Optimal configuration of multilayer perceptron neural network classifier for recognition of intracranial epileptic seizures

Background: This paper presents optimized configuration of multilayer perceptron neural network (MLP-NN) as a pattern classifier for recognition of intracranial electroencephalogram (iEEG) epileptic seizures. Though qualitative analysis of intracranial recordings from the patients involves cumbersome procedure, it provides significant neuronal activities of the brain that is essential for clinical decision making. Methods: This study considered three epileptic seizures stages, namely, pre-ictal (set C), post-ictal (set D) and epileptic (set E) iEEG from the University of Bonn, Germany database. Four entropies, Shannon, log energy, spectral and Renyi entropies were considered as features to evaluate MLP-NN network. Four classification tasks with the dataset were formed, namely CE, DE, CDE, and CD. In order to identify the optimal configuration of MLP-NN for classification, network parameters such as input-hidden layer activation/transfer functions, hidden-output layer activation/transfer functions, network training/learning functions, the number of hidden neurons were considered and the optimality was arrived based on the mean square error (MSE), classification accuracy (CA). The efficiency of the entropy features was exploited by the parameters, mutual range of coefficient (gamma), p-value, and z-score, which showed the band discrimination for various classification tasks. Results: Simulations results showed that the tan sigmoid, pure linear were found to be optimal input hidden, hidden-output transfer function and Levenberg-Marquardt learning algorithm as optimal training function for all the four classification tasks. It was inferred from the proposed study that the CA indirectly varies with gamma value, p value, and MSE, and directly varies with z-score. From the experimental study, the best CA of 97.68%, 94.56%, 84.58%, and 57.8% was obtained for case CE, DE, CDE, and CD respectively. It can be concluded that proposed features with optimally configured MLP-NN found to be helpful for real-time iEEG classification. (C) 2017 Elsevier Ltd. All rights reserved.