Contents I. Introduction
Epileptic seizure is a common neurological disorder of the brain. The study of electroencephalogram (EEG) is a well established technique to sense the epileptic seizures. The EEG data is usually analyzed through visual inspection by a neurologist. However, visual inspection of seizure activity in EEG signals is a time taken and inaccurate process. Thus, the automated seizure detection system is in high demand to assist neurologists, speed up the diagnosis process and, thereby, improve the life span of the patient [1]. Hence many researchers have presented these automatic seizure detection studies by analyzing the time and frequency domain features of EEG [2]. The automatic seizure detection system comprises two key sequential steps: feature extraction from EEG signals and classification. In the first phase, feature extraction has a direct impact on both the precision and complication of the entire system. Thus the EEG signal has to be processed with a more desirable technique to extract the features. Earlier techniques based on empirical mode decomposition (EMD) [2], short-time Fourier transform [3], and wavelet transforms [4] provided successful results, but all these techniques are handcrafted feature extraction methods. In recent times, deep learning techniques are given promising results in the field of image and signal processing [5], [6]. The main advantage of the deep learning model is the automatic feature extraction. For seizure detection using EEG, the features obtained by the deep learning algorithm such as CNN [7], [8] have been shown impressive results in detection efficiency than the traditional handpicked features. The second phase, classification directly depends on the previously extracted features. Some of these classification algorithms are support vector machines (SVM), naive Bayes (NB), K-nearest neighbour (K-NN). The features are considered as a higher priority because the entire system performance and the classification results will depend on the features extracted. The past researches [2] shown that the spectral characteristics of the EEG signal are having more consistency in epileptic seizure detection. Considering the spectral information of epileptic EEG signals is the finest method for the analysis of seizure movement in EEG signals. The spectral information of EEG is a mixture of different frequency sub-bands. The frequency sub-band signals of EEG are delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz) and gamma (30–80 Hz). The features obtained from the frequency sub-bands are shown better results compared with the entire frequency band [9], [10]. Adeli et.al [9] applied a Wavelet-Chaos Methodology on frequency sub-bands of EEG signal for seizure detection. The authors obtained EEG variations such as correlation component and the largest Lyapunov exponent for the study of different seizure EEG classes. In [10], the time-frequency features are extracted from the frequency sub-bands of EEG signals for the analysis of seizure EEG signals and they have achieved an overall accuracy of 98.7%. These works of literature used manual feature extraction techniques to design an automatic seizure detection system. Although the above methods have improved EEG-based seizure detection, there are still few problems such as manual feature extraction with the entire frequency band of EEG signal which may not give always the best results because it adopts that the entire frequency bands share the same importance [11]. The proposed method is different from these existing methods. In this study, we adopted a deep learning CNN architecture with frequency sub-band signals as an input for automatic feature extraction. The features are obtained from multi-scale frequency sub-bands and are classified as different seizure and healthy classes.
