An Explainable Machine-Learning Framework for Early Cardiovascular Risk Screening in Resource-Constrained Clinics

Abstract

There is still a challenge in clinics lacking laboratory testing, specialist review and computational infrastructure to perform cardiovascular risk assessment. This study details an explainable machine-learning framework for screening cardiovascular mortality over ten years using five cycles of the National Health and Nutrition Examination Survey (NHANES) from 1999-2000 to 2007-2008, that were linked to the 2019 public-use mortality files. Adults who have no self-reported cardiovascular disease (40-79 years of age) are eligible. All nine of the common predictors available are harmonized and tested with logistic regression, random forest and histogram gradient boosting with stratified five-fold cross validation. The following are used to assess performance: ROC-AUC, precision-recall AUC, Brier score, sensitivity, specificity, predictive values, calibration and threshold based clinical utility. Addressing explainability using permutation importance, PDP and patient level attribution. The framework is not based on the requirement for laboratory testing but rather is for use in an offline setting. Empirical cohort counts and predictive estimates are not reported since the NHANES files were not run in the current document-generation environment. The MS is consequently a full and repeatable analysis specification without bogus clinical findings.