An explainable machine learning model for predicting depressive symptoms among Chinese older adults with chronic diseases.

backgroundDepression is a multifactorial disorder whose coexistence with chronic diseases imposes a substantial health burden and complicates the identification of at-risk older adults. To address this, we analyzed seven-dimensional data derived from the China Health and Retirement Longitudinal Study (CHARLS) and applied machine learning (ML) to predict depressive symptoms among older adults with chronic diseases.

methodsParticipants from wave 4 of CHARLS were included. Depressive symptoms were assessed using the 10-item Center for Epidemiologic Studies Depression Scale (CESD-10), with a total score > 12 indicating clinically significant depressive symptoms. Five ML models were trained using 70% of the dataset and validated on the remaining 30%. Hyperparameters were optimized using grid search combined with five-fold cross-validation. Feature importance was calculated, ranked, and visualized using SHapley Additive exPlanations (SHAP).

resultsTotally, 5,564 participants from CHARLS Wave 4 who met the eligibility criteria were enrolled in the analysis, of whom 1,862 exhibited depressive symptoms. The prevalence of depressive symptoms in this cohort was approximately one-third to one-half of that reported in populations with other chronic diseases. Among the 55 features, 17 potential predictors were selected. The eXtreme Gradient Boosting (XGBoost) model outperformed the others, achieving the highest area under the curve (AUC, 0.756) and lowest binary Brier score (0.180). This model also obtained an accuracy of 0.736, precision of 0.657, F1 score of 0.517, and specificity of 0.890. Based on SHAP values, nighttime sleep duration was the most influential of the 17 features in predicting depressive symptoms, followed by bodily pain, life satisfaction, self-reported health, and gender. The optimal XGBoost model, validated via five-fold cross-validation, was ultimately deployed as a web application for predicting depressive symptoms.

conclusionsUsing multidimensional data, the XGBoost, emerged as the top-performing model for predicting depressive symptoms among elderly individuals with chronic conditions, and this model was translated into an open-access, user-friendly interactive application. This work establishes a viable technical framework for precision public health strategies to enhance mental health in older adults with chronic comorbidities, enabling large-scale, low-cost population-level proactive health management. CLINICAL TRIAL NUMBER: Not applicable.