Enhancing the Reliability of Electrocardiogram Signals for Stress Management: Learning Transferable Models from High-Resolution Electroencephalogram Supervision

Stress, though often overlooked, is a significant driver behind the high suicide rates in South Korea, which leads among OECD countries with 24.1 suicides per 100,000 people. This issue is not unique to South Korea; in the United States, adolescent suicide rates have surged by 35% since 1999, now affecting 12 per 100,000 individuals. To address these alarming trends and prevent stress-related health issues, schools frequently rely on traditional checklist tests. However, these tests are often inaccurate, are time-consuming, and lack scientific rigor. Recent studies have started to explore the use of electroencephalogram (EEG) technology, which provides a scientific and quantitative measure of stress by closely monitoring brain activity. Despite its promise, electrocardiogram (ECG) signal, which also correlates with stress, has not been as thoroughly investigated. To address this problem, I propose an integrated approach that combines EEG and ECG assessments to develop a more reliable and cost-effective method for stress detection using machine learning. During the training phase, the proposed system takes both EEG and ECG signals as input and learns to map these signals into an emotion-related feature space. After training, the pre-trained network is then used to predict arousal and valence from ECG signals. Extensive experiments demonstrated that the proposed approach significantly improved performance, reducing RMSE by 8.24.