Model card
Tuned gradient-boosted trees (XGBoost) estimating ~4-year risk of clinically relevant depressive symptoms (PHQ-9 ≥ 10).
Data
- The Maastricht Study; baseline 2010–2020; ages 40–75.
- Analytic cohort: 6,004 participants; 880 incident cases.
- Outcome: first PHQ-9 ≥ 10 over ~4 years (annual assessments).
Features
- Sleep minutes (weekly average; spline transform)
- Sleep fragmentation (quartile Q1–Q4; ≥1-min posture-break count)
- Age, sex
- Advanced flags: prior depression / antidepressant use (
any_vuln), neuropathy
Model & training
- XGBoost, Optuna-tuned;
tree_method="hist" - One-hot encoding; train n = 11,153; test n = 2,789
Performance (hold-out)
AUROC 0.71 · AUPRC 0.29.
Limitations & ethics
Research prototype; not a medical device. Trained on ages 40–75; no external validation yet. Results should not be used for diagnosis or treatment.