Comparative Evaluation of Machine Learning Models for Residential PM₁ Prediction in Zagreb (Croatia): Identifying Key...

Indoor exposure to particulate matter (PM) is increasingly recognized as a major contributor to respiratory and cardiovascular risk, yet the relative contributions of outdoor pollution, building characteristics, and occupant behavior remain poorly resolved. PM₁ (aerodynamic diameter < 1 μm) warrants focus due to its higher alveolar deposition. "Evidence driven indoor air quality improvement" (EDIAQI) project aims to enhance indoor air quality guidelines and increase awareness by providing accessible data on exposure, pollution sources, and related risk factors. As part of the Zagreb pilot within the project, 103 paired indoor/outdoor PM₁ samples were analyzed. Seasonal analysis revealed substantial wintertime outdoor PM₁ spikes, while indoor medians remained stable. Chemometric analysis identified factors such as dwelling size, outdoor pollution, resuspension, building age/heating type, and urban context. Among the tested models, the validated gradient-boosted regressor (GBR) achieved the strongest performance, explaining ~65% variance in indoor PM₁ (test R² ≈ 0.65). Explainable machine learning analysis (SHAP) identified outdoor PM₁ levels, infiltration, and resuspension as the most influential predictors. Findings underscore wintertime outdoor emissions (e.g., residential heating and traffic) and dwelling-related and behavioral factors as key drivers, with the machine learning-environmental data integration enabling targeted residential IAQ management: optimized ventilation protocols, resuspension mitigation via behavior, and infiltration reduction through retrofits.