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

Area of Science:

• Digital health
• Machine learning applications
• Mental health technology

Background:

• Mental health challenges are a global concern, with traditional assessments lacking objectivity and real-world applicability.
• Passive sensing using wearables and smartphones offers a noninvasive method for continuous mental health monitoring.
• Machine learning (ML) enhances passive sensing by enabling objective data analysis for mental health assessment.

Purpose of the Study:

• To comprehensively review passive sensing and ML technologies for mental health monitoring.
• To summarize technical approaches and identify associations between behavioral features and mental disorders.
• To explore future research directions in this domain.

Main Methods:

• A scoping review following PRISMA-ScR guidelines, searching 7 databases for studies from January 2015 to February 2025.
• Included 42 peer-reviewed studies utilizing passive sensing (wearables/smartphones) and ML for monitoring diagnosed mental disorders like depression and anxiety.
• Synthesized data on technical aspects (data collection, ML models) and clinical associations, categorizing behavioral features into 8 domains.

Main Results:

• Most studies (55%) focused on depression, primarily using wrist-worn devices to collect heart rate, movement, and step count data.
• Deep learning models (e.g., CNN, LSTM) achieved high accuracy, while traditional ML models (e.g., random forest) offered better interpretability.
• Significant limitations include small sample sizes (76% <100 participants), short monitoring durations (45% <7 days), and scarce external validation (2%).

Conclusions:

• Passive sensing and ML show high accuracy for detecting conditions like anxiety (92.16% with CNN-LSTM), but methodological heterogeneity and small sample sizes limit current evidence.
• Key limitations include inconsistent study designs (76% single-device, 45% <7-day monitoring), high risk of bias due to small samples and lack of validation, and ethical concerns regarding data anonymization (14%).
• Clinical translation necessitates standardized protocols, larger longitudinal studies (≥3 months), robust ethical frameworks, multimodal sensor fusion, and explainable AI to ensure reliable and deployable mental health solutions.