Stress Detection Through Sleep-Based Physiological Signals: An Integrated Semi-Supervised and Supervised Approach

Abstract

Stress significantly impacts both physical and mental health, often manifesting in physiological indicators such as elevated heart rate, reduced heart rate variability, and poor sleep quality. Wearable devices, such as smartwatches, provide continuous monitoring of these stress-related metrics but face challenges such as the lack of robust methods to process multi-modal sensor inputs and the difficulty in generalizing models across diverse populations due to limited labeled data and individual variability in stress responses. To address these challenges, this study introduces an integrated stress detection framework, combining semi-supervised learning for data refinement and supervised learning for stress classification. In the data pre-processing phase, feature selection and mathematical transformations were used to process a public sleep heart health dataset. By integrating a self-training classifier, the proposed approach enhanced stress label quality, mitigating the impact of limited labeled data and physiological variations in stress detection models. We trained multiple machine learning models, including Random Forest, Decision Tree, and XGBoost, and achieved up to 99% accuracy and a 97% F1-score, demonstrating high reliability in stress detection. %This study presents a reliable and accurate stress detection using a public dataset and machine learning techniques. The results emphasize the importance of data quality improvement, and demonstrate the potential of combining semi-supervised learning with supervised algorithms for enhanced stress detection.