Stress detection using the phase controlled Bi-channel adaptive features from the brain EEG signals

This work proposes a stress classification system from the electroencephalogram (EEG) signals collected from the stress subjects. The scheme extracts the phase-controlled Bi-channel adaptive features using a pair of EEG signals. The proposed adaptive features are derived from the least mean square adaptive filtering approach, in which a phase variation between the signal pair is used to control the feature amplitude. The adaptive feature that was estimated from the signal pair is trained by a 1D-convolutional neural network (1D-CNN) model, which classifies the intensity of induced stress, namely medium, low, and high, for the induced stress types: Stroop colour-word test, arithmetic test, and mirror image recognition test. The paper also proposes a dual activation and dual pooling structured 1D-CNN that contains two different activations, namely positive activation and negative activation functions. Further, the positive activation drives the max-pooling, while the negative activation drives the min-pooling. Thus, the dual activation and dual pooling structure based on 1D-CNN preserves both the positive and negative features. For the evaluation of the proposed stress classification, the SAM-40 dataset is used. The use of bi-channel adaptive features in stress classification results in accuracy, recall, and precision of 95.78%, 94.57%, and 93.72% respectively.