Using an Interpretable Machine Learning Framework for Managing Fragmented and Unbalanced Mobile Monitoring Datasets in Spatiotemporal Modeling

Particulate black carbon (BC) is an important air pollutant linked to adverse health effects in susceptible populations. This study developed a spatiotemporal exposure model for BC for a pregnancy cohort in Beijing, China, addressing the challenge of a temporally and spatially fragmented and unbalanced dataset obtained from a mobile monitoring campaign. Ambient particulate BC was measured using a mobile vehicle platform from 2023 through 2024 in urban residential areas. The resulting dataset, characterized by high spatial but limited temporal coverage at an hourly resolution, was processed using an interpretable tree-based machine learning model (XGBoost) with SHapley Additive exPlanations (SHAP) to generate a temporally continuous daily dataset. This dataset was subsequently integrated into a hierarchical geostatistical model incorporating multiple geographic covariates. Ten-fold cross-validation demonstrated good predictive performance, with an R²ₘₛₑ of 0.90 for the final spatiotemporal exposure model. The quartile misclassification of the BC exposure assessments, as estimated by different exposure models, was evaluated. By integrating an interpretable machine learning approach, this work provides a robust framework for managing a fragmented and unbalanced mobile monitoring dataset and supports high-resolution spatiotemporal exposure assessments for a pregnant cohort study.