Modeling air pollution-related hospital admissions employing remote sensing and geographical information systems

Abstract

Land-use regression models and remote sensing data have been widely employed to forecast atmospheric aerosol levels. Recently, these methodologies have been combined to predict the influence of this pollutant on human health. However, traditional land-use regression models do not often consider the complex interactions between predictors, and most of these do not include socioeconomic variables. Thus, in the present study, we aimed to estimate suspended particle-related hospital admissions by employing remote sensing, meteorological, environmental, and demographic parameters. In this cohort study, we analyzed 1,612,049 hospital admissions from Córdoba city, Argentina, from 2005 to 2011, and developed several regression and machine learning land-use models to compare their predictive powers. We found that childhood was the age group with the highest number of hospital admissions related with upper respiratory tract diseases. When predicting population-normalized hospital admissions, the machine learning models, in particular the generalized boosted machine, revealed a better performance than regression models, exhibiting the lowest root mean square error (0.4264) in the test data set. This model also achieved the best R2adj (0.6088) when plotting predicted vs. reported normalized cases. The most important predictors were the meteorological variables, followed by the aerosol optical depth and the planet boundary layer height. Some other predictors, such as educational level, land value, and unsatisfied basic needs, showed less relevance but enhanced the model's prediction power. Furthermore, the predictive power increased after a 1-day lag in hospital admissions (RMSE = 0.4121), highlighting the importance of meteorological and environmental variables in the onset of respiratory diseases.