Screening for Probable Undiagnosed Hypertension in US Adults Using Interpretable Machine Learning: An NHANES 2017-2018 Study

Background Hypertension remains one of the most challenging healthcare problems in the community. It is a common, measurable, and treatable condition that is nonetheless responsible for millions of preventable deaths each year. Hypertension affects approximately 1.28 billion adults worldwide, yet fewer than half (46%) are aware of their condition. Undiagnosed hypertension is a critical public health gap, contributing to preventable cardiovascular morbidity through silent end-organ damage. Machine learning can enable community-level screening using routinely available, non-invasive data, without the requirement of laboratory investigations. Methods We conducted a cross-sectional ML study using the National Health and Nutrition Examination Survey (NHANES) 2017-2018 cycle. Adults aged ≥18 years were included; those with a self-reported prior hypertension diagnosis (n=1,930) were retained as negative controls. Probable undiagnosed hypertension was defined as mean blood pressure ≥130/80 mmHg among participants reporting no prior hypertension diagnosis, consistent with the ACC/AHA 2017 threshold. Three classifiers: Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) were trained on eight non-invasive predictor variables. Performance was assessed using AUC-ROC, sensitivity, specificity, and F1 score with bootstrap 95% confidence intervals (CIs). Stratified 5-fold cross-validation was applied. Results Of 9,254 NHANES 2017-2018 participants, 5,237 adults were included after exclusion criteria were applied; 1,072 (20.5%) had probable undiagnosed hypertension. LR achieved the highest test AUC of 0.611 (95% CI: 0.571-0.652), with sensitivity 0.535 (95% CI: 0.473-0.600) and specificity 0.594 (95% CI: 0.560-0.626). RF demonstrated near-absent sensitivity (0.047) despite adequate AUC (0.607), while XGBoost performed intermediately (AUC: 0.599; sensitivity: 0.279). Diabetes status, sex, and age were the most influential predictors by permutation feature importance. Conclusion ML classifiers trained on eight non-invasive variables demonstrated modest but consistent discrimination for identifying probable undiagnosed hypertension, supporting the feasibility of laboratory-free community screening. External validation in diverse populations is warranted before clinical implementation.