Predictive Models of Neonatal and Postneonatal Mortality Based on Machine Learning in Honduras

Abstract

Child mortality remains a critical public health indicator in low- and middle-income countries, and risk stratification using machine learning represents a promising approach to target targeted interventions. This study aimed to develop and validate predictive models for neonatal (0–27 days) and postneonatal (28–364 days) mortality in Honduras, and to identify their main determinants using interpretability methods. A retrospective cohort study was conducted with data from ENDESA/MICS 2019, comparing three classification models (logistic regression, Random Forest and Gradient Boosting). Performance was assessed with AUC-ROC, AUC-PR, and Brier score, and interpretability was addressed using SHAP values. Among 9,579 live births, neonatal mortality was 14.2 per 1,000 and postneonatal mortality was 8.6 per 1,000. Gradient Boosting showed the best calibration (Brier=0.029 for neonatal; 0.019 for postneonatal). The most influential determinants of neonatal mortality were parity, wealth quintile and access to improved water; for post-neonatal mortality, the wealth quintile, parity and rural residence stood out. In conclusion, machine learning models interpretable using SHAP allow the identification of differential risk profiles for neonatal and postneonatal mortality, supporting the prioritization of maternal and child health interventions.