Development and Evaluation of Machine Learning Models for Predicting Infant Mortality in Malawi Using 2024 Malawi Demographic and Health Survey Data

Abstract Infant mortality remains an important public health concern in Malawi despite substantial progress in child survival over recent decades. The 2024 Malawi Demographic and Health Survey reported an infant mortality rate of 35 deaths per 1,000 live births, with neonatal deaths accounting for a large proportion of deaths during infancy. This study developed and evaluated machine learning models for predicting infant mortality in Malawi using the 2024 Malawi Demographic and Health Survey data. A quantitative secondary data analysis design was used. The analytical sample included 50,261 child records with a valid infant survival outcome. Infant death was defined as death before the first birthday and coded as a binary outcome. Descriptive analysis, missing data assessment, chi-square tests, Cramer’s V and multicollinearity assessment were conducted before model development. Logistic Regression was used as the baseline model, while Random Forest and XGBoost were developed as machine learning alternatives. Models were evaluated using accuracy, precision, sensitivity, specificity, F1-score, ROC-AUC, PR-AUC and Brier score. Class weighting and SMOTE sensitivity analysis were applied to address class imbalance. The analysis showed that infant death was a rare outcome, affecting 2,187 children, representing 4.35% of the unweighted sample and 4.72% of the weighted sample. Short gestation, birth interval, parity, maternal age at birth and multiple birth status were the most influential predictors. The improved main class-weighted XGBoost model was selected as the primary predictive model because it provided the best balance between predictive performance, data completeness and practical usefulness among the low-missingness models. The improved extended class-weighted XGBoost model achieved the highest overall PR-AUC, but it was treated as a sensitivity model because it included predictors with very high missingness. The study concludes that machine learning can add value to infant mortality prediction in Malawi, although the improvement over Logistic Regression was modest. The findings highlight the importance of balancing predictive performance with interpretability, data quality and public health usefulness. The model should not be used as a stand-alone decision-making tool without external validation and further calibration, but it provides useful evidence on infant mortality risk patterns in Malawi.