Does Lifestyle Information Improve Hypertension Prediction? A Leakage-Controlled Evaluation on NHANES
Organizations: Columbia University, New York, NY, USA
Abstract
Does readily available lifestyle information improve machine-learning prediction of hypertension beyond routine demographic and anthropometric data? On NHANES 2017-March 2020, five model families (Logistic Regression, Random Forest, Gradient Boosting, Bagged Decision Trees, and XGBoost) were evaluated with fold-local preprocessing and paired bootstrap confidence intervals. Adding physical activity, alcohol use, smoking, and sleep information showed no consistent detectable improvement in held-out discrimination: every paired ROC-AUC confidence interval included zero, including under a stricter measured-blood-pressure phenotype and without synthetic oversampling. Mean cross-model SHAP rank agreement fell from 0.873 to 0.806 over all encoded columns, but this decline was largely attenuated over the 19 baseline columns shared by both arms (0.873 vs. 0.863), and likewise under both robustness checks; attribution-agreement comparisons across feature sets of different sizes therefore require a matched comparison space. In exploratory analysis of untreated adults (measured-blood-pressure outcome), all five models showed small positive differences, two with unadjusted intervals excluding zero, meriting confirmation. These results do not imply that lifestyle factors are clinically unimportant; within this cohort, the evaluated lifestyle block added little incremental discrimination beyond an already-informative baseline.
Clinical Prediction
Clinical prediction uses patient records, signals, and images to estimate diagnoses, risks, outcomes, and treatment responses for care.
See all 228 papers