Abstract P276: De Novo Exposomic Geospatial Assembly of the Stroke Belt with Machine Learning and Network Analysis

Introduction: Geographical chronic disease patterns, such as the high stroke rates in the southeast United States that characterize the Stroke Belt, are known to vary with chronic disease rates and changes in the social determinants of health (SDOH). However, the role of regional variations in the exposome (quantified by pollution measures) is less clear. Hypothesis: We investigated the hypothesis that there is a geographical relationship between high-dimensional pollution indicators and stroke, and examined if this relationship could be identified using a novel machine-learning pipeline called aPEER (algorithm for Projection of Exposome and Epidemiological Relationships). Methods: aPEER assembles geographical regions from pollution and chronic disease data at the county and census-tract level using machine-learning, and validates findings with network analysis (Figure 1A). Multiple rounds of aPEER analyses were completed with different pollution datasets to iteratively identify the most important set of pollution measures that best predict the geographical distribution of the Stroke Belt and high stroke mortality regions (Figure 1B). Results: Using aPEER, the pollutants formaldehyde, methanol, acetaldehyde, benzene and carbon tetrachloride predicted high stroke mortality regions including the Stroke Belt with superior accuracy (area under the curve (AUC)=0.8417-0.9931) compared to SDOH (AUC=0.6853-0.8860) and chronic disease (AUC=0.8367-0.9731) indicators. Conclusion: In conclusion, we identified five pollutants associated with stroke geography using a combination of machine-learning methods comprising the aPEER pipeline. These pollutants showed superior predictive power compared to conventional SDOH and chronic disease risk factors, highlighting the potential role of the exposome in stroke pathogenesis.