Neighborhood-level stroke and its neighborhood-level determinants: a Bayesian machine learning approach

Abstract

The substantial health and economic burden of stroke has prompted numerous studies examining the risk factors of stroke using individual patient level data. Increasingly, place-based evidence is recognized as a critical part of stroke management, but descriptions of risk factors and their associations with stroke at the same neighborhood level smaller than counties are lacking.We applied a novel Bayesian machine learning approach to an integrated data set consisted of 24 potential variables to identify factors with high predictive power for stroke and quantify the associations between predictors and stroke prevalence. Both the stroke outcome and its predictors were measures at neighborhood level, defined as census tracts (n=26697) from 500 large cities in the United States. Variable selection was conducted using Bayesian Additive Regression Trees (BART)-Machine, and compared with conventional models: step-wise regression and regression with all predictors. The exposure-outcome associations were further assessed using Bayesian linear regression. We identified six high importance tract-level predictors for tract-level stroke prevalence. They were prevalence of no leisure-time physical activity (%LPA), proportions of population who were aged 65 years and above (%OlderAdults) and who were non-Hispanic black (%NHB), median household income, and ozone level, as well as the interaction term between %LPA and %OlderAdults. Among these factors, %OlderAdults, %NHB, %LPA, and ozone were positively associated with stroke prevalence, while median household income was inversely associated with stroke prevalence. The interaction term showed an exacerbated adverse effect of aging population structure and low physical activity on tract-level stroke prevalence.High-performance machine learning identified the most important determinants of neighborhood-level cardiovascular health from a wide-ranging variables in an agnostic, data-driven and reproducible way. The identified neighborhood-level predictors were consistent with known patient-level risk factors. The results can be used to prioritize and allocate resources to develop targeted neighborhood-level interventions for stroke prevention and control.