Data-driven approaches for context-specific proactive and reactive stress management

• Main Author: Xia, Ziqing
• Other Authors: Chen Chun-Hsien
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Stress management; Stress detection; Data-driven approaches
• Online Access: https://hdl.handle.net/10356/173669
• Institution: Nanyang Technological University

Stress, a significant mental health issue, is challenging to manage due to its complex and dynamic nature. This thesis explores data-driven approaches for proactive and reactive stress management, focusing on two specific contexts: English Online Education and Vessel Traffic Service (VTS), unfolded by four sub-studies. To facilitate proactive stress management in English Online Education, understanding the impact of stress on cognitive performance is crucial, enabling teachers to design targeted teaching strategies and specialized training for students. The first sub-study explores how stress impacts text processing abilities among first language (L1) and second language (L2) English readers, utilizes eye-tracking technology in an online learning context. The study reveals that stress reduces the efficiency of syntactic parsing and sentence integration in both L1 and L2 groups and uniquely impairs global text processing in L2 readers. Trait anxiety affects higher-level text processing in L1 readers, while L2 readers' processing relates mainly to their reading proficiency level. Addressing the limitations of existing stress detection methods in online education, the second sub-study develops a non-intrusive approach for simultaneously detecting stress states and cognitive performance based on eye-movement data and machine learning. Using LightGBM (Light Gradient Boosting Machine), a tree ensemble model, the method achieves an accuracy of 82.5% for detecting stress states and 79.3% for reading performance. An interpretable model, SHAP (SHapley Additive exPlanation) is used to identify the most influential eye-movement indicators and their correlation with stress and reading performance. For VTS scenarios, proactive stress management methods such as redesigning management systems, updating procedures, and developing training programs are critical. However, existing approaches lack methods that can inform stress analysis at the design stage. The third sub-study develops a neural network-enabled stress prediction model that can predict an operator's stress perception of a working scenario based on task, human, and environmental information. The model achieves a mean absolute error of 0.33 in a case study of VTS, overcoming individual differences through the integration of a personalization module and a Mixture-of-Expert module. The implementation of stress detection in VTS requires careful consideration of various factors including unobtrusiveness, privacy, timeliness, and accuracy. Wrist-mounted peripheral sensors emerge as an ideal stress detection tool for VTS, offering unobtrusive, real-time stress detection. Nonetheless, algorithms based on practical peripheral signals often struggle to achieve high stress recognition accuracy. To tackle this challenge, the fourth sub-study proposes a cross-modal knowledge transfer approach that transfers the knowledge from rich information modalities (RIMs), such as electroencephalogram (EEG), to a model established based on Practical Application Modalities (PAMs). The model achieves an impressive accuracy of 99.21% on the DEAP dataset using 6-channel PAMs (tEMG, zEMG, RESP, PLE, TEMP, and GSR) and a 99.99% accuracy on the AMIGOS dataset using 3-channel PAMs (ECG Right, ECG Left, and GSR), outperforming state-of-the-art models. In conclusion, this research investigates into data-driven approaches for proactively and reactively managing stress in two distinct contexts. Leveraging eye movement data, physiological data, and questionnaire data with machine learning or deep learning techniques, the study offers fresh insights into the complex problem of stress. It deepens the understanding of stress effects and development processes, providing potential solutions for diverse settings.