Smartphone-Based Digital Phenotyping for Identifying Elevated Depressive Symptom Levels Using Machine Learning

Depressive symptoms among university students are a growing public health concern, motivating unobtrusive monitoring with smartphone-based digital phenotyping. This study examined whether passively collected behavioral features can identify elevated depressive symptom risk. We collected smartphone sensing data from 36 university students over two weeks, yielding 551 participant-days. We extracted 21 digital phenotyping features, including 14 additional behavioral features and 7 baseline features. Multiple machine learning models were evaluated using repeated bootstrap validation, and participant-independent generalization was further assessed with Leave-One-Subject-Out (LOSO) validation. We also examined PHQ-9 thresholds, class-imbalance mitigation, and feature importance using SHAP. Tree-based ensemble models achieved the best performance under bootstrap validation, and the proposed feature set consistently outperformed the baseline set. However, performance dropped substantially under LOSO validation, indicating that participant-independent generalization remains challenging. Class weighting and SMOTE did not meaningfully improve performance. SHAP analysis identified home-stay behavior and time-of-day smartphone use as the most influential predictors. These findings highlight that participant-independent prediction remains challenging and underscore the importance of rigorous participant-level validation when developing smartphone-based digital phenotyping models.