EEG Stress Detection Using Deep Learning

The paper introduced MSA-CBL (Multi-Scale Attention CNN-BiLSTM) which is an end-to-end network for robust stress detection using EEG signals. It merges four complementary parts together into one differentiable pipeline, which are: multi-scale parallel convolution, squeeze and excite channel attention, bi-directional LSTM temporal modeling, and additive self-attention. MSA-CBL's performance is competitive with the highest state of the art method in literature and surpasses the other 9 baseline methods with accuracy of 92.4% and F1-score of 96.1% on SAM-40 dataset. In future work, multi-level stress estimation in continuous-label datasets and generalization across datasets through domain adaptation and transfer learning will receive attention. Additionally, multimodal extensions that utilize other physiological measurements, such as heart rate variability (HRV), electrodermal activity (EDA) and respiration measurements will be explored, to improve the development of a more holistic system of stress monitoring. Future research will aim at individual-level explainability using formal attribution techniques, like Integrated Gradients or SHAP; clinical validation of wearable EEG systems with dry electrodes in ambulatory environments; and strategies for personalization, reducing inter-subject variability.

[1] D. Adiele et al., “Association of academic stress, anxiety and depression with sociodemographic among medical students,” Int. J. Soc. Sci. Stud., vol. 6, no. 6, 2020. [2] H. M. Afify et al., “Stress detection based EEG under varying cognitive tasks using convolution neural network,” Neural Comput. Appl., vol. 37, no. 7, pp. 5381–5395, 2025. [3] J. Agrawal, M. Gupta, and H. Garg, “Early stress detection and analysis using EEG signals in machine learning framework,” IOP Conf. Ser. Mater. Sci. Eng., vol. 1116, no. 1, p. 012134, 2021. [4] A. Akella et al., “Classifying multi-level stress responses from brain cortical EEG using machine learning autoencoder,” IEEE J. Transl. Eng. Health Med., vol. 9, p. 2200109, 2021. [5] A. N. Alghamdi, “Stress and emotion recognition in natural speech in the work and family environments,” Ph.D. dissertation, RMIT Univ., Melbourne, Australia, 2023. [6] A. Author et al., “A multimodal approach for stress detection through questionnaire and emotion analysis,” in Proc. ICCCNT, pp. 1–6, 2024. [7] S. Bhatnagar, S. Khandelwal, S. Jain, and H. Vyawahare, “A deep learning approach for assessing stress levels in patients using electroencephalogram signals,” Decis. Anal. J., 2023. [8] R. Brown and D. Wilson, “Stress and emotion open access data: A review on datasets, modalities, methods, challenges, and future research perspectives,” J. Healthcare Inform. Res., vol. 9, pp. 233–257, 2025. [9] M. Chen and L. Zhang, “Real-time stress detection using LSTM networks for emotional and psychological insights,” Comput. Res. Dev., vol. 25, no. 2, pp. 101–112, 2023. [10] K. Cooper and J. Wojcik, “The effects of physical activity classes on stress in college students,” Winthrop McNair Res. Bull., vol. 4, no. 4, 2018.