Assessment of cumulative inhalation risks of PAHs and metals in an industrial city using machine learning–based concentration reconstruction

Abstract

The evaluation of long-term cumulative exposure to hazardous air pollutants (HAPs) in industrial cities is constrained by the insufficient spatiotemporal coverage of routine monitoring networks. To address this limitation, a Random Forest model was developed to reconstruct historical concentrations of 13 polycyclic aromatic hydrocarbons (PAHs) and eight metals in Ulsan, South Korea. The model was trained using national HAPs monitoring data for PAHs and metals during 2013–2023, together with criteria air pollutants (CAPs) and meteorological variables, and subsequently applied to generate reconstructed concentrations for 2001–2023. Concentration distributions were reconstructed across the air quality monitoring network, providing denser spatial coverage than HAP monitoring stations. The reliability of the model was evaluated using routine national monitoring data employed for model training and further assessed using independent passive and active air sampling conducted in 2024, confirming that temporal trends and spatial contrasts were accurately captured. Major pollution sources were identified through Positive Matrix Factorization (PMF), indicating traffic emissions and the influence of petrochemical, non-ferrous metal, and heavy manufacturing industrial complexes, with distinct spatial hotspots. Cumulative exposure estimates derived from the reconstructed time series were compared with those calculated using a conventional fixed-year approach. Cumulative PAH exposure in urban residential areas was found to be underestimated by the conventional method, as historically elevated PAH concentrations were not accounted for. Conversely, cumulative metal exposure near industrial complexes was overestimated because recent short-term concentration variability was applied across the entire exposure duration. In addition, residential scenario analyses demonstrated that residential relocation can be incorporated into cumulative exposure assessment. These findings indicate that cumulative risk assessment is more reliably supported by machine learning–based reconstruction than by static concentration assumptions, offering practical insights for long-term public health management in cities with limited data.