Evaluation of random forest regression and multiple linear regression for predicting indoor fine particulate matter concentrations in a highly polluted city.

BACKGROUND: Indoor and outdoor fine particulate matter (PM2.5) are both leading risk factors for death and disease, but making indoor measurements is often infeasible for large study populations. METHODS: We developed models to predict indoor PM2.5 concentrations for pregnant women who were part of a randomized controlled trial of portable air cleaners in Ulaanbaatar, Mongolia. We used multiple linear regression (MLR) and random forest regression (RFR) to model indoor PM2.5 concentrations with 447 independent 7-day PM2.5 measurements and 87 potential predictor variables obtained from outdoor monitoring data, questionnaires, home assessments, and geographic data sets. We also developed blended models that combined the MLR and RFR approaches. All models were evaluated in a 10-fold cross-validation. RESULTS: The predictors in the MLR model were season, outdoor PM2.5 concentration, the number of air cleaners deployed, and the density of gers (traditional felt-lined yurts) surrounding the apartments. MLR and RFR had similar performance in cross-validation (R2 = 50.2%, R2 = 48.9% respectively). The blended MLR model that included RFR predictions had the best performance (cross validation R2 = 81.5%). Intervention status alone explained only 6.0% of the variation in indoor PM2.5 concentrations. CONCLUSIONS: We predicted a moderate amount of variation in indoor PM2.5 concentrations using easily obtained predictor variables and the models explained substantially more variation than intervention status alone. While RFR shows promise for modelling indoor concentrations, our results highlight the importance of out-of-sample validation when evaluating model performance. We also demonstrate the improved performance of blended MLR/RFR models in predicting indoor air pollution.