Seizure detection using regression tree based feature selection and polynomial SVM classification

This paper presents a novel patient-specific algorithm for detection of seizures in epileptic patients with low hardware complexity and low power consumption. In the proposed approach, we first compute the spectrogram of the input fragmented EEG signals from three or four electrodes. Each fragmented data clip is one second in duration. Spectral powers and spectral ratios are then extracted as features. The features are then subjected to feature selection using regression tree. The selected features are then subjected to a polynomial support vector machine (SVM) classifier with degree of 2. The algorithm is tested using the intra-cranial EEG (iEEG) from the UPenn and Mayo Clinic's Seizure Detection Challenge database. It is shown that the proposed algorithm can achieve a sensitivity of 100.0%, an average area under curve (AUC) of 0.9818, a mean detection horizon of 5.8 seconds, and a specificity of 99.9% on using half of the training data for classification. The proposed approach also achieved a mean AUC of seizure detection and early seizure detection of 0.9136 on the testing data.