Use of explainable machine learning in risk classification of Cesarean section delivery: A cross-sectional analysis of Demographic and Health Surveys from ten Sub-Saharan African countries (2016–2024)

Cesarean section (CS) delivery is an important surgical intervention for reducing maternal and neonatal morbidity and mortality when medically indicated; however, substantial inequalities in its utilization persist across Sub-Saharan Africa (SSA). This study evaluated the performance of explainable machine learning (ML) algorithms in classifying cesarean section delivery patterns using Demographic and Health Survey (DHS) data from ten SSA countries collected between 2016 and 2024. A weighted sample of 388,015 women aged 15–49 years was included, among whom 7,369 (1.82%) had undergone cesarean section delivery. Data preprocessing included handling missing values, feature encoding, normalization, and balancing the minority class using the Synthetic Minority Over-sampling Technique (SMOTE). Multiple ML classifiers, including LightGBM, XGBoost, Random Forest, Decision Tree, Logistic Regression, Naïve Bayes, K-Nearest Neighbor, AdaBoost, and Artificial Neural Networks, were trained and evaluated using repeated 10-fold cross-validation. Model performance was assessed using accuracy, recall, F1-score, and area under the receiver operating characteristic curve (AUC). LightGBM achieved the best classification performance with an AUC of 0.89 (95% CI: 0.887–0.893), accuracy of 85% (95% CI: 83.7–86.3%), and recall of 91% (95% CI: 89.8–92.2%), significantly outperforming logistic regression. SHapley Additive exPlanations (SHAP) identified maternal age, child size at birth, maternal education, wealth index, antenatal care visits, and media exposure as the most influential features associated with cesarean section classification outcomes. The findings demonstrate that explainable ML approaches, particularly LightGBM, improve classification performance compared with conventional regression models when applied to large population-based DHS datasets. However, because the study used cross-sectional survey data without external validation or detailed clinical predictors, the findings should be interpreted as associative classification patterns rather than tools for prospective clinical prediction.