis of electroencephalogram signals has generated considerable research interests. A reliable seizure prediction algorithm with minimal computational requirements is prominent issue for medical facilities; however, it has not been addressed correctly. In this study, an optimized novel method is proposed in order to remove computational complexity, and predict epileptic seizures clinically. It is based on the univariate linear features in eight frequency sub-bands. It also employs principal component analysis (PCA) for dimension reduction and optimal feature selection. Class unbalanced problem is tackled by K-nearest neighbor (KNN)-based undersampling combined with support vector machine (SVM) classifier. To find out the best results two types of postprocessing methods were studied. The proposed algorithm was evaluated on seizures and 434.9h of interictal data from 18 patients of Freiburg database. It predicted 100% of seizures with average false alarm rate of 0.13 per hour ranging between 0 and 0.39. Furthermore, G-Mean and F-measure were used for validation which were 0.97 and 0.90, respectively. These results confirmed the discriminative ability of the algorithm. In comparison with other studies, the proposed method improves trade-off between sensitivity and false prediction rate with linear features and low computational requirements and it can potentially be employed in implantable devices. Achieving high performance by linear features, PCA, KNN-based undersampling, and SVM demonstrates that this method can potentially be used in implantable devices.