LAUSR

BACKGROUND Despite mounting evidence, the impact of the interplay between weather and pollution features on the risk of acute cardiac and cerebrovascular events has not been entirely appraised. The aim of this study was to perform a comprehensive cluster analysis of weather and pollution features in a large metropolitan area, and their association with acute cardiac and cerebrovascular events. METHODS Anonymized data on acute myocardial infarction (AMI) and acute cerebrovascular events were obtained from 3 tertiary care center from a single large metropolitan area. Weather and pollution data were obtained averaging measurements from several city measurement stations managed by the competent regional agency for enviromental protection, and from the Metereological Center of Italian Military Aviation. Unsupervised machine learning was performed with hierarchical clustering to identify specific days with distinct weather and pollution features. Clusters were then compared for rate of acute cardiac and cerebrovascular events with Poisson models. RESULTS As expected, significant pairwise correlations were found between weather and pollution features. Building upon these correlations, hierarchical clustering, from a total of 1169 days, generated 4 separate clusters: mostly winter days with low temperatures and high ozone concentrations (cluster 1, n=60, 5.1%), days with moderately high temperatures and low pollutants concentrations (cluster 2, n=419, 35.8%), mostly summer and spring days with high temperatures and high ozone concentrations (cluster 3, n=673, 57.6%), and mostly winter days with low temperatures and low ozone concentrations (cluster 4, n=17, 1.5%). Overall cluster-wise comparisons showed significant overall differences in adverse cardiac and cerebrovascular events (p<0.001), as well as in cerebrovascular events (p<0.001) and strokes (p=0.001). Between-cluster comparisons showed that Cluster 1 was associated with an increased risk of any event, cerebrovascular events, and strokes in comparison to Cluster 2, Cluster 3 and Cluster 4 (all p<0.05), as well as AMI in comparison to Cluster 3 (p=0.047). In addition, Cluster 2 was associated with a higher risk of strokes in comparison to Cluster 4 (p=0.030). Analysis adjusting for season confirmed the increased risk of any event, cerebrovascular events and strokes for Cluster 1 and Cluster 2. CONCLUSIONS Unsupervised machine learning can be leveraged to identify specific days with a unique clustering of adverse weather and pollution features which are associated with an increases risk of acute cardiovascular events, especially cerebrovascular events. These findings may improve collective and individual risk prediction and prevention.