In the chapter, a novel yet simple method for classifying EEG signals associated with normal and epileptic seizure categories has been proposed. The proposed method is based on empirical wavelet transform (EWT). The non-stationarity in the EEG signal has been captured using EWT, and subsequently, the common minimum number of modes have been determined for each EEG signal. Features based on amplitude envelopes of EEG signals have been computed. The Kruskal-Wallis statistical test has been used to confirm the discrimination ability of feature space. For classification, various classifiers, namely K-nearest neighbor (KNN), support vector machine (SVM), and decision tree (DT), have been used. The maximum classification accuracy of 98.67% is achieved with the K-nearest neighbor (KNN) classifier. The proposed approach has utilized only two features, which makes the proposed approach simpler. The proposed approach thus can be used in real-time applications.
TopIntroduction
Epileptic seizures are human brain disorders that can affect individuals form all ages. Globally, approximately 45.9 million people have epilepsy (Beghi et al., 2019). To access neurological activities, electroencephalogram (EEG) signals are used. The EEG signal’s spikes are frequently considered as the indicator of brain disorders like epilepsy (Mukhopadhyay & Ray, 1998; Ray, 1994). The classical strategy to diagnose epilepsy from EEG signals involves the expert’s visual inspection, which is often time-consuming and cumbersome. Therefore, various methods to process EEG signals have been reported for detecting epileptic seizures automatically in the literature.
Broadly, these methods make use of time-domain analysis, frequency domain analysis, and time-frequency analysis (Acharya et al., 2013). Line length and energy features have been used to detect epileptic seizures from EEG signals (Tessy et al., 2017). Features based on linear prediction (LP) have been proposed to detect epileptic seizures from EEG signals (Altunay et al., 2010). In (Ghosh-Dastidar et al., 2008), principal component analysis (PCA) has been used for the detection of epileptic seizures from EEG signals. With the presumption of stationarity in the EEG signals, authors in (Polat & Güneş, 2007; Srinivasan et al., 2005), have proposed time-domain and frequency-domain features to classify epileptic seizures from EEG signals. The EEG signals have non-stationary behavior (Boashash et al., 2003; Nishad & Pachori, 2020).
Taking the non-stationarity behavior of EEG signals into consideration, authors in (Tzallas et al., 2007, 2009) have developed time-frequency-based techniques to analyze and classify epileptic seizures from EEG signals. Numerous approaches based on wavelet and multi-wavelet transform have been developed to classify epileptic seizures from EEG signals (Akut, 2019; Anuragi et al., 2021; Khan et al., 2020; Nishad & Pachori, 2020; Sharan & Berkovsky, 2020; Uthayakumar & Easwaramoorthy, 2013; Zeng et al., 2020). Moreover, wavelet coefficients with phase space reconstruction (PSR) based features have been utilized to construct feature space to detect epileptic seizures from EEG signals (Lee et al., 2014). Various non-linear features inspired from chaotic signal processing, like, correlation dimension (Silva et al., 1999), Lyapunov exponent (Swiderski et al., 2005), and approximate entropy (ApEn) (Gupta & Pachori, 2019), are effective in detecting epileptic EEG signals (Srinivasan et al., 2007). Additionally, the Lyapunov exponent has been used to describe EEG signal’s chaotic nature; consequently, the Lyapunov exponent has been used to discriminate epileptic seizure and seizure-free EEG signals (Swiderski et al., 2005). To quantify the neural functioning during interictal (seizure-free) and ictal (seizure) activities, the correlation dimension from the EEG signal has been used (Lehnertz & Elger, 1995). Furthermore, the human brain’s ictal and interictal activities have been successfully discriminated against using fractal dimension parameters (Accardo et al., 1997). The higher-order spectral (HOS) analysis of EEG signals has been carried out to classify normal, ictal, and interictal activities in the human brain (Chua et al., 2007). Authors in (Acharya et al., 2011) have proposed features obtained from recurrence quantification analysis (RQA) to classify normal, ictal, and interictal EEG signals. To discriminate EEG signals in normal, pre-ictal, and epileptic seizure categories, authors in (Acharya et al., 2009) have used Hurst exponent, correlation dimension, and ApEn as features, and promising results were obtained. In (Wang et al., 2010), the ApEn parameter was computed, and a significant difference in ApEn values has been observed for epileptic seizures and normal EEG signals. In (Liang et al., 2010), the authors have used the spectral features with ApEn to classify EEG signals in normal, ictal, and interictal categories. Empirical mode decomposition (EMD) has been found to be useful in the analysis of EEG signals (Oweis & Abdulhay, 2011). In (Oweis & Abdulhay, 2011), it has been found that the weighted mean frequency of intrinsic mode functions (IMFs) differs significantly for EEG signals associated with normal and epileptic seizure categories. In (de la O Serna et al., 2020), the authors have suggested the filter bank approach based on the Tylor-Fourier series for analyzing EEG signals.