Comparative Analysis of Deep Learning Models for the Detection of Epileptic Seizure

Abstract

Electroencephalogram (EEG) is used to detect epilepsy, a common neurological disorder. Neurologists visually examine EEG results to make the diagnosis. Researchers have suggested automated technologies to diagnose the seizure because traditional method are lengthy and there is a dearth of professionals everywhere. The common symptoms of seizures, which are characterized by aberrant brain activity brought on by an epileptic disease, include bewilderment, loss of awareness, and strange behaviour. Sometimes it becomes very difficult to identify the seizure in persons. So, for determining seizures there are many deep learning models have been designed. Among those, three models have been chosen and compared in this paper. These three models are, Convolutional neural network-long short-term memory (CNN-LSTM), convolutional neural network-recurrent neural network (CNN-RNN), and convolutional neural network-gated recurrent unit (CNN-GRU) whose comparison study have been discussed in this paper by using three different types of optimizers, namely Rmsprop, Adam, and Nadam. After that the result of deep learning models have been compared with some previous machine learning work for the detection of epileptic seizure. Mainly three parameters such as accuracy, sensitivity and specificityof the models are found and compared to predict which model as well as which optimizer among Rmsprop, Adam and Nadam is best. For efficient removal of the features from an EEG sequence data, one dimensional convolutional neural network (CNN) is created. For further extraction of temporal characteristics, the features extracted are processed by the CNN-LSTM model's LSTM layers, CNN-RNN model's RNN layers, and CNN-GRU model's GRU layers. The last epileptic seizure recognition step involves feeding the output characteristics into a number of fully connected layers. I