Predicting Depressive Symptom Risk in Arthritis Patients Using the XGBoost Algorithm: A National Cross-sectional Study from CHARLS
| Author | Haibo Shi | en |
| Author | Xinyu Guo | en |
| Author | Yangzong Ciren | en |
| Issued Date | 2026-08-31 | en |
| Abstract | Background: Arthritis imposes substantial physical and psychological burdens on middle-aged and older adults, and depressive symptoms are common in this population. Objectives: This study aimed to develop and validate an interpretable machine learning model to predict depressive symptom risk among middle-aged and older Chinese adults with arthritis and to quantify sex differences in key predictors. Methods: Data were obtained from the 2020 Wave 5 China Health and Retirement Longitudinal Survey (CHARLS), including 6,030 participants with self-reported physician-diagnosed arthritis. Depressive symptoms were assessed using the 10-item Center for Epidemiological Studies Depression Scale (CES-D-10; range, 0 - 30), with a cutoff score of ≥ 10 indicating clinically significant depressive symptoms. Fifteen predictors were identified using LASSO regression and the Boruta algorithm. An XGBoost model was developed in a training set (70%) and evaluated in a held-out test set (30%), with temporal external validation performed using an independent Wave 4 sample (n = 5,918). SHAP (SHapley Additive exPlanations) was used to interpret feature importance and to explore sex-stratified risk profiles. Results: The model demonstrated moderate discriminative performance, with AUC values of 0.752 (95% CI: 0.738 - 0.767) in the training set, 0.687 (95% CI: 0.662 - 0.712) in the test set, and 0.712 (95% CI: 0.699 - 0.725) in the temporal validation set. The Brier scores were 0.202, 0.222, and 0.215, respectively, indicating acceptable calibration with minimal overfitting. Decision curve analysis demonstrated a clinical net benefit within the 0.15 - 0.50 threshold range. SHAP analysis identified ADL limitation (Mean |SHAP| = 0.273), nighttime sleep duration (Mean |SHAP| = 0.260), sex (Mean |SHAP| = 0.218), pain (Mean |SHAP| = 0.194), and the number of chronic diseases (Mean |SHAP| = 0.185) as the five most important predictors. Sex-stratified analyses revealed distinct risk profiles: physical function factors, including ADL limitation and sleep duration, predominated in males, whereas psychosocial factors, including residence type, memory function, marital status, and chronic disease burden, were more influential in females. Conclusions: The XGBoost-SHAP model demonstrated acceptable predictive performance and temporal stability in identifying depressive symptoms among patients with arthritis. Sex-specific risk factor profiles indicate the need for differentiated screening strategies, with an emphasis on functional and sleep assessments in males and psychosocial support in females. | en |
| DOI | https://doi.org/10.5812/healthscope-171757 | en |
| URI | https://brieflands.com/journals/healthscope/articles/171757 | en |
| Keyword | Machine Learning | en |
| Keyword | XGBoost | en |
| Keyword | SHAP | en |
| Keyword | Artificial Intelligence | en |
| Keyword | Arthritis | en |
| Keyword | Depressive Symptoms | en |
| Publisher | Brieflands | en |
| Title | Predicting Depressive Symptom Risk in Arthritis Patients Using the XGBoost Algorithm: A National Cross-sectional Study from CHARLS | en |
| Type | Research Article | en |