Passive Sensing for Mental Health Monitoring Using Machine Learning With Wearables and Smartphones: Scoping Review

BACKGROUND: Mental health issues have become a significant global public health challenge. Traditional assessments rely on subjective methods with limited ecological validity. Passive sensing via wearable devices and smartphones, combined with machine learning (ML), enables objective, continuous, and noninvasive mental health monitoring. OBJECTIVE: This study aimed to provide a comprehensive review of the current state of passive sensing-based and ML technologies for mental health monitoring. We summarized the technical approaches, revealed the association patterns between behavioral features and mental disorders, and explored potential directions for future advancements. METHODS: This scoping review adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines and was prospectively registered on the Open Science Framework. We systematically searched 7 databases (Web of Science, PubMed, IEEE Xplore, Embase, PsycINFO, Scopus, and ACM Digital Library) for studies published between January 2015 and February 2025. We included 42 peer-reviewed studies that used passive sensing from wearables or smartphones with ML to monitor clinically diagnosed mental disorders, such as depression and anxiety. Data were synthesized across technical dimensions (data collection, preprocessing, feature engineering, and ML models) and clinical associations, with behavioral features categorized into 8 domains. RESULTS: The 42 included studies were predominantly cohort designs (23/42, 55%), with a median sample size of 60.5 (IQR 54-99). Most studies focused on depression (23/42, 55%) and anxiety (9/42, 21%) using primarily wrist-worn devices (32/42, 76%) collecting heart rate (28/42, 67%), movement index (25/42, 60%), and step count (17/42, 40%) as key biomarkers. Deep learning models (eg, convolutional neural networks and long short-term memory) showed high accuracy, while traditional ML (eg, random forest) remained prevalent due to better interpretability. We identified critical limitations, including small samples (32/42, 76% with N<100), short monitoring periods (19/42, 45% <7 days), scarce external validation (1/42, 2%), and limited reporting on data anonymization (6/42, 14%). CONCLUSIONS: While passive sensing and ML demonstrate promising accuracy (eg, convolutional neural network-long short-term memory achieving 92.16% in anxiety detection), the evidence remains constrained by three key limitations: (1) methodological heterogeneity (32/42, 76% single-device studies; 19/42, 45% with <7-day monitoring), (2) high risk of bias from small samples (median 60.5, IQR 54-99 participants) and scarce external validation (1/42, 2%), and (3) ethical gaps (only 6/42, 14% addressing anonymization). These findings underscore the technology's potential to transform mental health care through objective, continuous monitoring-particularly for depression (heart rate and step count biomarkers) and anxiety (sleep and social interaction patterns). However, clinical translation requires standardized protocols, larger longitudinal studies (≥3 months), and ethical frameworks for data privacy. Future work should prioritize multimodal sensor fusion and explainable artificial intelligence to bridge the gap between technical performance and clinical deployability.