Machine learning-based identification of underdiagnosed hypertension: a multi-modal analysis of NHANES 2021–2023 data
V Polana, V Godishala, S Tanna, S Karipineni, N Patel, D Nanavaty, S Singh, S ReddyAbstract
Background
Hypertension remains a leading modifiable risk factor for cardiovascular disease, yet many individuals with elevated blood pressure remain undiagnosed. Identifying phenotypes linked to underdiagnosis can enhance targeted screening efforts.
Methods
We analyzed NHANES data from August 2021 to August 2023, selecting adults with measured elevated blood pressure (≥130/80 mmHg). Underdiagnosis was defined as the absence of a self-reported hypertension diagnosis or antihypertensive treatment. Demographic, socioeconomic, and cardiometabolic variables—including BMI, HbA1c, lipid profile, and albumin-to-creatinine ratio—were incorporated with phlebotomy weights (WTPH2YR) for national representation. Data preprocessing involved imputation, standardization, and a stratified 70/30 train-test split. Machine learning models (logistic regression, random forest, HistGradientBoosting, XGBoost) were evaluated using survey-weighted AUC with 95% confidence intervals. SHAP analysis identified key predictors.
Results
Among 2,465 adults with hypertension, 1,158 (47.0%) were underdiagnosed. Logistic regression achieved the highest predictive performance (weighted AUC: 0.731), followed by random forest (0.729) and XGBoost (0.725). At a 0.50 threshold, the logistic model showed weighted precision of 0.689, recall of 0.672, and F1-score of 0.681. SHAP analysis ranked predictors by importance: Age, Total Body Weight, BMI, Total Cholesterol, and Albumin-to-Creatinine Ratio.
Conclusion
Nearly half of adults with elevated blood pressure were underdiagnosed, with younger individuals exhibiting obesity and adverse metabolic profiles disproportionately affected. Elevated BMI and cholesterol increased the risk of underdiagnosis, while a higher albumin-to-creatinine ratio correlated with reduced underdiagnosis, likely reflecting increased clinical attention in renal risk patients. These findings support the use of machine learning to inform data-driven screening strategies for early hypertension detection and prevention of secondary complications.