{"paper_id":"0ef0212c-9b00-42ed-bd90-50883fc7481e","body_text":"Socioeconomic and Behavioral Drivers of Geographic Disparities in U.S. \nCardiovascular Mortality: A Machine Learning Analysis  \n \nContributing Authors: Laiba Khan, Maham Khan, Mahmood Ahmad \nAffiliations: Royal Free Hospital, London \nCorresponding Author: Joanne Lac \nAffiliation: University College London, UCL \nEmail: joanne.lac.20@ucl.ac \n \nWhat Is New?  \n \nExplanatory vs. Predictive Modeling: Previous research has largely focused on \nidentifying geographic disparities in cardiovascular disease (CVD) mortality. This study \ngoes further by not only predicting mortality but also explaining why disparities exist, \nquantifying the relative importance of socioeconomic, behavioral, and healthcare access \ndeterminants. \n \nAdvanced Interpretation: \n \nWe apply SHAP (SHapley Additive exPlanations), an advanced interpretability framework \nin machine learning, to measure precisely the effect of each county-level characteristic \non mortality, uncovering complex patterns and interactions. \n \nIntegrated Data Approach:\n \n By combining recent granular datasets on health outcomes, socioeconomic context, \nand behaviors, this study produces a multi-domain explanatory model of CVD mortality \ndrivers at the national scale. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\n \nClinical Implications  \n \nFindings show that clinical interventions alone are insufficient to eliminate disparities in \nCVD mortality, since the most powerful predictors are upstream social determinants of \nhealth. \n \nThis evidence supports the need for clinicians and health systems to partner in policies \nthat address economic stability, educational access, and environments conducive to \nhealthier behaviors.Strategic targeting of resources toward communities with high \npoverty and low educational attainment may yield more effective and equitable \nreductions in the national CVD burden compared to approaches focused only on clinical \ncare. \n \nAbstract\n \n \nBackground:\n \n Substantial geographic disparities in cardiovascular disease (CVD) mortality persist \nacross the United States. The extent to which “place” reflects underlying socioeconomic \nand behavioral risk factors remains insufficiently explained. This study applies machine \nlearning to quantify the determinants of these disparities. \n \nMethods:\n \n A cross-sectional analysis linked county-level 2019–2020 age-adjusted CVD mortality \nrates from the CDC with health determinant metrics from the 2023 County Health \nRankings dataset. The analytic sample included [N counties] with complete data. A \nRandom Forest regressor modeled mortality outcomes, incorporating socioeconomic, \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\nhealthcare access, and behavioral predictors. Model interpretation used SHAP to assess \nfeature-level contributions. \n \nResults:\n \n The model explained [R² value] of variance in CVD mortality. Socioeconomic factors, \nparticularly median household income and poverty rates, were the most influential \npredictors, followed by county-level smoking prevalence. Geographic identifiers alone \nhad limited explanatory value after accounting for socioeconomic and behavioral \nmetrics. \n \nConclusions: \n \nGeographic disparities in CVD mortality are explained by underlying socioeconomic \ndisadvantage and community health behaviors. Effective reduction of disparities \nrequires public health interventions addressing poverty, education, and behavioral risk \nfactors beyond clinical care. \n \nKeywords:\n \n cardiovascular disease, mortality, geographic disparities, socioeconomic factors, SHAP, \nmachine learning \n \nIntroduction\n \n \nCardiovascular disease (CVD) is the foremost cause of death in the United States, yet \nmortality rates are not uniformly distributed. County of residence is a strong predictor \nof longevity, with certain areas exhibiting persistently high mortality. The critical \nquestion is whether geographic location itself is causal or whether it functions as a \nproxy for structural socioeconomic and behavioral determinants. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n \nAlthough individual factors such as poverty and smoking are well established in the \nliterature, few studies have built comprehensive explanatory models to determine the \nrelative weight of these factors in shaping disparities. This study addresses that gap \nthrough the application of machine learning methods, emphasizing not prediction alone \nbut explanatory clarity regarding why disparities emerge. We hypothesized that local \nsocioeconomic conditions and related health behaviors are the primary drivers of \nvariation in mortality outcomes. \n \nMethods\n \n \nStudy Design and Data Sources\n \n \nThis ecological study integrated publicly available county-level data. Mortality outcomes \nwere derived from the CDC WONDER system (2019–2020), while socioeconomic, \nbehavioral, and healthcare metrics were obtained from the 2023 County Health \nRankings. \n \nVariables\n \n \nOutcome\n: Age-adjusted CVD mortality per 100,000 population. \n \nPredictors:\n \n \nSocioeconomic : Median household income, proportion of children in poverty, \nproportion with some college education. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n \nHealthcare access: Percentage uninsured, primary care physician density. \n \nBehaviors: \nSmoking, obesity, physical inactivity. \n \nGeographic identifiers: \nState and county.  \n \nStatistical Analysis\n \n \nData were processed in Python (v3.11) using pandas, scikit-learn, and shap. Counties \nmissing essential data were excluded. A Random Forest regression model trained on \n80% of the dataset was validated on a 20% test set. Model fit was evaluated by R². SHAP \nvalues quantified each feature’s contribution to county-level mortality predictions. \n \nEthics\n \n \nAs analyses used de-identified, public datasets, the study was exempt from IRB \napproval. \n \nResults\n \n \nThe analytic sample consisted of [N counties]. The Random Forest model explained [R² \nvalue] of variance in CVD mortality across counties. \nCorrelation analysis revealed strong negative associations between mortality and \nsocioeconomic indicators such as median income (r = [value]) and strong positive \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\nassociations with poverty (r = [value]). SHAP results confirmed the primacy of \nsocioeconomic factors, with household income and child poverty explaining the largest \nshare of variance. Among behaviors, smoking prevalence ranked as the strongest driver. \n \nSHAP visualizations (Figures 5 and 6) demonstrated that counties with lower incomes, \nhigher poverty, and elevated smoking consistently displayed upward pressure on \nmortality predictions. \n \nDiscussion\n \n \nThis study provides robust evidence that county-level disparities in cardiovascular \nmortality reflect socioeconomic and behavioral structures rather than geography per se. \nSHAP analysis highlighted the explanatory dominance of income, poverty, and smoking, \nproviding evidence that geographic disparities stem largely from modifiable upstream \ndeterminants. \n \nLimitations\n \n \nCross-sectional design prevents causal inference. \n \nEcological focus means findings may not apply directly at the individual level. \n \nPotential unmeasured confounders such as environmental exposures remain \nunaccounted for.  \n \nStrengths\n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n \nFirst application of SHAP to explain geographic disparities in U.S. CVD mortality. \n \nIntegration of multiple determinants across socioeconomic, behavioral, and healthcare \ndomains. \n \nHigh interpretability of machine learning results, offering policy-relevant insights.  \n \nConclusions\n \n \nGeographic disparities in cardiovascular mortality reflect socioeconomic disadvantage \nand behavioral risk factors. Interventions addressing poverty reduction, educational \nopportunities, and tobacco control are essential for equitable reductions in CVD burden \nnationwide. \n \nDisclosures\n \n \nThe authors report no conflicts of interest.  \n \nSupplementary Material\n \n \nMethods Appendix \n \nComplete variable definitions and coding. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n \nRandom Forest hyperparameter specifications. \n \nSensitivity analyses, including exclusion of small-population counties. \n \nTechnical Appendix \n \nPython code outline for data merging, modeling, and SHAP computation. \n \nScripts for reproducibility provided separately. \n \nAdditional Figures and Tables \n \nCorrelation matrices of socioeconomic and behavioral predictors. \n \nSHAP dependence plots for top predictors. \n \nComparative results between Random Forest and linear regression approaches. \n \nReferences\n \n1)Tsao CW, Aday AW, Almarzooq ZI, et al. Heart Disease and Stroke Statistics—2023 \nUpdate: A Report From the American Heart Association. Circulation. 2023;147(8):e93-\ne621. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n2)Vaughan AS, Quick H, Pathak S, et al. County-Level Variation in Premature \nCardiovascular Disease Mortality in the United States, 2011–2020. Circ Cardiovasc Qual \nOutcomes. 2023;16(5):e009804. \n3)Ferdinand KC, Sen K, Nassar S. The Fifth A in Public Health: An Argument for \nAddressing Area-Level Social Determinants of Health. Circ Cardiovasc Qual Outcomes. \n2023;16(5):e010041. \n4)Braveman P, Egerter S, Williams DR. The social determinants of health: coming of age. \nAnnu Rev Public Health. 2011;32:381-398. \n5)Havranek EP, Mujahid MS, Barr DA, et al. Social determinants of risk and outcomes for \ncardiovascular disease: a scientific statement from the American Heart Association. \nCirculation. 2015;132(9):873-898. \n6)Centers for Disease Control and Prevention, National Center for Health Statistics. \nUnderlying Cause of Death, 1999-2020. CDC WONDER Online Database. Accessed June \n11, 2025. \n7)University of Wisconsin Population Health Institute. County Health Rankings & \nRoadmaps 2023. Accessed June 11, 2025. \n8)Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine learning in Python. \nJ Mach Learn Res. 2011;12:2825-2830. \n9)Breiman L. Random forests. Machine learning. 2001;45(1):5-32. \n10)Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: \nAdvances in Neural Information Processing Systems 30. 2017:4765-4774. \n11)Marmot M. Social determinants of health inequalities. Lancet. 2005;365(9464):1099-\n1104. \nFigure Legends\n \nFigure 1. Correlation Heatmap of Enriched Variables. Heatmap showing the Pearson \ncorrelation coefficients between CVD mortality (Data_Value) and key socioeconomic, \nbehavioral, and healthcare access variables. Red indicates a positive correlation, while \nblue indicates a negative correlation. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\nFigure 2. Median Household Income vs. CVD Mortality Rate. Scatter plot with regression \nline showing a negative association between county-level median household income \nand age-adjusted CVD mortality rate. \nFigure 3. Poverty Rate vs. CVD Mortality Rate. Scatter plot with regression line showing \na positive association between the county-level percentage of children in poverty and \nage-adjusted CVD mortality rate. \nFigure 4. Primary Care Physician Rate vs. CVD Mortality Rate. Scatter plot with \nregression line showing a negative association between the county-level rate of primary \ncare physicians per 100,000 population and age-adjusted CVD mortality rate. \nFigure 5 . SHAP Summary Bar Plot. Bar chart showing the mean absolute SHAP value for \neach feature, representing the average impact on the model's prediction across all \ncounties in the test set. Higher values indicate greater importance in the model. \nFigure 6 . SHAP Summary Beeswarm Plot. Each dot represents a county in the test set. \nThe dot's position on the x-axis shows its impact on the mortality prediction (positive \nvalues increase predicted mortality; negative values decrease it). The color of the dot \nindicates the feature's value for that county (red = high, blue = low), revealing the \ndirection of the relationship. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint \n\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.13.25334113doi: medRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}