Design-consistent statistical learning for rare survey outcomes: a cardiovascular-kidney-metabolic multimorbidity plasmode benchmark in women
DOI:
https://doi.org/10.29244/currbiomed.5.1.7Keywords:
Cardiovascular-kidney-metabolic syndrome, calibration, complex surveys, multimorbidity, statistical learningAbstract
Background Cardiovascular-kidney-metabolic (CKM) disease integrates metabolic, renal, and cardiovascular systems. Learning in probability samples is challenged by rare outcomes, unequal sampling, clustered dependence, and miscalibration. This study aimed to develop a design-consistent, calibration-first framework for concurrent CKM multimorbidity classification in complex surveys and quantify the consequences of ignoring survey design.
Methods We analyzed the National Health and Nutrition Examination Survey (NHANES) 2017–March 2020 data for U.S. women aged 20–44 years. CKM multimorbidity was defined as the coexistence of at least two of the following: hypertension, diabetes mellitus, chronic kidney disease, and clinical cardiovascular disease. Five learners were compared using 20×5 whole-PSU design-aware cross-validation, fold-contained preprocessing, survey-weighted loss functions, and cross-fitted sigmoid recalibration. The evaluation used design-based and survey-bootstrap procedures across key performance metrics. A 6,000-replicate survey-plasmode benchmark was used to evaluate bias, RMSE, and 95% coverage against finite-population truth.
Results Among 1,743 classifiable women, 44 unweighted cases yielded a survey-weighted prevalence of 1.97% (95% CI, 1.30%–2.65%). The elastic net minimized the calibrated weighted log loss (0.089) and achieved a ROC AUC of 0.797 (95% CI, 0.765–0.830), PR AUC of 0.062 (0.023–0.100), and calibration slope of 0.90. Naive random cross-validation yielded an absolute log-loss bias of 0.0539 and 30.2% coverage; weighted design-aware cross-validation with recalibration achieved 95.0% coverage.
Conclusion For rare CKM outcomes in complex surveys, valid population-level performance estimation requires survey-aware resampling, proper scoring rules, cross-fitted calibration, and design-consistent uncertainties. The framework distinguishes investigator-defined multimorbidity from guideline-aligned CKM staging.
Downloads
Downloads
Additional Files
Published
Issue
Section
License
Copyright (c) 2026 Sunday A Adetunji, Rhoda O Oyewusi, Tomiisin Balogun, Bukola Adetunji, Oluwatimileyin C Adetunji

This work is licensed under a Creative Commons Attribution 4.0 International License.
This journal is published under a Creative Commons Attribution (CC BY 4.0) International License. This license enables reusers to distribute, remix, adapt, and build upon the material in any medium or format, so long as attribution is given to the creator.
