Enhancing Wildfire\nSmoke Exposure Assessment: A Machine\nLearning Approach to Predict Indoor PM<sub>2.5</sub> in British Columbia,\nCanada

Epidemiological studies\ntypically model wildfire smoke exposure\nby predicting outdoor fine particulate matter (PM<sub>2.5</sub>) concentrations,\noverlooking indoor environments where people spend most of their time.\nThis discrepancy can lead to exposure misclassification for wildfire\nsmoke and other air pollutants. We developed a machine learning (ML)\nmodel for estimating daily indoor and outdoor PM<sub>2.5</sub> concentrations\nin British Columbia, Canada, using an ensemble of nonparametric ML\nalgorithms during the 2022 and 2023 wildfire seasons. For model training,\nwe included daily PM<sub>2.5</sub> concentrations collected at 44\ncare facilities equipped with low-cost air quality sensors colocated\nindoors and outdoors. Model predictors for both indoor and outdoor\nPM<sub>2.5</sub> at the facilities included outdoor PM<sub>2.5</sub> and meteorological data from Canada’s National Air Pollution\nSurveillance Program and Purple Air sensors. The indoor and outdoor\nmodels were evaluated with cross validation and then used to compare\nexposure-response relationships for asthma inhaler dispensations,\nas an indicator of population respiratory health. Ensemble models\naccurately predicted PM<sub>2.5</sub> indoors (RMSE = 3.29 μg/m<sup>3</sup>; R<sup>2</sup> = 0.71) and outdoors (RMSE = 3.80 μg/m<sup>3</sup>; R<sup>2</sup> = 0.78). For the out-of-sample validation\nset (2023 wildfire season), the indoor model had a lower RMSE than\nthe outdoor one (RMSE<sub>Indoor</sub> = 6.65 μg/m<sup>3</sup> vs RMSE<sub>Outdoor</sub> = 9.64 μg/m<sup>3</sup>). The effect\nestimates for the relationship between indoor PM<sub>2.5</sub> and\ninhaler dispensations were higher than that for outdoor PM<sub>2.5</sub>. These results suggest that population-scale indoor PM<sub>2.5</sub> exposure assessment is feasible for wildfire smoke epidemiology\nresearch, and that using outdoor estimates may bias the true relationship\ntoward the null.