Abstract
This study investigates the longitudinal robustness of mental health prediction models using three years of data from the Healthy Minds Study, comprising 265,870 college students surveyed between 2022 and 2025. Multiple statistical techniques were applied to assess data drift, while three Machine Learning (ML) algorithms, namely Logistic Regression (LR), Random Forest (RF), and XGBoost, were evaluated under several temporal modeling strategies. Mental health risk was defined as the presence of moderate-to-severe depression or anxiety symptoms. Although statistically significant distributional changes were observed across variables, effect sizes remained small, indicating limited practical drift. Model performance remained strong over time (mean F1-score = 0.71, mean AUROC = 0.80), with minimal temporal degradation (1.8%). Well-being emerged as the most influential predictor, accounting for the dominant share of feature importance. These findings suggest that mental health prediction models can be reliably deployed when temporal stability is present and highlight the crucial role of well-being in mental health risk prediction.