Machine Learning for Depression Screening and Intervention: An Original Circadian Rhythm Score-based Methodology

This paper proposes a Circadian Rhythm Score (CRS) that compresses sleep, activity, and social-behavior indicators into a unified, nonnegative representation for depression screening. Using CHARLS data from 15,233 participants, the authors combine gradient-boosted trees with SHAP to model nonlinear and saturation-like associations. The reported ROC-AUC is 0.825. Interaction modeling and counterfactual regression are then used to estimate heterogeneous, dose-dependent behavioral effects, including an exercise threshold of approximately 300 MET-min/week and an approximately 65-minute restorative nap duration for sleep-deprived individuals. These intervention-oriented estimates should not yet be interpreted as causal clinical recommendations.