Evaluation of Deep Learning Architectures for EEG-Based Stress Detection with and Without SNR Augmentation

The classification of mental states using electroencephalogram technology is increasingly recognised for real-time stress monitoring. Nonetheless, the robustness of deep learning models could vary considerably when trained without augmentation methods like Signal-to-Noise Ratio (SNR) enhancement. This study investigates and compares the performance of three notable Convolutional Neural Network architectures-EEGNet, DeepConvNet, and ShallowNet on a binary stress classification problem utilising EEG inputs with and without SNR augmentation. Performance was evaluated utilising Accuracy, Area Under the Curve (AUC), and confusion matrices. Results demonstrate that ShallowNet surpasses the competitors, attaining the best classification accuracy (80.98%) and AUC (0.8932). DeepConvNet achieves an accuracy of 78 % and an AUC of 0.8662, whereas EEGNet records an accuracy of 67.93 % and an AUC of 0.811. Analysis of the confusion matrix indicates that ShallowNet exhibits the highest true positive and true negative rates, demonstrating exceptional generalisation capabilities even in the absence of data augmentation. These findings indicate that shallowNet may be effective for EEG-based stress detection tasks in low-preprocessing environments, rendering them appropriate for real-time and limited resource applications. without data augmentation. These findings demonstrate that shallow CNNs can be more effective for EEG-based stress detection tasks under lowpreprocessing conditions, making them suitable for real-time and resource-constrained applications.