An Extensive Analysis on Deep Neural Architecture for Classification of Subject-Independent Cognitive States

Abstract

Human mental state can be measured by analyzing and understanding EEG (Electroencephalogram) signal in various applications such as neuro-science, brain-computer interfaces, etc. It is an important area of research where machine learning algorithms are being used to develop tools for mental state classification. The modern deep learning algorithms can be used on large EEG data set after applying the data augmentation process on them. In this paper, we apply the Deep Belief Network (DBN) model based on the Restricted Boltzmann Machine (RBM) for unsupervised feature learning of EEG signals to extract salient features for classification. This DBN model provides an unsupervised taxonomy-based system without human intervention. The efficiency of this model is evaluated on the ambulatory EEG signal with other deep learning algorithms. Experimental results demonstrate that DBN with Recurrent Neural Network-Long Short Term Memory (DBN-RNN-LSTM) provides an accuracy of 98.3% which is better than RNN-LSTM and other classical machine learning algorithm.