Explanatory Machine Learning Model to Associate All-Cause Mortality with PM 2.5 by Sources and Components: A Nationwide Difference-in-Differences Study

Background: Long-term PM2.5 exposure is a major health risk, yet the heterogeneity of its chemical components and sources is often overlooked in risk assessments. Methods: Using county-level census data for 1990, 2000, and 2010 in China, we employed a difference-in-differences (DID) study design to associate all-cause mortality with long-term exposure to PM2.5 from five sources and by five components. We applied log-linear models and several mixed exposure models (weighted quantile sum (WQS), ridge regression, random forest, XGBoost, and an ensemble model) to capture effect heterogeneity with cross-validation for evaluating the prediction performance and Shapley additive explanations (SHAP) values for interpretability. Results: Each 10 μg/m³ increment in PM2.5 was associated with a 3.16% increase in all-cause mortality (95% confidence interval [95% CI]: 2.13-4.20%). Industry (3.46%, for each interquartile range increment), power (2.32%), and residential usage (1.46%) were the most hazardous sources; ammonium (6.46%), nitrate (4.23%), and black carbon (3.93%) were the most potent components. Random forest and the ensemble model showed superior predictive performance, identifying industrial-sourced black carbon as the leading contributor to PM2.5-associated mortality. Conclusion: The health effects of PM2.5 differ considerably by the source sector and chemical component. Industrial black carbon exhibits higher toxicity and should be given priority when designing the air-pollution-control strategies.