Perinatal Mortality Prediction and Risk Factor Identification Using Machine Learning on Recent Sub-Saharan African DHS Data Affiliations

Abstract Background Perinatal mortality stillbirths after 28 weeks of gestation and early neonatal deaths within seven days of birth persists as a major public health challenge in Sub-Saharan Africa. Despite global declines in neonatal mortality, progress remains uneven in low-resource settings. Limited use of advanced analytics, especially machine learning, has hindered predictive accuracy and insights into complex risk factors. Methods This cross-sectional study analyzed pooled Demographic and Health Survey data from 19 Sub-Saharan African countries (2016–2023), including a weighted sample of 802,470 women aged 15–49. Seven supervised machine learning algorithms (Random Forest, Decision Tree, Logistic Regression Voting Classifier, XGBoost, Naïve Bayes, Stacking, LightGBM) predicted perinatal mortality. Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance. Performance was assessed via accuracy, precision, recall, F1-score, and AUC-ROC; SHAP analysis interpreted feature importance. Results Perinatal mortality prevalence was 5.2%, highest in Côte d’Ivoire (9.1%), Tanzania (7.5%), and Ghana (7.0%), lowest in Gabon (2.5%), Zambia (3.4%), and Madagascar (3.6%). Random Forest excelled with 81.7% accuracy, 80% precision, 84% recall, 82% F1-score, and 0.91 AUC-ROC. SHAP analysis highlighted maternal age, education, and parity (number of living children) as top predictors. Higher maternal education offered strong protection; first and high-order births elevated risk. Conclusion Random Forest effectively predicted perinatal mortality. Low maternal education was the primary risk factor, with parity showing a U-shaped pattern. Prioritizing female education, antenatal care quality, and high-risk parity groups could substantially reduce regional perinatal deaths.