Incorporating geospatial environmental exposure indicators in individual hypertension risk prediction: a multi-stage machine learning pipeline

Abstract Background Environmental exposures are known contributors to chronic disease but are rarely incorporated into risk prediction models. Objective We demonstrate a staged machine learning approach to incorporating geospatially measured neighborhood social and ambient environmental exposure measures into hypertension risk prediction. Methods We analyzed data from 10,491 adults in the Gulf Long-Term Follow-Up (GuLF) Study. Hypertension was defined by measured blood pressure and medication use. We assessed incremental predictive performance across three stages: Stage 1 (age, sex, race, BMI); Stage 2 (Stage 1 + neighborhood social factors); Stage 3 (Stage 2 + ambient exposures). Variable selection combined Boruta and bootstrapped area under the precision-recall curve (AUPRC). Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGB) were trained and evaluated for discrimination, calibration, and classification. SHAP (Shapley Additive Explanations) was used to interpret variable contributions. Results Participants had a mean age of 43.6 years with a standard deviation of 13.02 years; 78.4% were male, 52.3% were non-Hispanic White, and 35.5% had hypertension. Model-selected environmental predictors included neighborhood disadvantage, community resilience, social vulnerability (Stage 2), vegetation, PM₂.₅, NO₂, and formaldehyde (Stage 3). AUC and AUPRC showed minimal change across stages; in the XGB model, sensitivity was 0.775 (Stage 1), increased to 0.797 (Stage 2), and was 0.784 (Stage 3), with a corresponding precision trade-off (0.525→0.517→0.524). SHAP identified vegetation, social vulnerability, area deprivation, PM₂.₅, formaldehyde, and community resilience scores as leading environmental contributors. Impact Environmentalexposures have been linked to an increased risk of hypertension but have rarely been incorporated into risk prediction models. Leveraging data from a large prospective cohort, we developed an interpretable machine learning pipeline that screens and incorporates relevant geospatial socioeconomic and environmental exposures to individual-level risk prediction models. While adding these selected exposures provided modest improvements in model sensitivity, improved sensitivity identified additional hypertension cases that would have been missed. Even small gains in sensitivity can translate to earlier identification of additional at-risk individuals who might benefit from interventions, which can lead to public health improvements.