Association between long-term air quality index exposure and multidimensional study-defined cognitive impairment in middle-aged and older adults

Cognitive impairment is a major public health challenge in aging populations. Air pollution has been associated with adverse central nervous system outcomes, and PM₁ has attracted increasing attention because its small particle size may facilitate biologically relevant exposure pathways. However, epidemiological evidence concerning PM₁ and multidimensional cognitive outcomes remains limited. The within-model attribution of PM₁ relative to other pollutant inputs also remains incompletely characterized. We conducted a pooled, multi-wave, retrospective observational analysis of 29,468 unique CHARLS participants aged ≥ 45 years. Each participant contributed one analytical record. City-level AQI and seven pollutant metrics summarized over 2015–2020 were linked to the cognitive assessment retained for each participant. Descriptive statistics, multivariable logistic regression, and restricted cubic spline analysis examined associations with three study-defined cognitive outcomes. Six machine-learning algorithms were evaluated using five-fold cross-validation. Their AUCs represent geographically non-independent internal discrimination, and SHAP analyses describe within-model feature attribution. After full adjustment for the measured covariates, each 1-unit increase in AQI corresponded to odds ratios of 1.002 (95% CI, 1.001–1.003) for episodic memory impairment, 1.007 (95% CI, 1.005–1.008) for mental status impairment, and 1.009 (95% CI, 1.007–1.011) for comprehensive cognitive impairment (all P < 0.001). Restricted cubic spline models indicated statistically significant nonlinear patterns, although curve shapes differed across outcomes and adjustment sets and uncertainty increased at the exposure extremes. The machine-learning models showed moderate discrimination. Across the pollutant inputs, PM₁ had the largest mean absolute SHAP value, while age and education generally had greater importance in the full predictor set. Long-term AQI was associated with study-defined multidimensional cognitive impairment among middle-aged and older Chinese adults. The prediction results describe internal discrimination under geographically non-independent individual-level five-fold cross-validation. Among pollutant inputs, PM₁ had the largest mean absolute SHAP value. This within-model ranking does not establish temporal precedence, a causal effect, or incremental predictive value beyond non-pollutant predictors.