ABSTRACT:
Exposure to air pollution is associated with an increased risk of respiratory diseases (RD) and cardiovascular diseases (CVD). The abundant epidemiological evidence on the health effects of air pollution often focuses on individual pollutants, rather than complex mixtures of simultaneous exposure to multiple air pollutants. Traditional statistical approaches to investigate the effects of air pollution often predict the health outcome of a specific pollutant, and control for confounding by other pollutants. This is a limitation which could result in an underestimation of the health risks. Recently, the application of Machine Learning (ML) has been used to predict the health effects of complex air pollution mixtures. The aim of this study was to use unsupervised ML clustering methods to determine the joint effects of SO2, NO2, O3, PM2.5, and PM10 on hospital admissions for RD and CVD in the Vaal Triangle Airshed Priority Area (VTAPA), South Africa. Due to the missing air pollution from January 2011 to February 2020, data were imputed using the multiple imputation by chain equations (mice) method. K-means and spectral clustering methods were then applied to the air pollution data. A three-cluster spectral clustering model using the normalised Laplacian matrix, showed that the risk of RD hospital admission increased when exposed to SO2, NO2, PM2.5, and PM10 in higher concentration levels, and lower levels of O3 by 1.04 (95 % CI 1.01–1.08). No cluster models showed an increased risk of CVD hospitalisation. Although the clustering methods used in this study were quick to run and analyse, it is evident that more studies need to be done before considering unsupervised ML a reliable and definite tool to study joint effects of air pollution on different health outcomes.