Year-Long Digital Phenotyping and Natural Language Processing of Daily Voice Diaries Reveal Affective and Behavioral Signatures of Real-World Life Stress

Experiencing stressful life events has been associated with poor mental health outcomes, but less is known about daily stressful events’ more proximate impacts on daily behavior and affect. Here we leverage mobile and wearable technology and recent advances in natural language processing for a fine-grained examination of first-year college students’ daily life stress, with a focus on academic and social domains, over a full academic year (8,000+ total daily observations). Experiences of stress were characterized from participants’ daily voice diaries narrating the main events of the day, using a combination of expert human labeling and large language models fine-tuned for sentiment analysis and topic modeling. Bayesian hierarchical models assessed within-person associations between diary-derived instances of academic and social stressful events and same-day sleep, physical activity, social activity, and negative affect measured with actigraphy wristbands and phone surveys. Days with academic stressful events were associated with shorter sleep duration, decreased physical activity, reduced desire to be around others, and modestly increased negative affect. Additionally, days with academic stressful events had significantly reduced social interaction and increased time spent on schoolwork relative to days with social stressful events. Meanwhile, days with social stressful events stood out by especially heightened negative affect, above and beyond the effect of academic stressful events. Our results suggest that academic and social dimensions of life stress may present distinct signatures in daily affect and behavior, with potential implications for long-term wellbeing.