Machine learning analysis of drivers of differences in PAH content between PM1 and PM10 in Zagreb, Croatia

Abstract

Particulate matter (PM) is a critical component of air pollution, with its size fractions, PM₁ and PM₁₀, posing varying risks to human health. Among its constituents, polycyclic aromatic hydrocarbons (PAHs) are of particular concern due to their toxicity and ability to affect respiratory health. PM₁ particles, being smaller, can penetrate deeper into the respiratory system, making the distribution of PAHs across PM₁ and PM₁₀ especially relevant. This study examines the seasonal patterns and factors influencing PAH concentrations in PM₁ and PM₁₀ collected in Zagreb, Croatia. A combination of machine learning (ML) techniques, including Random Forest (RF) regression and Principal Component Analysis (PCA), and statistical approaches like t-tests and Cohen's d, were applied to explore these relationships. Post-hoc interpretation using SHapley Additive exPlanations (SHAP) clarified the contribution of various predictors in the models. Results indicated that PAH concentrations contributions were higher in PM₁ than PM₁₀, posing greater health risks associated with finer particles. Seasonal trends showed increased PAH levels during winter and spring, primarily driven by heating activities and temperature inversions. The study also highlighted the "J_Curve_Day" variable as the most critical predictor in RF regression models, capturing the influence of seasonal changes on PAH levels through its representation of meteorological conditions and atmospheric processes. These insights shows the importance of understanding seasonal variability in PAH distributions within PM to inform air quality management and public health strategies.