Evaluating Cross-Subject Generalization in Multimodal Wearable Stress Sensing for Hospital Nurses

Abstract

Healthcare professionals, particularly nurses, experience sustained occupational stress that can negatively affect both their well-being and patient safety. Wearable devices can continuously monitor physiological indicators associated with stress in real-world clinical environments. Using a publicly available multimodal dataset collected from 15 nurses via wrist-worn sensors during the COVID-19 pandemic, this study compares subject-independent and subject-personalized machine learning models for stress detection. Physiological time-series signals were processed using non-overlapping window segmentation, resulting in a 30-dimensional feature space. Three ensemble-based algorithms—Random Forest, XGBoost, and LightGBM—were evaluated under a Leave-One-Subject-Out (LOSO) framework, while personalized modeling was evaluated separately for each subject. The results reveal a substantial personalization gap in wearable stress sensing. Subject-independent models achieved mean balanced accuracies between 0.529 and 0.538, whereas personalized models achieved a mean balanced accuracy of 0.903 and a mean ROC-AUC of 0.989. Statistical analysis confirmed that the improvement was highly significant (p = 1.49 × 10⁻8, Cohen's d = 4.03). The findings indicate that inter-individual physiological variability substantially limits cross-subject generalization, and that subject-specific adaptation may be important for reliable wearable stress monitoring.