Detecting Prolonged Stress in Real Life Using Wearable Biosensors and Ecological Momentary Assessments: Naturalistic Experimental Study

Increasing efforts toward the prevention of stress-related mental disorders have created a need for unobtrusive real-life monitoring of stress-related symptoms. Wearable devices have emerged as a possible solution to aid in this process, but their use in real-life stress detection has not been systematically investigated. We aimed to determine the utility of ecological momentary assessments (EMA) and physiological arousal measured through wearable devices in detecting ecologically relevant stress states. Using EMA combined with wearable biosensors for ecological physiological assessments (EPA), we investigated the impact of an ecological stressor (ie, a high-stakes examination week) on physiological arousal and affect compared to a control week without examinations in first-year medical and biomedical science students (51/83, 61.4% female). We first used generalized linear mixed-effects models with maximal fitting approaches to investigate the impact of examination periods on subjective stress exposure, mood, and physiological arousal. We then used machine learning models to investigate whether we could use EMA, wearable biosensors, or the combination of both to classify momentary data (ie, beeps) as belonging to examination or control weeks. We tested both individualized models using a leave-one-beep-out approach and group-based models using a leave-one-subject-out approach. During stressful high-stakes examination (versus control) weeks, participants reported increased negative affect and decreased positive affect. Intriguingly, physiological arousal decreased on average during the examination week. Time-resolved analyses revealed peaks in physiological arousal associated with both momentary self-reported stress exposure and self-reported positive affect. Mediation models revealed that the decreased physiological arousal in the examination week was mediated by lower positive affect during the same period. We then used machine learning to show that while individualized EMA outperformed EPA in its ability to classify beeps as originating from examinations or from control weeks (1603/4793, 33.45% and 1648/4565, 36.11% error rates, respectively), a combination of EMA and EPA yields optimal classification (1363/4565, 29.87% error rate). Finally, when comparing individualized models to group-based models, we found that the individualized models significantly outperformed the group-based models across all 3 inputs (EMA, EPA, and the combination). This study und