Article Preview
TopIntroduction
Brain injury disturbs the nervous system and is considered the main cause of a well-known neurological condition called epilepsy. It has been reported that medication is of no help to 30% of those affected by the epileptic disease (about 1% of the world population) (Adeli et al., 2003). Depending on the patient condition, epilepsy, also known as a seizure disorder, can be determined after the occurrence of one seizure. Electroencephalogram (EEG) signals can provide significant information regarding the electrical activities generated by the cerebral cortex nerve cells. As such, the EEG signals are considered as one of the most important physiological signals used for the purpose of epilepsy detection (Sadati et al., 2006). However, the visual inspection of EEG recordings for epileptic-related features is difficult and time consuming. Moreover, different bio-signals may exhibit disagreements regarding the same condition. Consequently, machine intelligence techniques are proposed to enhance the epileptic seizure detection process. To obtain useful features, most of the techniques reported in the literature first transform the EEG signals into the frequency domain, which contains some characteristics of waveforms that fall primarily within four frequency bands (Sadati et al., 2006). In Pregenzer and Pfurtscheller (1999), a one-second time window was used for classification. The data in the window was first transformed into the frequency domain by fast Fourier transformation (FFT), where the frequency components from 9 to 28 Hz (20) were studied. The features were then used by a modified version of Khonen’s learning vector quantization classifier to discriminate between a limited number of different brain states. Poliker et. al have proposed the use of an ensemble based data fusion approach for early diagnosis of Alzheimer’s disease (Poliker et al. 2006). The event related potentials recorded from the Pz, Cz, and Fz electrodes of the EEG were decomposed into different frequency bands using multi-resolution wavelet analysis (discrete wavelet transform DWT). The data fusion approach includes generating multiple classifiers trained with selected subsets of the training data from each source, which are then combined through a modified weighted majority voting procedure. The best classification accuracy reported was 79.2%. For the fuzzy neural network classifier presented in Sadati et al. (2006), the signals decomposed into five levels using Daubechies 4 (DB4) wavelet filter. The energy of details d1, d2, d3, d4, d5 and approximation a5 (in total 6 features) were used as the input features. The classification accuracy achieved, using artificial neural networks, is 85.9%.
EEG time-domain feature-based approaches have also been reported in the literature. In a Brain Computer Interface (BCI) application (Vidaurre et al., 2009), a feature called Time Domain Parameter (TDP) is introduced to a Linear Discriminant Analysis (LDA) classifier. The TDP is a generalized form of the Hjorth parameters and can be computed efficiently. In Valderrama et al. (2012), time and frequency domain features are used for seizure classification. Time domain features such as the first, second, third, and fourth statistical moment of EEG amplitudes, (i.e., mean, variance, skewness, and kurtosis, respectively), along with long term energy and other frequency domain features were calculated and fed to a support vector machine classifier.