Dataset and codes for Machine Learning-Based Classification of Self-Reported Cardiovascular Disease History in Africa Using Harmonised Multi-Country WHO STEPS Surveys: 2014–2019

This repository contains the data documentation, derived analytical datasets and outputs, and reproducibility materials supporting the study “Machine Learning-Based Classification of Self-Reported Cardiovascular Disease History in Africa Using Harmonised Multi-Country WHO STEPS Surveys: 2014–2019.” The study used harmonised World Health Organization STEPwise Approach to Noncommunicable Disease Risk Factor Surveillance (WHO STEPS) data from 60,294 adults across 12 African countries: Algeria, Benin, Botswana, Eswatini, Ethiopia, Kenya, Malawi, Morocco, São Tomé and Príncipe, Sudan, Uganda, and Zambia. Surveys were conducted between 2014 and 2019. The primary outcome was self-reported prevalent cardiovascular disease (CVD) history, defined as a reported previous heart attack, angina or chest pain from heart disease, or stroke. The analytical framework incorporated demographic, socioeconomic, behavioural, dietary, and biological characteristics, including age, sex, residence, education, occupation, marital status, tobacco use, alcohol-related harm, physical activity, dietary indicators, hypertension, diabetes, cholesterol status, country, and survey year. The study compared four classification approaches: Elastic Net logistic regression (LASSO), Random Forest, XGBoost, and unpenalized logistic regression. Model development used an 80:20 training-test split, with SMOTE restricted to the training data. Five-fold cross-validation was used for model tuning. Performance was evaluated using area under the precision-recall curve (AUC-PR) as the primary discrimination metric, together with AUC-ROC, sensitivity, specificity, positive predictive value, F1 score, balanced accuracy, calibration, and Brier score. Additional analyses examined detection bias associated with clinically ascertained predictors, SMOTE versus class weighting, probability recalibration, complete-case analysis, chained-equations multiple imputation, survey-weight sensitivity, PSU-level cluster bootstrap analysis, and country-held-out transportability. Adapted Framingham and WHO/ISH cardiovascular risk scores were included as contextual benchmarking approaches. The repository is intended to facilitate transparency, methodological reproducibility, secondary methodological research, and evaluation of machine-learning approaches for cardiovascular disease surveillance in African population-based surveys. Important data-access statement: The original individual-level WHO STEPS microdata are governed by WHO data-access and use conditions and are not redistributed through this repository. Researchers wishing to obtain the underlying participant-level WHO STEPS datasets should apply through the WHO STEPS data-access process. Materials deposited here should therefore be understood as reproducibility resources, derived data products, metadata, analytical outputs, and/or code that can be used with appropriately authorised WHO STEPS data. The cardiovascular disease outcome used in the study is cross-sectional and self-reported. Consequently, the models classify prevalent self-reported CVD history and should not be interpreted as prospective cardiovascular risk-prediction models or clinical decision-support tools.