EEG-based stress identification using oscillatory mode decomposition and artificial neural network

Abstract

The stress identification offers an understanding of the user's mental state. Therefore it has the potential to greatly improve the Human-Machine Interaction (HMI). In the recent era of industrial, assistive, and healthcare applications, a new trend is to use the “Artificial Intelligence” (AI) powered multimodal signal processing to detect the stress level. This work presents an original approach for classification of stress and nonstress subjects. We have used the multichannel Electroencephalogram (EEG) signals for this categorization. Intrinsic mode function (IMF) is obtained from these EEG signals using the “Empirical Mode Decomposition” (EMD). The “Variational Mode Decomposition” (VMD) is also applied to get the modes from EEG signals. For the selected IMFs and modes, the “Second-Order Difference Plots” (SODPs) are traced. The shape of these SODPs is used to distinguish the stress and without stress categories. The feature space is derived from first 7 IMFs and modes. This includes areas, “Central Tendency Measures” (CTMs), and means of SODPs. A publicly available EEG signals dataset is used to test the applicability of the work. We have used the machine learning algorithms such as the “Support Vector Machine” (SVM), “Multilayer Perceptron Neural Network” (MLPNN), and Boosting with “Random Forest” (RF) for an automated categorization. A detailed performance analysis is conducted for individual channels, subset of channels, and lobe-wise. The highest attained accuracy at subset level is 99.89%, channel-wise is 98.89%, and lobe-wise is 99.99%.