Machine learning-enhanced prediction of size-resolved gas-particle partitioning quotient: Implication for health risk assessment of polycyclic aromatic hydrocarbons

De-Qi Wang, Wan-Li Ma, Ru-Peng Wang, Pu-Fei Yang, Peng-Tuan Hu, Shi-Ming Jia, Meng Qin,
Machine learning-enhanced prediction of size-resolved gas-particle partitioning quotient: Implication for health risk assessment of polycyclic aromatic hydrocarbons,
Journal of Hazardous Materials,
Volume 499,
2025,
140047,
ISSN 0304-3894,
https://doi.org/10.1016/j.jhazmat.2025.140047.
(https://www.sciencedirect.com/science/article/pii/S0304389425029668)
Abstract: Gas-particle partitioning plays an important role in environmental fate and health risk assessment of polycyclic aromatic hydrocarbons (PAHs), especially for multi-size particles. In this study, machine learning models were applied for predicting size-resolved gas-particle partitioning quotient (KPi) and health risk of PAHs. It was found that the XGBoost model performed better than the other three machine learning models for KPi prediction. The SHapley Additive exPlanations indicated that octanol-air partition coefficients of PAHs, particle sizes and ambient temperature were key impact factors of KPi. Then, the XGBoost model of KPi was successfully integrated with the International Commission on Radiological Protection (ICRP) model for directly predicting benzo[a]pyrene equivalency (BaPeq) of particulate PAHs. The integrated model showed good predictive performance, with predicted BaPeq values reaching over 90 % of the measured values. The integrated model was further employed to estimate BaPeq reduction rates of particulate PAHs under different controlling strategies. The result emphasized the importance of controlling PM1 for reducing BaPeq of particulate PAHs in the alveolar region. With the successful integration of the machine learning model and the ICRP model, this study provided new insights for health risk assessment of size-resolved particulate PAHs.
Keywords: PAHs; Size-resolved gas-particle partitioning; Benzo[a]pyrene equivalency; Machine learning; Model integration