Identifying mortality-related PM2.5 components and high-risk regions using machine learning

Fine particulate matter (PM 2.5 ) is a major air pollutant associated with increased risks of respiratory and cardiovascular diseases. Because the health impacts of PM 2.5 may vary depending on its chemical components, this study applied machine learning to identify key PM 2.5 components linked to mortality and regions with high threshold exceedance frequency. PM 2.5 chemical components, air pollutant concentrations, meteorological variables, and mortality data for the Seoul Metropolitan Area were collected from national monitoring networks and public databases. Four machine learning models, including Random Forest, eXtreme Gradient Boosting, Light Gradient-Boosting Machine, and Categorical Boosting, were applied to predict the age-standardized mortality rate. SHapley Additive exPlanations (SHAP) analysis was then used to identify key components and derive concentration thresholds. Among the models, Categorical Boosting showed the best performance, achieving the highest predictive accuracy under the 7-day moving average condition (R 2 = 0.943). Feature importance and SHAP analyses identified nitrate (NO 3 - ) and arsenic (As) as key components associated with mortality prediction. Threshold-based spatial analysis showed relatively high exceedance ratios for NO 3 - in Ansan (50.3%) and for As in Pocheon (47.8%), with both components also showing elevated ratios in Seoul and Ansan. These findings provide a practical framework for component-based air quality management and targeted mitigation strategies.