A Deep Learning Framework for Quantitative Stress Estimation using Wearables
Loading...
Authors
Moustafa, Israa Ashraf Abdelrahman Abdou
Date
2025-07-08
Type
thesis
Language
eng
Keyword
Stress , Wearable sensors , Deep learning , Physiological signals , Electrodermal activity (EDA) , Heart rate variability (HRV) , Affective computing , Exploratory Factor Analysis
Alternative Title
Abstract
Repeated exposure to everyday stressors without sufficient recovery can progressively lead to chronic stress and depression, highlighting the importance of early stress monitoring and prevention. Although advances in wearable devices have enabled stress detection, most methods adopt discrete classification, introducing several critical limitations. Discretization masks individual differences and reduces statistical power by ignoring within-group variance when studying physiological stress associations. It also weakens reliability by averaging out fluctuations and reduces temporal sensitivity for real-time monitoring.
This thesis presents a framework for estimating a continuous Quantitative Stress Index (QSI) by integrating psychometric modeling with physiological features from wearables.
As a first contribution, a continuous questionnaire-based score, QSI_Quest, is constructed using Exploratory Factor Analysis (EFA) on standardized self-report measures, capturing four factors representing stress dimensions. These are aggregated using variance weighting to yield a robust, personalized stress score that outperforms the discriminative power and reliability of conventional questionnaire-derived scores.
Building on QSI_Quest as a reference, two deep learning architectures, QSI-MLP and QSI-CNN, are proposed and evaluated for continuous stress estimation from physiological signals, including electrocardiogram (ECG), blood volume pulse (BVP), and electrodermal activity (EDA). The models are benchmarked against baseline machine learning methods and established quantitative indices, demonstrating improved session-level discrimination and smoother temporal trends. In classification tasks, the ECG-EDA and BVP-EDA configurations achieved high mean accuracies of 98.81% and 98.58%, respectively, in distinguishing stress from non-stress sessions. In regression on a 0–100 stress scale, both models yielded mean absolute errors below 10, with QSI-CNN performing slightly better (ECG-EDA: 9.58; BVP-EDA: 9.82). The comparable performance of BVP-EDA to ECG-EDA underscores the practical potential of wrist-worn systems for real-world stress monitoring.
By modeling stress as a continuous, multidimensional construct and integrating psychometric and physiological modalities, the proposed framework enables more realistic, personalized, and temporally sensitive stress estimation. Physiological signals capture intra-session variability, reflecting natural fluctuations in stress responses, while psychometric inputs provide subject-specific baselines. This integrated approach supports high-resolution monitoring and has broad applicability in clinical decision-making, adaptive human-computer interaction, and personalized health management, particularly in affective computing and telehealth environments.