Seizure prediction using cost-sensitive support vector machine

Abstract

Approximately 300,000 Americans suffer from epilepsy but no treatment currently exists. A device that could predict a seizure and notify the patient of the impending event or trigger an antiepileptic device would dramatically increase the quality of life for those patients. A patient-specific classification algorithm is proposed to distinguish between preictal and interictal features extracted from EEG recordings. It demonstrates that the classifier based on a Cost-Sensitive Support Vector Machine (CSVM) can distinguish preictal from interictal with a high degree of sensitivity and specificity, when applied to linear features of power spectrum in 9 different frequency bands. The proposed algorithm was applied to EEG recordings of 9 patients in the Freiburg EEG database, totaling 45 seizures and 219-hour-long interictal, and it produced sensitivity of 77.8% (35 of 45 seizures) and the zero false positive rate using 5-minute-long window of preictal via double-cross validation. This approach is advantageous, for it can help an implantable device for seizure prediction consume less power by real-time analysis based on extraction of linear features and by offline optimization, which may be computationally intensive and by real-time analysis.