There is a rich literature available on epilepsy detection system-based electroencephalography (EEG) signals and computational techniques. The various epileptic seizure detection techniques have been developed based on EEG signals using linear and nonlinear methods.5,6 These techniques mainly include feature extraction methods along with machine-learning algorithms to discriminate between seizure, seizure-free, and normal EEG patterns. These methods include frequency domain analysis (e.g., subbands analysis), entropy analysis, wavelet analysis, largest Lyapunov exponent, correlation dimension, fractal dimension, Hurst exponent, and higher-order cumulants.6,7 Machine-learning algorithms used in these techniques include artificial neural network (ANN), k-nearest neighbor (k-NN), support vector machine (SVM), naive Bayes classifier, the Gaussian mixture model (GMM), the fuzzy classifier, and the decision tree.

Time and frequency domain techniques are normally used for feature extraction from EEG signals; extracted features are further used as inputs to the classifiers.8 In addition to frequency domain features, researchers have commonly used discrete wavelet transform (DWT) to analyze EEG signals. DWT decomposes EEG signals into time-frequency representations. Gotman et al.9 initially proposed an automated seizure prediction technique that was widely employed. After EEG signal decomposition, features such as peak amplitude, time-duration, sharpness, and slope were used to detect epileptic seizure activity. Khan et al. used the DWT to decompose EEG signals into subbands from which features including energy and coefficient variation were computed and used to detect seizure activity. Adeli et al.10,11 used the wavelet transform to analyze and characterize the epileptic discharge as a 3 Hz spike. Using wavelet decomposition, transient features were accurately computed and localized in time and frequency domains. Subasi et al.12 proposed a method based on wavelet transform and ANNs to classify EEG seizure signals. Subasi’s team13 later improved their method by using dynamic fuzzy neural networks. Guo et al.14 employed automatic feature extraction based on genetic programming (GP) to classify epileptic seizure signals. In 2014, Kumar et al.15 classified EEG signals based on features computed via fuzzy approximate entropy (fApEn) and DWT. Senhadji et al.8 used wavelet transform for time-duration analysis and a time-frequency approach to analyze the spectral content of EEG signals as a function of time. The same authors utilized both approaches to detect interictal spikes and to determine the ictal period. Tzallas et al.16 employed time-frequency analysis using many time-frequency distributions including short-time Fourier transform (STFT) with four different classification algorithms for epileptic seizure detection. However, ANNs achieved high classification results. Das et al.17 proposed a seizure detection method using symmetric normal inverse Gaussian parameters of subbands of EEG computed in dual-tree complex wavelet transformation. SVM with radial-basis function (RBF) kernel was used for classification of seizure from nonseizure EEG patterns. Besides, chaotic feature extraction techniques have been used for epileptic seizure detection, such as largest Lyapunov exponent, Hurst exponent, and fractal dimension, reported by Hosseini et al.18,19 The combination of chaotic features and adaptive neuro fuzzy inference system classifier achieved 98.6% accuracy for classification of normal from interictal EEG patterns and 98.1% accuracy for separating ictal from normal signals. Chen et al.20 proposed a method to reduce the computational load for the classical wavelet transform. Further, ANNs and logistic regression were used to classify epileptic seizure activity. Xie et al.21 developed a sparse functional linear model based on wavelets to represent EEG signals using wavelet variance to capture discriminative components in EEG signals. Acharya et al.22 decomposed EEG signals into wavelet coefficients by using wavelet packet decomposition. They removed eigenvalues from these coefficients with the help of principal component analysis (PCA). ANOVA of eigenvalues then identified those of significance. Further classifiers were trained using 10-fold cross-validation techniques. The GMM classifier was also used to obtain high classification accuracy for epileptic signals. Wang et al.23 used wavelet packet entropy for feature extractions that achieved a hierarchical classification approach. Recently, Kumar et al.15 proposed a wavelet-based fuzzy-approximate entropy (fApEn) method to SVM for classification purposes. Discrete wavelet transforms decomposed EEG signals into subfrequency bands. The fApEN of each subband was then computed to assess the chaotic behavior of EEG signals. The highest classification level achieved used SVM with the RBF kernel.

Upadhyay et al.24 computed wavelet fractal features from wavelet coefficients of four subbands extracted using different wavelets, such as Haar, biorthogonal, Coiflets, and Daubechies, for epileptic seizure detection. Three classifiers, such as least square SVM, ANN, and random forest tree, were tested for classification of normal EEG (sets A, B) versus epileptic seizure (set E). The ANN achieved 100% accuracy with feature computed with db3 wavelet. Kumar and Kolekar25 employed different time- frequency domain feature extraction methods such as fractal dimension, zero-crossing, subbands energy, and variance. The two classes C (interictal) versus E (ictal) were tested with SVM and reported 98% performance of seizure selection. DWT with db4 wavelet was used for feature extraction and classifying the normal and epileptic epochs using ANNs by Kulasuriya and Perera.26 The data was decomposed up to the fifth level and 10 hidden layer neurons were used. The achieved classification accuracy was 86.67%. However, the authors employed other than Bonn datasets.

In clinical applications the diagnostic or detection system should have a high accuracy of detection. The existing literature suggests that the majority of studies were unable to achieve perfect results (100%) when attempting to detect and differentiate seizure activity from seizure-free EEG signals or seizure-free signals from normal EEG segments.

< Prev   CONTENTS   Source   Next >