The Canadian Optimized Statistical Smoke Exposure Model (CanOSSEM): A machine learning approach to estimate national daily fine particulate matter (PM2.5) exposure

• DOI: 10.1016/j.scitotenv.2022.157956
• Publication Date: 2022-08-15T05:22:25Z
• Authors: Naman Paul; Jiayun Yao; Kathleen E. McLean; David M. Stieb; Sarah B. Henderson
• Keywords: Exposure Assessment

Exposure to biomass smoke has been associated with a wide range of acute and chronic health outcomes. Over the past decades, frequency intensity wildfires increased in many areas, resulting longer episodes higher concentrations fine particulate matter (PM2.5). There are also communities where seasonal open burning residential wood heating have short- long-term impacts on ambient air quality. Understanding effects exposure requires reliable estimates PM2.5 during wildfire season throughout year, particularly areas without regulatory quality monitoring stations. We developed machine learning approach estimate across all populated regions Canada from 2010 2019. The random forest model uses potential predictor variables integrated multiple data sources daily mean (24-hour) at 5 km × spatial resolution. training prediction datasets were generated using observations National Air Pollution Surveillance (NAPS) network. Root Mean Squared Error (RMSE) between predicted observed was 2.96 μg/m3 for entire set, more than 96 % predictions within NAPS measurements. evaluated 10-fold, leave one-region-out, leave-one-year-out cross-validations. Overall, CanOSSEM performed well but performance sensitive removal large events such as Fort McMurray interface fire May 2016 or extreme 2017 2018 seasons British Columbia. will be useful epidemiologic studies exposure, especially populations affected by routine measurements not available.