Development and validation of an interpretable risk prediction model for depressive symptoms among community-based patients with chronic obstructive pulmonary disease using machine learning techniques

Chronic obstructive pulmonary disease (COPD) is frequently accompanied by depressive symptoms, which are often under-recognized in community settings. Although machine-learning-based prediction models for depression risk in COPD have recently emerged, evidence from community-dwelling Chinese COPD populations remains limited. This study aimed to compare commonly used supervised machine learning algorithms and to develop an interpretable, parsimonious model for identifying community-based COPD patients at high risk of depressive symptoms. We analyzed data from 809 participants aged > = 40 years with self-reported physician-diagnosed COPD/chronic lung disease in the 2020 China Health and Retirement Longitudinal Study (CHARLS). Depressive symptoms were defined using the 10-item Center for Epidemiologic Studies Depression Scale (CES-D-10), with a total score > = 10 indicating depressive symptoms. COPD status was based on self-reported physician diagnosis rather than spirometry. Missing values were handled using multiple imputation by chained equations. Least absolute shrinkage and selection operator (LASSO) regression and multivariable logistic regression were used for feature selection. Six models, including logistic regression (LR), multilayer perceptron, K-nearest neighbors, random forest (RF), XGBoost, and LightGBM, were compared using 10-fold cross-validation within the training set and an independent test set. RF impurity-based and permutation feature-importance analyses were added, and LR and RF were further compared across 50 repeated random training-test splits. Results: Six predictors were retained: Self_assessed_health, Sleeping_time_at_night, Memory_disease, Gender, ADL_score, and Age. LR showed the most stable generalization performance, with an independent test-set AUC of 0.713 (95% CI: 0.648–0.778). Across 50 repeated stratified random splits, LR achieved higher average sensitivity (0.698 ± 0.060) and F1 score (0.706 ± 0.035) than RF (sensitivity, 0.641 ± 0.036; F1 score, 0.671 ± 0.021), while specificity was comparable between LR and RF (0.606 ± 0.061 vs. 0.616 ± 0.046). In this CHARLS-based community sample, a simple LR model showed stable and interpretable performance for predicting depressive symptoms among self-reported COPD patients. The model should be regarded as a screening-support tool rather than a diagnostic instrument, and further external validation with spirometry-confirmed COPD cohorts is warranted.