EEG stress classification based on Doppler spectral features for ensemble 1D-CNN with LCL activation function

The paper proposes an induced stress classification algorithm that uses features from the Doppler spectrum. In this approach, a reference signal source is used to obtain the quadrature and in-phase components of the EEG signal. The higher frequency components from the in-phase and quadrature are then eliminated using a pair of low-pass filters. The Doppler spectrum was then constructed from which the Doppler frequency features are then estimated. The features that were thus obtained are trained using an ensemble 1D D-CNN (one-dimensional Convolutional neural network) which uses two sections of 1D D-CNN. The first section 1D-CNN trains the features based on the EEG signal classes namely Stroop test, arithmetic task, and mirror tasks, while the second 1D D-CNN section trains the features based on the EEG signal intense classes namely high, low, and medium stress. We also propose a linear-cosine-linear (LCL) activation function for the ensemble 1D D-CNN which was derived from the cosine signal. The proposed stress classification scheme was evaluated using the SAM-40 datasets with induced stress classes namely arithmetic task, Stroop color-word test, and mirror image recognition task with stress levels namely high, low, and medium with the evaluation metrics such as precision, F1-score, accuracy, specificity, and recall. The proposed stress classification scheme attains an average accuracy, precision, and recall of 95.25%, 95.22%, and 92.9% when evaluated in 9 classes of EEGs.