Classification of EEG signals for epileptic seizures using hybrid artificial neural networks based wavelet transforms and fuzzy relations

Highlights

• Classification of EEG signals for detection of epileptic seizures.

• Feature extraction with discrete wavelet transforms.

• Reduction the dimensionality of features using Fuzz Relations.

• Artificial neural networks based cross-entropy and information criteria.

• Genetic and gradient based algorithms.

Abstract

Epilepsy is one of the most common central nervous system disorders. Epileptic people suffer from recurrent seizures depending on many trigger factors such as genetic, physiologic, brain damage, etc. Epileptic seizures occur at unpredictable times and mostly without warning. During seizures, epileptic people have distraction or involuntary spasms that might even result in serious physical injuries or death. Therefore, the detection of epilepsy still challenges the neurologists in terms of prediction of seizure times and classifying the brain signals received from different zones of brain. This study presents an efficient procedure that provides an accurate classification of Electroencephalogram (EEG) signals for early detection of epileptic seizures. Essentially, this procedure hybridizes many tools such as artificial neural networks (ANNs), gradient based algorithms, genetic algorithms (GAs), feature extraction with discrete wavelet transforms (DWT) and fuzzy relations for reducing dimensionality of features. In analysis, ANNs are trained by the gradient based algorithms and GAs considering early stopping, cross-validation and information criteria. In order to ensure an accurate classification performance, the automated multi-resolution signal processing technique splits EEG signals into the detailed partitions with different bandwidths, and then decomposes them into detail and approximation coefficients by means of DWT at the different decomposition levels. Thus, some specific latent features that characterize the nonlinear and dynamical structures of EEG signals are acquired from these coefficients. The fuzzy relations bring out the significant components by reducing the dimension of feature matrix. To detect the epileptic behaviors in EEG signals, these selected components are processed by ANNs based cross-entropy and information criteria. According to analysis results, this approach not only allows making deeply analysis of EEG signals for detection of epilepsy, but also provides the best model configurations for ANNs in terms of reliability and complexity.

Introduction

Epilepsy is one of the most common neurological diseases and it takes fourth place in terms of incidence after migraine, stroke and Alzheimer's. Approximately 50 million people worldwide suffer from epilepsy, and nearly two out of every three new symptoms are detected in developing countries (Sharanreddy & Kulkarni, 2013). According to 2015 Annual Report of Epilepsy Foundation, 8.4 out of 1000 people have active epilepsy in the United States. However, these numbers are growing to 16.5 per 1000 people when people are asked if they have ever had epilepsy (England, Liverman, Schultz, & Strawbridge, 2012). Generally, epilepsy is a central nervous system disorder, so the patients suffer from recurrent seizures that occur at unpredictable times and usually without warning (Shoeb & Guttag, 2010). During seizures, the epileptic people have distraction or involuntary spasms concerned the whole-body. Also, these seizures may even result in serious physical injuries or death. To find out whether if a person has epilepsy, the specialists investigate the type of seizure or epilepsy syndrome. Actually, this process requires the detailed evaluations including the medical history, blood tests, Electroencephalography (EEG) tests, and brain imaging tests such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. These tools are able to provide useful information about the electrical activity of the brain and possible types of seizures. In the computer diagnostic techniques, EEG is still a useful test, because it is capable of reading the electrical activity of the brain without painful and invasive. Also, EEG is a very powerful tool with the voltage range 3-100 µV which is 100 times weaker than ECG (Electrocardiography) signals (Kaur, 2012). To record the brain activities, the electrodes connected the wires are directly attached the scalp. However, these wires do not transmit any electrical current to the scalp; they only make multi-electrode recording of time-varying the electrical activity in the brain. For this reason, EEG signals are widely used to examine the different brain activities safely.

