Epileptic Activity Detection in EEG Signals using Linear and Non-linear Feature Extraction Methods

The aim of this study is to obtain an automated medical diagnosis-support system about epilepsy by classifying EEG signal epochs as ictal, inter-ictal and normal. EEG signals were analyzed in their sub-bands obtained via discrete wavelet transform. Linear and non-linear methods are used for extracting features of normal, ictal and inter-ictal states. Support vector machine classification is realized by using time domain features which are autoregressive coefficients and linear prediction error energy; and information theory based features which are Shannon entropy and approximate entropy. In order to improve accuracy, linear and non-linear features are combined and then SVM trained by these features. By the proposed method, 99.0%, 96.0%, 100% accuracy, sensitivity and specificity are obtained for epileptic and non-epileptic classification, while accuracy, sensitivity and specificity of 98.2%, 95.0 and 99.0% are obtained for normal, ictal, and inter-ictal activity classification, respectively.