Benchmarking Classical Machine Learning Models for College Student Mental Health Screening Using Passive Sensing Data

Abstract

Mental health risk screening among college students can be supported by passive smartphone sensing, but model selection remains difficult because these data are heterogeneous, incomplete, and strongly imbalanced. This study establishes a reproducible benchmark of five classical single-model machine learning algorithms—Logistic Regression, Support Vector Machine, k-Nearest Neighbors, Naive Bayes, and Decision Tree—using the publicly available StudentLife dataset. The final analysis included 45 participants and five daily behavioral features, with depression risk defined as PHQ-9 ≥ 10. A uniform preprocessing pipeline was applied within training folds, and 5-fold GroupKFold cross-validation was used to prevent cross-subject information leakage. Performance was measured using AUROC, AUPRC, F1-score, precision, and recall, with AUPRC and recall emphasized because the positive-risk class represented approximately 11.1% of participants. Robustness was additionally examined by simulating 10% and 30% missing feature values. Decision Tree achieved the highest AUROC (0.564 ± 0.144), F1-score (0.210 ± 0.087), and recall (0.480 ± 0.179), while Decision Tree and SVM obtained the same mean AUPRC (0.164). The overlapping variability and modest absolute performance indicate that these results should be interpreted as preliminary benchmark evidence rather than definitive model superiority. The benchmark provides a transparent baseline for future work on temporal modeling, explainability, and privacy-aware student mental health screening.