Multi Class Classification of EEG Signal for Epilepsy Disease Detection Using Deep Learning

Abstract

This work is related to a real life health issue i.e. Epilepsy. Epilepsy disease is a neurological disorder which causes generation of an abnormal signal in brain causing uncontrollable jerking movements, temporary confusion and loss of awareness and may even cause death. Electroencephalogram (EEG) is used to read brain’s behavior by appens due to a sudden generation of abnormal electrical signal so it is difficult to diagnose. Different techniques and approaches have been introduced to overcome this problem which is generally classified into two groups: conventional methods (Machine learning) and Deep learning methods (CNN &LSTMs). From the previous literatures we find that conventional methods achieve remarkable accuracy on seizure dataset so rather than using the same method used previously we researched and analyzed that different stages of epilepsy have never been categorized before specifically the seizures during the non conscious state and the conscious state. Identification of seizures during NCES is important because comma or absence state causes no focal seizure symptoms hence it cannot be analyzed visually such long term activity in a brain of a coma patient may cause life time brain damage or even cause death. To extract complete information of epilepsy from EEG five stages including convulsive epileptic seizures , Non convulsive epileptic seizures (NCES) , pre-ictal stages and post ictal stages which have never been categorized thus hinders the power to actually understand these signals as different stages of Epilepsy. Keeping in view these different stages of epilepsy might contribute in better understanding of disease and help in inventing treatments that can be effective during any of this stage multi classification is carried out. As time series is complex in nature, and it’s difficult to identify the important features out of them because not all of this information contributes towards predicting the outcomes. Through learning the various techniques of feature extraction from literatures we focused on feature minimization and applied entropy based feature extraction approach. We use a single feature named SampEn for Classification. The calculated feature is fed to SVM as an input for classification. The CNN and LSTMs models are trained for classification are highly optimized [both in terms of Accuracy and speed] than the previous state of the art architectures. The error rate calculated on the output our model is compared with the results of previously implemented models and it is noticed that the results of the previously implemented techniques are same or less as resulted by our proposed approaches. Comparative analysis is carried out between all the applied approaches where CNNs performed better and they achieves an accuracy as high as 99.0 ± 0.1 in case of binary classification for all the stages and 87.5 99.0 ± 0.1 in case of multi five class classification. A summary of the experimental approach used in our work is shown in blocked