Machine-learned modeling of PM_(2.5) exposures in rural Lao PDR

摘要

This study presents a machine-learning-enhanced method of modeling PM2.5 personal exposures in a data-scarce, rural, solid fuel use context. Data collected during a cookstove (Africa Clean Energy (ACE)-1 solar-battery-powered stove) intervention program in rural Lao PDR are presented and leveraged to explore advanced techniques for predicting personal exposures to particulate matter with aerodynamic diameter smaller than 2.5 mu m (PM2.5). Mean 48-h PM2.5 exposure concentrations for female cooks were measured for the pre- and post-intervention periods (the "Before" and "After" periods, respectively) as 123 mu g/m(3) and 81 mu g/m(3). Mean 48-h PM(2.5 )kitchen air pollution ("KAP") concentrations were measured at 462 mu g/m(3) Before and 124 mu g/m(3) After. Application of machine learning and ensemble modeling demonstrated cross-validated personal exposure predictions that were modest at the individual level but reasonably strong at the group level, with the best models producing an observed vs. predicted r(2) between 0.26 and 0.31 (r(2) = 0.49 when using a smaller, un-imputed dataset) and mean Before estimates of 119-120 pg/m(3) and After estimates of 86-88 mu g/m(3). This offered improvement over one typical method of predicting exposure - using a kitchen exposure factor (the ratio of exposure to KAP)-which demonstrated an r(2) similar to 0.03 and poorly estimated group average values. The results of these analyses highlight areas of methodological improvement for future exposure assessments of household air pollution and provide evidence for researchers to explore the advantages of further incorporating machine learning methods into similar research across wider geographic and cultural contexts. (C) 2019 Elsevier B.V. All rights reserved.