Spatiotemporal associations between air pollution and emergency room visits for cardiovascular and cerebrovascular diseases in Korea using a multivariate graph autoencoder modeling approach: an ecological study.

Abstract

Purpose: This study aimed to assess the spatiotemporal associations between air pollution and emergency room visits for cardiovascular and cerebrovascular diseases in South Korea using a graph autoencoder (GAE). A multivariate graph-based approach was used to uncover seasonal and regional variations in pollutant-disease relationships.

Methods: We collected monthly data from 2022 to 2023, including concentrations of 6 air pollutants (SO2, NO2, O3, CO, PM10, and PM2.5) and emergency room visits for 4 disease types: cardiac arrest, myocardial infarction, ischemic stroke, and hemorrhagic stroke. Pearson correlation coefficients were used to construct adjacency matrices, which, along with normalized feature matrices, were used as inputs to the GAE. The model was trained separately for each month and region to estimate the strength of pollutant-disease associations.

Results: The pollutant-disease network structures exhibited clear seasonal variations. In winter, strong associations were observed between O3, NO2, and all disease outcomes. In spring, PM2.5 and PM10 were strongly linked to cardiac and stroke-related visits. These connections weakened during summer but became more pronounced in autumn, especially for NO2 and cardiac arrest. Urban areas displayed denser and stronger associations than non-urban areas.

Conclusion: Our findings underscore the necessity for season- and region-specific air quality management strategies. In winter, focused control of O3 and NO2 is needed in urban areas, while in spring, PM mitigation is required in urban and selected rural regions. Autumn NO2 control may be especially beneficial in non-urban areas. Spatiotemporally tailored interventions could reduce the burden of air pollution-related emergency room visits.