Stress Classification From Wearable Multimodal Physiological Signals: Deep Learning Approach

Abstract

Stress is a common problem with serious consequences for health and well-being, especially among healthcare workers such as nurses. Chronic stress in this population can lead to burnout, affecting both their personal well-being and patient care. Traditional stress assessment methods are limited in their ability to provide continuous, objective monitoring, highlighting the need for innovative solutions. Wearable technology, such as the `Empatica E4` smartwatch, offers a promising avenue for real-time stress monitoring through the measurement of physiological signals. In our study, we developed a stress classification model utilizing data collected from the `Empatica E4` smartwatch worn by nurses in a hospital setting. The model uses physiological signals to effectively classify stress levels using Long Short-Term Memory (LSTM) neural networks. The study begins by exploring the physiological and psychological aspects of stress, emphasizing the challenges faced by nurses in high-stress environments and the potential of wearable technology for stress monitoring. Utilizing a publicly available dataset consisting of physiological signals, including heart rate, electrodermal activity (EDA), skin temperature, and the physical activity of nurses, and preprocessing it for analysis. The results demonstrate the model's effectiveness in both binary and multiclass classification tasks, with notable performance differences observed across stress categories. In multiclass classification, the model exhibits high precision and recall for "no stress" (precision: 0.88, recall: 0.90) and "high stress" (precision: 0.94, recall: 0.94) categories but shows reduced performance for "low stress" (precision: 0.74, recall: 0.71). This gap could be related to signal ambiguity and class imbalance. In binary classification, merging "low stress" with "no stress" simplifies the task and results in consistent performance across both classes, with stress detection achieving precision of 0.91 and recall of 0.96 and "no stress" classification precision of 0.90 and recall of 0.82. This research not only advances our understanding of stress monitoring in the healthcare environment but also furthers the exploration of stress diagnosis using smartwatch data. The developed stress classification model offers potential applications in real-time stress monitoring systems, personalized feedback mechanisms, and intervention strategies to mitigate stress among healthcare professionals. These findings have broader implications for improving well-being and patient care quality in high-stress environments beyond healthcare