Generalizable Prediction of Ultrafine Particle Concentration Using Combined Multi-Site Measurements

• Activity Number: 534 - Tradeoff Between Risks and Benefits When Transporting Model Under Distribution Shift
• Type: Invited
• Date/Time: Thursday, August 11, 2022 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Prasad Patil* and Sean Mueller and Neelakshi Hudda and John Durant and Jon Levy and Kevin Lane
• Companies: Boston University School of Public Health and Boston University School of Public Health and Tufts University and Tufts University and Boston University School of Public Health and Boston University School of Public Health
• Keywords: air pollution; environmental health; machine learning; ensembling; heterogeneity

Abstract

Atmospheric concentration of ultrafine particles (UFP; particles < 100nm) has the potential to affect human health in a distinct manner from larger particles such as fine particulate matter mass (PM2.5). In a source apportionment study of aviation-attributable UFP Particle Number Concentrations (PNC) from flight activity at Boston Logan Airport, our team implements both stationary monitoring at varying distances along flightpaths from the airport and a mobile monitoring campaign that completes a circuit around the airport at different times of day. One goal of collecting this data is to develop a predictive model for UFP concentration using meteorological, land use, and automobile and air traffic measurements that are generalizable to regions where monitoring cannot be conducted. We explore the efficacy of combining data and ensembling predictors from multiple stationary sites in a manner that incorporates varying spatial configurations of monitors relative to the source. We validate and assess the generalizability of these multi-site combination techniques using both leave-one-site-out cross-validation and externally with the mobile monitoring data.