2.3 Review of the ECG and EEG Classification
2.3.2 Epileptic Classification Methods
The key challenge of any automated epileptic seizure detection method is the extraction of the distinguishing features from EEG signals as it significantly affects the perfor- mance of the classifier. Representative characteristics or features extracted from EEG data can describe the key properties or morphologies of the signals for perfect detection of epileptic seizure [106]. In past two decades, a variety of methods has been developed for feature extraction in epileptic EEG data. These feature extraction techniques can be divided into four groups [107]: time domain, frequency domain, timefrequency domain andnonlinear methods.
The simplest features are extracted in the time domain. Three popular time do- main methods are PCA, independent component analyses (ICA) and linear discriminant analysis (LDA). PCA is an unsupervised method that transforms high-dimensional data into a low-dimensional feature subspace. In epileptic EEG signals classification, PCA can be used for feature enhancement [108] and feature reduction [109, 110]. Xie and Krishnan [111] developed an efficient feature extraction based on dynamic PCA with nonoverlapping moving window to separate EEGs from interictal and ictal for seizure detection. ICA is a feature extraction method that transforms a multivariate random signal into a signal with independent components [109]. It also can be utilised to re- move artefacts and to separate brain signals sources from EEG signals [112]. LDA is another method for dimension reduction. It finds a linear combination of features that can serve as a classifier to separate two or more classes [107]. Subasi and Gursoy [109] employed all three methods: PCA, ICA, and LDA and compared the effectiveness of them in epileptic EEG classification.
TheFrequency domainmethods are based on spectral analysis methods. The spec- tral analysis methods estimate power spectral density (PSD) which is the power dis-
tributions of a signal across the frequency. These methods can be divided into two categories [107]: the non-parametric methods which calculate spectra directly from sig- nals by using Fourier Transform (FT) method, and the parametric methods which derive spectral quantities from a statistical model of the signals. The periodogram and the cor- relogram are two common non-parametric methods. These methods can provide high resolution for adequately long data lengths. However, they suffer from spectral leakage effects and poor frequency resolution due to the use of windows and the finite length of data [17]. These limitations are overcome by the parametric method. Three parametric models are: the moving average (MA) model which is suitable for modelling spectra with broad peaks and sharp nulls, the autoregressive (AR) model which is suitable for modelling spectra with narrow peaks, and the autoregressive moving average (ARMA) model which is suitable for modelling spectra with both sharp peaks and deep nulls [17]. Polat and Gunes [113] used spectral parameters based on the Fourier transformation of EEG signals to detect epileptic seizures in EEG signals. ¨Ubeyli [114] classified nor- mal and ictal EEG time series using the AR method for feature extraction. Srinivasan et al. [115] proposed a two-class epilepsy detection methods based on EEG features extracted in bothtime domainandfrequency domain.
Since the EEG signals are non-stationary [116], many timefrequency based meth- ods have been developed for epileptic seizure classification from EEG signals. Unlike time-andfrequency-domainmethods, thetime-frequencybased methods do not impose the quasi-stationarity assumption on the data [117]. These methods include wavelet transform [118, 117, 109, 113, 119], time-frequency distribution [120, 121], multi- wavelet transform [122], and empirical mode decomposition (EMD) [123, 124]. As the wavelet transform provides a representation of the signal in both the time and frequency domains, they can accurately detect and localise transient features in the signals like epileptic spikes [125]. Discrete Wavelet Transform (DWT), Wavelet Packet Decompo-
sition (WPD), and Continuous Wavelet Transform (CWT) are the three types of wavelet transforms [107]. Among these methods, the DWT stands out in terms of efficiency and algorithmic elegance [126]. Lee et al. [127] used features based on the Euclidian dis- tance calculated from phase space representation (PSR) of wavelet coefficient to detect normal and epileptic seizure EEG signals. Pachori and Patidar [128] proposed a method for classification of ictal and seizure-free EEG signals using the 95% confidence ellipse area measured from the second-order difference plot (SODP) of intrinsic mode func- tions (IMFs) as a feature. Samiee et al. [129] proposed a feature extraction technique based on a rational discrete short-time Fourier transform for classification of normal and epileptic segments.
Although frequency domain methods can detect rhythmic oscillations in a signal, they have some limitation in capturing phase locking and nonlinear coupling between harmonics in the spectrum [130]. To better represent and reflect the characteristics of the EEG signals, various nonlinear parameters have been proposed as features using different nonlinear methods for classification of epileptic seizure EEG signals. The nonlinear parameters namely, Higher Order Spectra (HOS) [131, 132], Correlation Di- mension (CD) [133, 133, 134], Lyapunov Exponent (LE) [19, 135], Hurst Exponent (H) [130], Fractal Dimension (FD) [136], Approximate Entropy (ApEn) [137, 138, 139], Sample Entropy (SampEn) [140, 138, 141], and Recurrence Quantification Analysis (RQA) [142, 143, 143] have been used in classification of epileptic seizure EEG sig- nals. The HOS is the spectral representation of higher order statistic i.e. moments and cumulants of third and higher order [116]. Acharya et al. [132] used features extracted from HOS for classification of the normal, interictal, and ictal EEG signals. The mea- sure of CD has been used to quantify the complex neural activity of the human brain during interictal and ictal activities [133]. The LE has been used for discrimination of epileptic seizure EEG signals as it provides significant details about changes in EEG
activity [19]. The FD parameter has been used effectively to discriminate the chaotic nature of EEG signals in interictal and ictal activities [136]. Acharya et al. [130] used the CD, H, and ApEn parameters for the automated detection of epilepsy. The value of ApEn has been found to show a strong connection with the synchronous discharge of large groups of neurons during seizure [139]. Acharya [142] used the RQA features for classification of the normal, interictal, and ictal EEG signals.
Once features are extracted from EEG signals, a classifier is employed to dif- ferentiate between normal and epileptic EEG. In past two decades, many classifi- cation methods have been used for seizure detection such as artificial neural net- works (ANNs) [144, 141], multilayer perceptron neural network (MLPNN) [144], re- current neural networks (RNNs) [135], Fuzzy Sugeno Classifier (FSC) [138, 145], SVM [146, 132, 147, 109], LS-SVM [148, 149], Gaussian mixture model (GMM) [130, 110], DT [123, 150],k-NN [151, 152], Mixture expert model [119], RF [20].