K-means cluster analysis of cooperative effects of CO, NO2, O3, PM2.5, PM10, and SO2 on incidence of type 2 diabetes mellitus in the US

Air pollution (AP) has been shown to increase the risk of type 2 diabetes mellitus, as well as other cardiometabolic diseases. AP is characterized by a complex mixture of components for which the composition depends on sources and metrological factors. The US Environmental Protection Agency (EPA) monitors and regulates certain components of air pollution known to have negative consequences for human health. Research assessing the health effects of these components of AP often uses traditional regression models, which might not capture more complex and interdependent relationships. Machine learning has the capability to simultaneously assess multiple components and find complex, non-linear patterns that may not be apparent and could not be modeled by other techniques. Here we use k-means clustering to assess the patterns associating PM 2.5 , PM 10 , CO , NO 2 , O 3 , and SO 2 measurements and changes in annual diabetes incidence at a US county level. The average age adjusted annual decrease in diabetes incidence for the entire US populations is −0.25 per 1000 but the change shows a significant geographic variation (range: −17.2 to 5.30 per 1000). In this paper these variations were compared with the local daily AP concentrations of the pollutants listed above from 2005 to 2015, which were matched to the annual change in diabetes incidence for the following year. A total of 134,925 daily air quality observations were included in the cluster analysis, representing 125 US counties and the District of Columbia. K-means successfully clustered AP components and indicated an association between exposure to certain AP mixtures with lower decreases on T2D incidence. • Epidemiological studies tend to focus on a single air pollution component, potentially missing synergistic effects. • The concurrent impact of PM 2.5 , PM 10 , CO, NO 2 , O 3 , and SO 2 on type-2 diabetes incidence was assessed at the US county level. • The cluster with the greatest CO, NO 2 , PM 2.5 , PM 10 , and SO 2 concentrations had the smallest decrease in diabetes incidence. • Individual cluster components did not trend with diabetes incidence, possibly indicating cooperative or threshold effects. • Machine learning methods can assess concurrent effects of air pollution components, revealing potentially non-linear patterns.