Automated Epilepsy Seizure Detection from EEG Signals Using Deep CNN Model
Introduction
Roughly 50 million people are suffering from epilepsy globally, according to the study by the WHO (World Health Organization) in 2017 [I]. Approximately 10% of people are affected with epilepsy every year [2]. Epilepsy is a neurological disorder wherein there is an uncontrolled electrical discharge of neurons. Our whole brain is a biological neural network. The primary unit of the neurons system is the cell. Every neuron is made of two parts: axon and cell body dendrites. Neurons transmit signals throughout the body. Epilepsy can affect anyone at any stage of life. Epileptic patients experience a vast range of symptoms which largely depend on the portion and the area of the brain that is affected. Epileptic seizures are of potential harm since they are often responsible for physical, social consequences and psychological disorders, which may result in loss of consciousness, injury to the brain and, in certain cases, abrupt death [1].
Normally if seizure is found active in one section of the brain cell or tissue, then it may spread to remaining sections of the brain. If a clinical seizure were experienced by the patient, the neurologist would directly visualize and inspect whether the EEG signals are normal or abnormal [3, 4, 5, 6]. However, this process consumes too much time. Even an entire day may be insufficient to adequately visualize and inspect the patient reports. It may also require secondary neurology experts to help them in this area [7]. Although we are not replacing the neurologist expert, we can help them in reducing the time consumed for visualizing the report.
All activities occurring in our brain signals are detected by EEG. There are small metal discs or electrodes, which consists of wires that are to be placed on the scalp. The electrodes are mainly responsible for acquiring the electrical activity in form of signals from brain cells and hence mapping the pattern in the brain due to electrical activity. This is very efficient for diagnosis of several conditions ranging from minimal to severe harm like headaches, dizziness, epilepsy, sleeping disorders and deadly brain tumors. The very first EEG measurements were done by Hans Berger, who was a German psychiatrist, in the year 1929 [8]. EEG signals prove to be of great use in the detection of epileptic seizures since any rapid changes from the normal pattern could indicate the presence of a seizure. EEG signals get small- scale amplitudes of the order 20 pV [9]. The signals observed in the scalp are divided into four bands based on their frequencies namely: Delta in the range of 0.3—4 Hz; theta in the range of 4-8 Hz; alpha in the range of 8-13 Hz; and beta in the range of 13-30 Hz [8, 10]. An electrode is always placed in accordance with the 10-20 international system which is depicted in Figure 2.1. This system is used for electrode placement, where the respective electrode is placed at either (10 or 20) % of the total distance among the notion and the inion. In 10-20 placement, every electrode is marked with a letter and that followed by a number. Electrodes contain odd and even numbers. Odd numbers show that electrode is located on the left side of head. Even number shows that electrode is located on the right side of head. The letter indicates the area of the brain where the electrode is placed: F letter show that frontal lobe, T letter show that temporal lobe, P letter show that parietal and О letter show that occipital lobes, Letter Z denotes that the electrodes are positioned on the midline of the brain [11, 12].
Artificial Neural Networks (ANNs) were first developed several decades ago. by researchers attempting to develop the learning process of the human brain. ANN is typically composed of interconnected “units” which denote the modeled neurons [11, 12] in 2004, Nigam and Graupe [13, 14, 15, 16] presented a novel approach for the detection of epileptic seizures from the EEG recordings. It employs the Diagnosis of EEG signals of ANN in combination with a multistage nonlinear preprocessing filter. Kannathal et al. [12] Compared different entropy measures that are tested on EEG data and it has been proven that EEG data using ANFIS classifier have obtained an accuracy of 90%. Guo et al. [17] put forward a technique which uses Relative Wavelet Energy (RWE) for analysis of EEG signals which are then classified using ANNs. This method has achieved an accuracy of 95.20%. Homan et al. [18] in 2000 proposed an Epileptic Seizure Prediction system which depends on RNN. Guler et al. [60] proposed an EEG signal classification system that depends on Recurrent Neural Networks using Lyapunov exponents which has achieved an accuracy of up to 97.38%. Talathi [19] in 2017 presented an Epileptic
FIGURE 2.1 Standardized 10-20 electrode placement system.
Seizure Detection and classification established on Deep RNN which has obtained an accuracy of 99.6%. Nowadays, when millions of data comes into the clinical area for better accuracy, they use deep learning algorithms. Further Pereira et al. [20] have done automatic segmentation of brain tumors w'ith the help of the convolutional neural network. Acharya et al. [21] in 2017 proposed an application that detection automated infraction using ECG signal that given 92.50% accuracy. Acharya et al. [22, 23] showed a deep Convolutional neural network in ECG signals for finding coronary artery disease that has acquired an accuracy of 94.95%.
ANN, which works on the idea of neural networks, is the backbone on which DL is based. For handling a gradient descent algorithm effectively, a back propagation algorithm is a very good approach w'hich can be applied on any dataset. Although it has a very promising training accuracy, the testing accuracy seems detrimental. BPA faces the problem of local optima when its application is put to effect on random initialize node. In huge datasets, there is a problem of over-fitting as well. Flenceforth, popular learning algorithms namely SVM, KNN. decision tree, and logistic regression are employed to achieve the global optima [3, 4, 24, 25].
The main contribution in the view of this paper, a novel deep CNN model is proposed for classifying EEG signals. In this model feature selection and extraction processes have been done automatically by convolutional and max-pooling layers. There is no requirement of any specific handcrafted feature extraction and selection technique, which also reduces the computational complexity of our model. In this model, first the Bonn university EEG database is normalized and then split into training and testing datasets. We have used 10-fold cross validations for dense layer and back propagation algorithm to classify the dataset into three classes normal, pre-ictal and epilepsy.
