Early Detection of Cardiovascular Disease Using Machine Learning: A Sex-Stratified Fairness-Aware Study for Low-Resource Clinical Settings in Sub-Saharan Africa

Cardiovascular disease (CVD) is the leading cause of mortality worldwide and the fastest-growing non-communicable disease burden across sub-Saharan Africa, where late clinical diagnosis continues to drive preventable deaths.

This study presents a comprehensive, fairness-aware comparative evaluation of eight machine learning algorithms for the early prediction of cardiovascular disease, using a combined primary dataset of 1,100 patient records validated against established clinical
feature distributions from the cardiovascular disease literature.

The eight evaluated models are Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, XGBoost, LightGBM, and a Multi-Layer Perceptron. A rigorous preprocessing pipeline incorporating median imputation, SMOTE
oversampling, and Standard Scaler normalization was applied.

SVM achieved the highest AUC-ROC of 0.9270, while Random Forest achieved the highest accuracy of 86.36 percent and recall of 0.8515, and SVM recorded the strongest cross-validated accuracy at 87.68 percent. SHAP analysis identified thalassemia status, ST
depression, maximum heart rate, exercise-induced angina, and number of major vessels as the five dominant predictive features.

A three-dimensional fairness analysis stratified by sex, age group, and cholesterol level revealed that female patients exhibit lower recall than male patients (80.0 percent vs 85.92 percent), empirically confirming the well-documented clinical phenomenon of female CVD underdiagnosis and raising critical concerns for equitable AI deployment in diverse clinical populations.

All findings directly inform Nexora's AI-driven preventive health intelligence platform for underserved communities across Ghana and West Africa, extending the scientific program initiated in the author's prior work on Type 2 diabetes prediction
(DOI: https://doi.org/10.5281/zenodo.20348652).