Predictive clustering methods to identify differential toxicity of air pollutant mixtures

Abstract

BACKGROUND AND AIM: Spatial misalignment between the relatively small number of locations where air pollution mixtures are measured and the locations of cohort members poses a challenge to assessing health effects. This talk presents a clustering-based approach to defining multipollutant exposure profiles that can be well predicted at unobserved locations. METHODS: Clusters are defined by combining multipollutant measurements with spatial and spatiotemporal characteristics of the monitoring location in a Gaussian mixture model. Cluster membership for cohort locations and times are then predicted using parametric and machine learning classification algorithms. RESULTS: We demonstrate this approach using PM2.5 components on a national scale and on-road multipollutant measures (including ultrafine particles and volatile organic compounds) on a regional scale. We identify mixtures associated with differing effects of air pollution exposure on atherosclerosis, blood pressure, and cancer risk. CONCLUSIONS: Cluster-based approaches for defining pollution mixtures provide an effective approach to assessing multipollutant exposures using spatially misaligned data. Incorporating location characteristics can improve predictive accuracy and power for assessing differences in health effects of ambient air pollution exposure. KEYWORDS: ambient air pollution, fine particulate matter, k-means, multipollutant exposures, spatiotemporal statistics