Associations between air pollution and depression risk in older adults: An improved interpretable prediction approach integrating CNN-BiLSTM-Attention with grouped SHAP

Although air pollution is considered a potential risk factor for depression in older adults, its underlying associations and predictive patterns remain unclear. Existing studies often rely on linear models or basic machine learning methods that neglect temporal dynamics and individual heterogeneity, resulting in suboptimal predictive accuracy and limited interpretability. This study employed a Convolutional Neural Network-Bidirectional Long Short-Term Memory-Attention (CNN-BiLSTM-Attention) model combined with a grouped SHapley Additive exPlanations (SHAP) method to develop a predictive risk stratification tool for depression risk in older adults associated with air pollution exposure. The analysis was based on data from the China Health and Retirement Longitudinal Study (CHARLS). The SHAP method accounted for individual variables and quantified the model-based feature importance of air pollution variables, comparing them with eight other submodels to assess model performance and predictive contributions. The findings showed that the CNN-BiLSTM-Attention model outperformed the other eight submodels in prediction performance in population screening settings, with an area under the curve (AUC) of 0.6968 (p < 0.05), a sensitivity of 0.7986, and an F1 score of 0.6425. Accounting for individual variables, air pollution variables were generally positively correlated with depression risk, with sulfur dioxide (SO 2 ), fine particulate matter (PM 2.5 ), and black carbon (BC) showing the greatest contributions. This relationship was consistent across all nine models. This study integrates deep learning and interpretability techniques to characterize the robust associations between air pollution and depression risk in older adults. The findings provide evidence to inform public health policy and guide future environmental mental health research, particularly in longitudinal analyses of pollutant interactions. Not applicable.