Investigation of Optimal Components and Parameters of the Incremental PCA-based LSTM Network for Detection of EEG Epileptic Seizure Events

Abstract

Prediction of Epileptic seizures is highly imperative to improve the epileptic patient’s life. Epileptic seizures occur due to brain cells excessive abnormal activity that leads to unprovoked seizures and may occur without prior notice. Therefore, preventive measure that monitor and alert the possible occurrence of the seizures is paramount. Commercial and clinical available epileptic seizure computer aided detection system that utilized deep learning algorithms suffers from many challenges. These challenges ranges from low accuracy and precision, sensitive to artifacts and noise, among others. To enhance and increase the accuracy and optimal performance of these networks, this paper endeavor to investigate various optimization algorithm to optimized the network components and parameters in the developed incremental Principal Components Analysis based Long Short-Term Memory (Inc-PCA-LSTM) network for the detection and classification of Electroencephalograph (EEG) epileptic seizure signals based on the big data scenario. The model proved to be effective in the characterization of seven seizure events. The Adam, Elu, Orthogonal, and L1/L2 performed better than their counterparts in optimization functions, activation functions, initialization functions, and regularisation techniques respectively. The accuracy values of 97.5%, 97.5%, 98.4%, and 98.5% was obtained for each of the mentioned core components receptively.