Understanding Health Effects of Particulate Matter Sources by Using Stochastic Machine Learning Models

Abstract

Particulate matter (PM) is a complex mix of organic and inorganic compounds of distinct sources, with a range of physical and chemical properties, which might have a different harmful effects to health. Disentangling total ambient PM concentration into its sources is key for developing strategies to reduce PM through targeted actions. Current methods to identify sources of particulate pollution typically require \emph{a priori} specification of the number of sources and do not include information on covariates in the source allocations. In this work, we develop a comprehensive approach for source apportionment of airborne particles by using machine learning probabilistic models. We proposed a Bayesian nonparametric approach through a Dirichlet process mixture models that enables the better understanding of hidden structures in multi-pollutants and allow to accommodate complex patterns of temporal dependencies as well as for concomitant processes (e.g. meteorology) in the prediction of the source contributions. Then we evaluate the health effects of the sources. To illustrate our model framework, we applied it to the PM$_{10}$ chemical composition data measured at an urban background site (North Kensington) in London, UK, from 2011 to 2012. The health data will be related to cardio-respiratory hospital admission.