A Comparative Machine Learning Analysis of Neural Networks and XGBoost for Predicting Maternal Mortality in Maiduguri, Nigeria

Abstract Maternal mortality continues to pose a major public health challenge in low resource and conflict affected settings. This study builds and compares two machine learning classifiers, an Artificial Neural Network (ANN) and Extreme Gradient Boosting (XGBoost), for predicting maternal mortality outcomes using 158 maternal health records collected in Maiduguri, Nigeria (118 survivals and 40 deaths). During data preparation, the Severe Morbidity (SM) variable was found to behave as a deterministic proxy for the outcome and was removed to avoid data leakage, together with a Health Facility Type variable that showed no variation across the sample. Model performance was assessed through repeated stratified 5-fold cross-validation (20 repetitions, 100 folds in total), using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC). XGBoost attained a mean cross-validated accuracy of 0.980 (precision 1.000, recall 0.920, F1-score 0.953, AUC 1.000), while the ANN reached a mean accuracy of 0.999 (precision 0.997, recall 1.000, F1-score 0.998, AUC 1.000); both exceeded a logistic regression baseline (accuracy 0.902, AUC 0.958). Once the leaked variable was removed, eclampsia, mode of delivery, parity, and gravidity emerged as the strongest predictors of maternal mortality. Although both models classified the available data very well, the limited sample size relative to the number of predictors means the results are best read as evidence of a strong fit to this particular dataset rather than proof of generalizable performance, and external validation is recommended before any clinical application. The findings nonetheless suggest that XGBoost and ANN hold real promise as maternal-mortality risk-stratification tools in resource-constrained settings, pending further validation.