Smartphone-based digital phenotyping for detection of high-risk depression and anxiety in Korean community settings

Smartphones generate continuous behavioral signals such as mobility and activity patterns, offering scalable opportunities for monitoring mental health in community settings. Digital phenotyping approaches that integrate passive sensing with brief self-report measures may enable early identification of individuals at high risk for depression and anxiety without reliance on additional wearable devices. We prospectively evaluated a smartphone-based digital phenotyping framework in 455 community-dwelling adults in Korea who contributed 28 days of passive Global Positioning System and accelerometer data, daily self-report microsurveys, and weekly PHQ-9/GAD-7 assessments for screening high-risk depression and anxiety. Machine learning models were compared across active-only, passive-only, and combined feature sets. After applying predefined coverage criteria (≥60% passive-data coverage and ≥ 60% corresponding active-data availability), 277 participants were included in the depression cohort and 275 in the anxiety cohort. Passive features capturing mobility, activity regularity, and sleep-related behaviors were derived, and machine learning models were trained using active-only, passive-only, and combined feature sets. For depression, combined models achieved the best performance, with AUCs ranging from 0.77 to 0.83 and APs ranging from 0.86 to 0.91 across classifiers. Similar patterns were observed for anxiety, with AUCs up to 0.86 and APs up to 0.95. Ablation analyses identified robust deployment conditions relevant to clinical screening, including tolerance to missing data and short look-back windows. These findings support the practical utility of smartphone-based digital phenotyping pipelines that integrate passive behavioral signals with brief self-reports for scalable screening for high-risk depression and anxiety in real-world environments, and they may inform future just-in-time mental health intervention systems. • Smartphone-based digital phenotyping enables screening for high-risk depression and anxiety in community settings. • Passive GPS and accelerometer data fused with brief self-reports improved prediction accuracy. • Combined models outperformed single-modality approaches for both depression and anxiety screening. • Robust performance was maintained under realistic levels of missing smartphone data. • Smartphone-only frameworks support symptom monitoring in community and outpatient settings without wearables.