Detection of Stress in Naturalistic Settings Through Passive Mobile Sensing

Abstract

Unobtrusive stress detection using wearable sensors could enable scalable, continuous mental-health monitoring. However, stress is an inherently subjective state that can only be inferred indirectly from physiological signals, making generalizable detection in naturalistic settings challenging. Although prior work has focused on improving model performance, it remains unclear whether wearable physiology supports a shared cross-individual mapping to subjective stress or whether this relationship is fundamentally person-specific. We evaluated feature-based and deep-learning models across multiple physiological modalities using ecological momentary assessment (EMA) as the reference standard, comparing within- and between-individual modeling approaches. Within-individual models achieved modest but consistent improvements in stress detection, whereas between-individual models consistently failed to generalize, yielding negative R2 values despite multimodal fusion and high-capacity architectures. Error analyses revealed regression to the mean, reduced sensitivity to high-stress states, and residual associations with general physiological activation, highlighting the limited stress specificity of wearable physiology. These findings suggest that wearable stress detection is fundamentally a personalized inference problem and that future systems should prioritize individual adaptation and contextual modeling over universal stress predictors.