Breaking Barriers in Stress Detection: An Inter-Subject Approach Using ECG Signals

Shahane Tigranyan, Arman Martirosyan

Abstract: Stress has become a ubiquitous challenge, affecting individuals mental and physical health significantly. Early detection of stress can aid in timely interventions, mitigating long-term health risks. Several studies focus on inter-subject stress classification. However, these studies often face limitations due to the small size and non-diversity of datasets. This paper introduces a novel framework based on deep learning for inter-subject stress detection utilizing electrocardiography data. We propose a two-step methodology to increase stress classification accuracy. The study involves the application of self-supervised contrastive learning and a data pseudo-labeling approach. We utilized PTB-XL, a large ECG dataset, to create a diverse training dataset. The model shows 93% by the G-mean metric on the publicly available WESAD dataset. This approach does not require the presence of ECG samples from the participant in the training set in order to correctly identify stress manifestations in the ECG.