A Random Forest Model for PM2.5 Personal Exposure Assessment for a Chinese Cohort

Abstract

BACKGROUND AND AIM: Errors in air pollution exposure assessment are often considered a major limitation in epidemiologic studies. However, it is difficult to obtain accurate personal level exposure on cohort populations due to the often prohibitively expensive costs. Personal exposure estimation models are used in lieu of personal exposures, but still suffer from the issues of availability and accuracy. We aim to establish a personal PM2.5 exposure assessment model for a cohort population, and assess its performance by applying our model on cohort subjects. METHODS: We analyzed data from representative sites selected from the Sub-Clinical Outcomes of Polluted Air in China (SCOPA-China) cohort study, and established a random forest model to estimate PM2.5 personal exposure. Parameters obtained from questionnaires and outdoor and meteorological monitoring sites were pre-screened using the Boruta algorism, and parameter contributions were evaluated using the rank of variable importance. We also applied the model among subjects recruited in the above project within the same area and study period to estimate the reliability of the model. RESULTS:The established model showed good fit with an R2 of 0.81. Ambient PM2.5 contributed the most to personal exposure concentrations, and indoor passive smoking, meteorological parameters, and durations in different microenvironments also ranked high in feature importance. The model application results showed similar patterns with empirically measured data, supporting the performance of our established model. CONCLUSIONS:Our pilot study provided a validated and feasible modeling approach to assess personal PM2.5 exposure for large cohort populations, and the preliminary application among cohort subjects proved satisfactory. The promising model framework can improve PM2.5 exposure assessment accuracy for future environmental health studies of large populations. KEYWORDS: Particulate matter, Exposure assessment, Methodological study design, Modeling