Explainable machine learning classification of 7-year longitudinal depressive symptom trajectories: evidence from CHARLS

Late-life depression exhibits heterogeneous symptom trajectories, yet the early-life experiences of persistent high-symptom groups remain poorly understood. This study examines the contributions of adverse childhood experiences (ACEs) and current health factors to the classification of longitudinal depressive symptom trajectories in Chinese middle-aged and older adults. Using 7-year longitudinal data from the China Health and Retirement Longitudinal Study (CHARLS, N = 9,427), we identified depressive symptom trajectories over four waves using latent growth curve modeling (LGCM) and latent class growth analysis (LCGA). A consensus feature selection strategy (combining LASSO, Elastic Net, and Boruta) screened 37 candidate predictors, including ACEs, parenting behaviors, cognition and self-reported health status at baseline. We trained four classifiers (XGBoost, Random Forest, SVM, and Logistic Regression) and interpreted model classifications using Shapley Additive Explanations (SHAP). We identified two distinct trajectories: “low-stable trajectory” (73.35%) and “persistent high-symptom trajectory” (26.65%). Seventeen features were used to train four machine learning models. The four classifiers showed broadly comparable discrimination, with the RF model achieving the numerically highest AUC of 0.783 (95% CI: 0.760–0.807) and an accuracy of 0.733 (95% CI: 0.712–0.752). SHAP analysis identified trouble pain, gender, sleep duration, self-reported memory, self-reported health, and household mental illness as the leading contributors to trajectory classification. Longitudinal depressive symptom trajectories were associated with baseline characteristics across current health, sociodemographic, family, and childhood factors. Interpretable machine learning may provide a useful approach for exploratory risk stratification, but external validation is needed before clinical application. Not applicable.