Neonatal mortality remains a major challenge in resource-limited settings, where delayed recognition of high-risk cases and inconsistent clinical decisions hinder timely and targeted interventions, which are essential for reducing preventable deaths. In response, this study developed and evaluated Machine-Learning (ML) models to predict neonatal death using maternal and neonatal features collected both before and after delivery. To this end, guided by the CRISP-DM data-mining framework, we analyzed a dataset of 7,214 births (5,000 survivors and 2,214 deaths) from 2021–2022, derived from routinely collected records in the Iranian Maternal and Neonatal (IMaN) registry. As a result, among the data-mining models—Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Machine—trained with imbalance-sensitive techniques, XGBoost achieved the best performance (ROC-AUC = 0.967, PR-AUC = 0.940). Feature importance analysis identified gestational age (importance = 0.179) and birth weight (0.109) as the dominant predictors, followed by nervous system malformations (0.035), musculoskeletal malformations (0.033), high-risk delivery indicators (0.032), and other congenital malformations (0.031). The contribution of this study is in twofolds, first, these findings demonstrate that accurate, real-time prediction of neonatal mortality is achievable. Seconds, beyomd a prognostic tool, the final model can serve as an operational lever within neonatal services; when embedded into a clinical Decision Support System (DSS), it can enhance early risk detection, improve triage accuracy, facilitate timely NICU preparedness, and strengthen overall process reliability and system performance in resource-limited care settings.