Explainable Machine Learning for Hypertension Status Classification Using NHANES: Performance and Cross-Model SHAP Analysis

Abstract

Hypertension is a major noncommunicable disease and an important public health concern. Its health burden is relevant to public-health risk management and broader disaster-risk contexts, supporting Sustainable Development Goal 3 (SDG 3), particularly Target 3.4 on reducing premature mortality from noncommunicable diseases. This study compares Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) for classifying reported hypertension status, evaluates hyperparameter optimization, and examines predictor contributions using SHapley Additive exPlanations (SHAP). Using 5,008 NHANES 2021–2023 respondents and eight predictors, hypertension status was defined from BPQ020 as self-reported physician diagnosis. The data were divided using an 80:20 stratified split into training and independent test sets, with five-fold cross-validation applied to the training set for hyperparameter optimization. At a fixed threshold of 0.50, LR achieved the highest baseline ROC-AUC (0.7921), while XGBoost achieved the highest baseline Recall (0.5968) and F1-Score (0.6233). Tuning improved RF Recall from 0.5332 to 0.6048 and F1-Score from 0.5793 to 0.6316. SHAP identified Age, Waist circumference, and BMI as the top three predictors in both tuned models, with a Spearman rank correlation of 0.9286 and 100% top-three agreement. Bootstrap analysis yielded a median correlation of 0.9286 (95% CI: 0.9048–1.0000), with 100% top-three agreement across 2,000 replications. At the fixed threshold, tuned RF achieved the highest Recall and F1-Score, whereas LR yielded the highest ROC-AUC.