Automated attention deficit classification system from multimodal physiological signals

Abstract

Lack of attention, if it could not be taken care of and persists for a long time then may lead to a severe issue. Analysis of Electroencephalogram (EEG) signals can effectively measure attention and its deficit. This paper proposed an efficient classification system to analyse and predict cognitive attention or its deficit with less computational power and adaptable in real-time. EEG signals have been split into six windows of varying time duration. Robust and computationally less expensive features hurst and power have been used for the designing of feature space. Objective of this proposed work is to provide robust methodology for classification of attentive and non-attentive category of subjects for real time screening. The robust classifier has been designed by multi-layer perceptron neural network and tuned with primary parameters and hyper-parameters using Adam optimisation. Gradient descent has been used for backpropagation. Hurst component of the signal has provided the self-similar characteristics. The features’ significance has been tested using the Wilcoxon signed-rank test. The experimental results have revealed that the proposed hybrid classification model could distinguish between an individual’s cases not being attentive and being attentive with accuracy of 88.04% at temporal lobe.