StreXNet: A Novel End-to-End Deep-Learning-Based Improved Multilevel Mental Stress Classification From EEG Sensors

Mental stress has emerged as a psychosocial issue that leads to functional impairment in routine work. It raises the risk of stroke, coronary disease, depression, and hypertension. Evidence suggests a significant impact on cortical brain activity during stressful events. If these brain activities are detected timely, the negative outcomes of stress can be reduced by coping mechanisms, personalized interventions, and timely support. In addition, detection of multilevel stress may facilitate early interventions. Electroencephalography (EEG) sensors have been a potential marker for mental stress classification. However, the heterogeneity in manual feature extraction across researchers makes it difficult to standardize the classification pipeline. Moreover, the feature extraction process requires expert knowledge. To eliminate the need for feature extraction, the present study proposes an end-to-end multilevel mental stress classification from 64-channel EEG sensors using the fusion of extreme gradient boosting (XGBoost), squeeze excitation (SE) block-enhanced convolutional neural network (CNN), and long short-term memory (LSTM) with self-attention-based deep learning model, StreXNet. The results illustrate the improved classification accuracy, in comparison to previous studies, of the proposed StreXNet model with 91.60 +/- 3.01% and 91.81 +/- 4.26% for two-class and four-class classification, respectively. The one-way repeated measure ANOVA shows a significant difference [F(4, 68) = 20.261; p = 0.000] and [F(4, 68) = 46.625; p = 0.000] in classification accuracy between the proposed StreXNet model and other network architectures during two-class and four-class classification, respectively. These findings indicate that the proposed StreXNet deep learning model is more robust and effective for end-to-end classification of multilevel mental stress from EEG sensors.