A Spatial Machine Learning Approach for Analyzing Population-Level Factors Associated with Emergency Department Visits with Diabetes Diagnoses

Emergency department (ED) visits for diabetes arise from complications related to both acute and chronic symptoms including, but not limited to, cardiovascular diseases, renal failure, neurological issues, and retinopathy. Research has shown that such complications disproportionately affect racial and ethnic minority populations. This is likely due to the complex pathogenesis of diabetes, which includes both individual-level factors such as race and family history and population-level drivers such as neighborhood walkability and access to healthy food outlets, which may be heavily influenced by residential segregation. We use spatial machine learning and statistical methods to estimate population-level factors associated with ED visits with diabetes diagnoses at the ZCTA level in Maryland as stratified by race and local measures of residential segregation. Using XGBoost and SHAP, we identified population-level factors associated with diabetes-coded ED visits and used logistic regression to estimate odds ratios. Our findings show that ED visits were associated with factors including housing vacancy, heavy traffic, greenspace, and health behaviors like physical leisure time and smoking across all subgroups. Further, we found that the factors associated with diabetes-coded ED visits differ across racial groups and across communities with different patterns of residential segregation. Heavy traffic, a proxy for traffic-related air pollution and barriers to walkability, was consistently linked to higher odds of diabetes-coded ED visits, especially among Black patients.