What does this research mean for the field?

A hybrid CNN-Transformer architecture significantly outperforms traditional machine-learning models in detecting stress and attention from wearable PPG and EEG data in noisy, real-world workplace environments, achieving 76.2% and 87.4% accuracy respectively. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This research aims to bridge the gap between laboratory-based detection of stress and attention and real-world applications using wearable technology.

April 19, 2026Open Access

Wearable Stress and Attention Detection: Bridging Lab Methods and Real-World Workplace Monitoring

Key Points

This research aims to bridge the gap between laboratory-based detection of stress and attention and real-world applications using wearable technology.
Developed a mobile platform for remote acquisition of PPG and EEG data.
Monitored data collection continuously during participants' office work over two weeks.
Enabled self-labeling of stress states through a mobile interface.
Applied a hybrid CNN–Transformer architecture for data analysis.
Achieved 76.2% accuracy for stress detection and 87.4% for six-level attention classification in real-world conditions.
Outperformed classical models such as SVM by at least 10% for comparable tasks.
Demonstrated the significance of temporal self-attention in noisy data environments.

Abstract

Compact, user-friendly wearable sensors enable unobtrusive, real-time monitoring of mental health and cognitive performance, supporting quality of life management, early detection and prevention of cognitive and mental health risks. Among wearable sensors, photoplethysmography (PPG) and electroencephalography (EEG) provide access to autonomic and cortical processes underlying stress and attention, reflecting cardiovascular autonomic regulation and neural dynamics, respectively. They have demonstrated reliable performance under controlled laboratory conditions; however, their deployment in real-world settings remains challenging due to increased noise, motion artifacts, limited sensor placement, and environmental interference. To address the performance gap between laboratory-based stress and attention detection and real-world deployment, we developed a custom mobile platform enabling fully remote acquisition of wearable PPG and EEG data from 25 participants over two weeks of everyday office work, without experimental control or direct intervention. The platform supported continuous acquisition of wearable PPG and EEG signals, remote monitoring of the data-collection process to detect recording failures or user errors, and participant-driven self-labeling of stress states through a mobile interface. The platform supported an integrated pipeline from data acquisition to analysis under real-world conditions. Here we show that a hybrid CNN–Transformer architecture significantly outperforms traditional machine-learning models in noisy, naturalistic environments. Our results demonstrate that while classical models, such as support vector machines (SVMs), achieve up to 96% accuracy in the laboratory, they fail to generalize to the workplace. In contrast, our proposed deep learning approach achieves 76.2% accuracy for stress versus non-stress detection and 87.4% for six-level attention classification, outperforming prior reported methods by at least 10% under comparable real-world conditions. These findings highlight the importance of temporal self-attention for modeling stress- and attention-related states in noisy, real-world data, beyond methods designed for clean laboratory signals. This work bridges laboratory studies and real-world deployment by presenting a wearable sensing and analysis pipeline for occupational health that identifies stress- and attention-related patterns using noninvasive, consumer-grade hardware, and supports early prevention of chronic stress and improved well-being in everyday work settings.

Bookmark

View Full Paper