Enhancing Wildfire Smoke Exposure Assessment: A Machine Learning Approach to Predict Indoor PM_2.5 in British Columbia, Canada

Abstract

Epidemiological studies typically model wildfire smoke exposure by predicting outdoor fine particulate matter (PM 2.5 ) concentrations, overlooking indoor environments where people spend most of their time. This discrepancy can lead to exposure misclassification for wildfire smoke and other air pollutants. We developed a machine learning (ML) model for estimating daily indoor and outdoor PM 2.5 concentrations in British Columbia, Canada, using an ensemble of nonparametric ML algorithms during the 2022 and 2023 wildfire seasons. For model training, we included daily PM 2.5 concentrations collected at 44 care facilities equipped with low-cost air quality sensors colocated indoors and outdoors. Model predictors for both indoor and outdoor PM 2.5 at the facilities included outdoor PM 2.5 and meteorological data from Canada's National Air Pollution Surveillance Program and Purple Air sensors. The indoor and outdoor models were evaluated with cross validation and then used to compare exposure-response relationships for asthma inhaler dispensations, as an indicator of population respiratory health. Ensemble models accurately predicted PM 2.5 indoors (RMSE = 3.29 μg/m 3 ; R 2 = 0.71) and outdoors (RMSE = 3.80 μg/m 3 ; R 2 = 0.78). For the out-of-sample validation set (2023 wildfire season), the indoor model had a lower RMSE than the outdoor one (RMSE Indoor = 6.65 μg/m 3 vs RMSE Outdoor = 9.64 μg/m 3 ). The effect estimates for the relationship between indoor PM 2.5 and inhaler dispensations were higher than that for outdoor PM 2.5 . These results suggest that population-scale indoor PM 2.5 exposure assessment is feasible for wildfire smoke epidemiology research, and that using outdoor estimates may bias the true relationship toward the null.