STRESS CLEAR-NET: DEEP LEARNING FOR EARLY MENTAL STRESS DETECTION USING FILTERED EEG

Abstract

Mental stress is a pervasive factor affecting cognitive health and productivity. Early detection of stress using electroencephalogram (EEG) signals offers a promising preventive approach, yet challenges such as signal noise and inconsistent prediction accuracy limit the effectiveness of current models. This study proposes StressClear-Net, a novel deep learning framework that systematically integrates adaptive signal preprocessing techniques with a convolutional neural network (CNN) for enhanced early-stage mental stress detection. Two benchmark EEG datasets, STEW and Neurocom, were utilized to validate the framework. Comparative analysis across six filtering techniques—Adaptive Noise, Butterworth, Gaussian, Moving Average, Savitzky-Golay, and Median—demonstrated that the Median filter combined with StressClear-Net achieved a test accuracy of 80%, significantly outperforming the baseline unfiltered model, which achieved only 35% accuracy. Notably, the Median filter consistently yielded an improvement of over 45% in prediction accuracy compared to models without preprocessing. These results underline the critical importance of targeted signal enhancement prior to deep learning classification. StressClear-Net paves the way for robust and real-time EEG-based mental stress detection, offering substantial improvements in prediction reliability. Future work will extend this methodology to the Neurocom dataset and explore optimization strategies for deployment in real-world, resource-constrained environments.