To detect the epileptic seizure, EEG signals showing the brain's electrical activity in real-time can be separated into seizure and non-seizure periods by means of unsupervised time-series segmentation techniques. However, these techniques have some shortcomings about separating them into seizure and non-seizure periods accurately (Agarwal, Gotman, Flanagan, & Rosenblatt, 1998). These shortcomings can be eliminated by supervised techniques in the discriminative framework (Shoeb & Guttag, 2010). In the context of supervised framework, Andrzejak et al. (2001) analyzed EEG signals that consists of five data sets with 100 EEG segments each, recorded extracranially from healthy subjects with eyes open and closed, and intracranially from epileptic patients both during seizure-free periods and epileptic seizures. In this analysis, they showed the existence of a strong indication of nonlinear deterministic dynamics in EEG signals during seizure activities, and no significant similar behavior for healthy subjects with eyes closed.

Although EEG test is a useful tool for monitoring the brain's electrical activity and diagnosing epilepsy, discriminating and interpreting EEG signals visually requires the skilled specialists in terms of making more accurate inspections. Also, this treatment may fail to figure out some latent features within EEG signals that carry significant information related to past or future of epileptic seizures. Actually, such information cannot be obtained directly from the recorded signals; it can be masked by other biologic signals. Therefore, the signal processing is inevitable to magnify the relevant information and to extract latent features from it (Mainardi, Bianchi, & Cerutti, 2012). To do this, the automated tools are mostly used to separate EEG signals into the more detailed sub-series with a certain window-width and decompose them into the transformed sub-bands by means of signal processing techniques such as Lyapunov exponent, Fourier, Hilbert, Wavelet transforms (WTs), etc. Generally, this process is known as sub-band coding or multi-resolution signal processing.

Recently, WTs are often used to extract the latent features from EEG signals. Essentially, this technique provides the decomposition of EEG signals into the sub-bands with different scales at the certain decomposition levels. These sub-bands include some useful features that characterize the dynamical structure of EEG signals. These features can be used for detection of epilepsy or classifying EEG signals (Rosso et al., 2004, Guler and Ubeyli, 2005, Haas et al., 2007, Yalcın et al., 2015, Sharma et al., 2017, Bhattacharyya et al., 2017, Bhattacharyya et al., 2017). In the literature, current approaches mostly tend to extract different types of features depending on addressed problem and methodology such as various statistical moments and metrics as well as entropy measures (Kumar et al., 2015, Acharya et al., 2015, Acharya et al., 2011).

In the context of supervised learning, there are remarkable studies based on the classical statistical and artificial intelligence (AI) techniques for detection of epilepsy. Generally, these studies utilized Logistic Regression, Discriminant Analysis, Regression Trees, k-Nearest Neighbor, Gaussian Mixture Models, Naive Bayes Classifier, Support Vector Machines (SVMs), ANNs and ANFIS classifiers. According to analysis results, apparently the hybrid AI techniques show superior performance than classical statistical approaches, because they provide a flexible framework to estimate efficient models in the high dimensional and excessive nonlinear environments.

Especially, WTs have been integrated with various classifiers in the context of detection of epileptic signals. Because of their flexible and adjustable structures, ANN classifiers have been mostly preferred for the mentioned problems. For instances, the following studies quite a remarkable in terms of integrating WTs with ANNs for classification of EEG signals: (Akin et al., 2001, Nigam and Graupe, 2004, Guler and Ubeyli, 2004, Guler and Ubeyli, 2005, Guler and Ubeyli, 2007, Patnaik and Manyam, 2008, Guo et al., 2010, Martis et al., 2013, Dehuri et al., 2013, Rivero et al., 2013, Yalcın et al., 2015; Bihati et al. 2017; Amorim, Moraes Fazanaroa, Silvaa, & Pedrini, 2017).

However, there exist many researches in which various feature extraction methodologies have been used with other classifiers rather than ANNs in the context of feature extraction and classification of EEG signals (Ocak, 2009; Wang, 2011; Acharya et al., 2011, Kumari and Jose, 2011, Acharya et al., 2013, Kaur, 2012, Nicolaou and Georgiou, 2012; Martis et al., 2013, Joshi et al., 2014, Kumar et al., 2015, Sharanreddy and Kulkarni, 2013, Pachori and Patidar, 2014; Fu et al., 2015, Sharma and Pachori, 2015, Pachori, 2008, Pachori et al., 2015, Amorim et al., 2017, Sharma et al., 2017, Tiwari et al., 2017, Bhattacharyya et al., 2017, Bhati et al., 2017).