Uncertainty-Aware Multi-Outcome Screening for Undiagnosed Cardiometabolic Disease in the United States Population Using NHANES

<title>Abstract</title> <p>Background Cardiometabolic diseases, including diabetes, hypertension, dyslipidemia, and chronic kidney disease, are leading causes of morbidity and mortality, and many affected adults remain undiagnosed until complications appear. Machine learning can support population screening, but model probabilities are often poorly calibrated and give no indication of when a prediction should be trusted. Using the National Health and Nutrition Examination Survey (NHANES) 2011 to 2014, we assembled an analytical cohort of 20,374 adults aged 18 years or older and developed a screening framework that pairs multi-outcome prediction with conformal prediction. Five outcomes representing previously undiagnosed disease were defined by combining the absence of a self-reported physician diagnosis with objective biomarker or examination evidence, and any biomarker used to define an outcome was excluded as a predictor for that outcome. Three nested deployment scenarios represented community, primary-care, and laboratory-supported settings, and six models were compared. Results In the analytical cohort, 26.3% of adults had evidence of at least one undiagnosed cardiometabolic condition, and the survey-weighted prevalence of any latent disease was 26.6% (95% CI 25.1 to 28.2). Discrimination improved as clinical information increased; for the composite outcome the Random Forest AUROC rose from 0.789 in the community scenario to 0.913 with examination data and 0.964 with laboratory data. Under routine clinical assessment, tree-ensemble models reached an AUROC of 0.984 for dyslipidemia, 0.977 for hypertension, 0.922 for diabetes, and 0.999 for possible chronic kidney disease risk, and repeated cross-validation confirmed that these estimates were stable across folds. Conformal prediction achieved empirical coverage close to the 90% target for every model, yet the fraction of individuals flagged as uncertain differed widely between models with similar discrimination, ranging from 0.3% for Random Forest to 59.9% for logistic regression on the composite outcome. Conclusions Reporting a prediction together with its uncertainty makes it possible to screen for multiple undiagnosed cardiometabolic conditions across realistic care settings while identifying the individuals for whom a screening result is not reliable and confirmatory testing is warranted. The uncertainty rate is a decision-relevant quantity that discrimination metrics alone do not reveal.</p>