Interpretable machine learning for identifying determinants of high hypertension burden under extreme heat vulnerability: evidence from Maryland, USA

Extreme heat poses increasing risks to cardiovascular health, yet fine-scale determinants of heat-related hypertension burden remain insufficiently understood. This study examines high hypertension burden in a heat-vulnerability context across 1,385 census tracts in Maryland using an interpretable machine learning framework. An XGBoost model was developed to classify census tracts with high hypertension burden using demographic, socioeconomic, built-environment, heat anomaly, and adaptive-capacity variables. SHapley Additive exPlanations (SHAP) were then used to identify key predictors, assess their contribution to model predictions, and examine non-linear effects. The model achieved strong predictive performance, with an accuracy of 0.819, balanced accuracy of 0.806, ROC-AUC of 0.821, and PR-AUC of 0.909. Results show substantial spatial heterogeneity in hypertension burden, with high-burden tracts concentrated in Baltimore City, Prince George's County, southern Maryland, western Maryland, and parts of the Eastern Shore. SHAP results indicate that African-American population share, older-adult population share, and low educational attainment were the strongest predictors, followed by summer maximum air temperature, non-vegetated land area, and lack of air conditioning. Dependence plots further reveal non-linear and threshold-like relationships, suggesting that predicted risk increases sharply beyond certain levels of demographic vulnerability, educational disadvantage, heat exposure, and limited cooling access. These findings indicate that high hypertension burden in a heat-vulnerability context is shaped by the intersection of structural social vulnerability, demographic susceptibility, environmental exposure, and household adaptive capacity. The study demonstrates the value of combining XGBoost and SHAP for tract-level heat-health risk assessment and provides policy-relevant evidence for targeted heat adaptation, cooling assistance, and public health preparedness in Maryland.