Machine learning⍰based prediction of cardiovascular disease risk in Africa using WHO Stepwise Surveys: 2014-2019

Abstract

ABSTRACT Introduction Cardiovascular diseases (CVDs) are the leading cause of death globally, with rising burdens in Africa due to ageing populations, lifestyle changes, and poor risk factor control. Conventional risk scores developed in high-income settings often perform poorly in African populations. Machine-learning (ML) approaches offer potential to improve prediction by capturing complex, non-linear interactions among demographic, behavioural, and biological factors. This study applies ML models to WHO STEPS survey data to generate context-specific CVD risk predictions across 12 African countries. Methods We analysed data from 60,294 adults collected in WHO STEPS surveys between 2014 and 2019 across 12 African countries. Three ML models; Elastic Net logistic regression (LASSO), Random Forest (RF), and XGBoost (XGB); were trained to predict self-reported CVD outcomes. Data were split into training (80%) and testing (20%) sets with five-fold cross-validation. Feature selection used the Boruta algorithm, and model performance was assessed via accuracy, sensitivity, specificity, AUC, F1 score, and Brier score. Results Overall CVD prevalence was 5%. Hypertension emerged as the strongest predictor across all models, followed by alcohol-related harm. Tree-based models outperformed regression approaches and conventional clinical scores, with XGBoost achieving the highest discrimination (AUC=0.769), balanced accuracy (0.699), and calibration (Brier score=0.195). Predicted risk trajectories were smoother and more clinically plausible than Framingham or WHO/ISH scores, particularly across age, sex, and hypertension status. LASSO and Random Forest performed moderately, while conventional risk scores showed poor discrimination and marked miscalibration. Conclusion Machine-learning approaches provide accurate, context-specific cardiovascular risk prediction in African populations. By highlighting modifiable risk factors such as hypertension and alcohol-related harm, these models support targeted interventions aligned with WHO PEN, HEARTS, and SBIRT strategies. The African CVD Risk Prediction Tool translates complex data into actionable insights, offering a scalable platform for prevention-focused, equitable cardiovascular care across diverse African settings.

Ähnliche Arbeiten