Performance of patient independent seizure detection system using time domain measures

The objective of this work is to design a patient independent system using time domain measures to detect the electrical onset of seizure in patients with temporal lobe epilepsy (TLE). We utilized the EEG data from 29 seizures of 18 patients who underwent multi-day video-scalp EEG monitoring as part of their presurgical evaluation. Seven time domain measures - signal energy, approximate entropy (ApEn), sample entropy (SampEn), mean, variance, skewness and kurtosis were calculated for each windowed signal. Among them, signal energy was selected as significant feature to discriminate normal and seizure condition. The performance of five classifiers - Linear Discriminant Algorithm (LDA), Naive Bayes (NB), Decision Tree (DT), Support Vector Machine (SVM) and k-Nearest Neighbour (KNN) using signal energy feature were analysed to test the suitability for automated seizure detection. Among the five, LDA and NB classifiers detected the unknown samples with sensitivity (specificity) of 44% (95%), 54% (90%) respectively. The other two, DT and KNN classifiers performed with sensitivity (specificity) of 74% (73%), and 74% (67%) respectively. The SVM classifier performed with sensitivity and specificity of 64% and 82% is found suitable for the design of generalized system to detect the onset of seizure.