Neonatal mortality prediction under extreme class imbalance: calibrated supervised models and exploratory phenotype analysis in a population-based Iranian registry

Neonatal mortality remains a leading contributor to under-five deaths, yet early, individualized risk prediction is hampered by the rarity of the outcome and by limited attention to calibration and interpretability in prior models. We developed and stress-tested machine learning (ML) models for neonatal mortality prediction in a large, population-based cohort, integrating probability calibration, explainable AI, and exploratory unsupervised phenotyping. The development cohort of 61,517 neonates (131 deaths; class imbalance 1:469) was used to train L2-regularized logistic regression (LR) and LightGBM models on 60 routinely available perinatal features. performance was assessed by discrimination (AUROC, AUPRC), calibration, and decision-curve analysis, while SHAP and four unsupervised methods (UMAP, K-means, Isolation Forest, Apriori) characterized the underlying risk structure. Locked models then underwent a one-time case-only external stress test on 604 deceased neonates from a separate center and later period. LR achieved an AUROC of 0.955 (LightGBM, 0.944); at the optimal threshold (τ*=0.004) it captured 87.8% of deaths at 97.3% specificity, and positive net clinical benefit. In the external stress test, LR flagged 69.5% of deaths. Exploratory unsupervised analyses, fitted without outcome labels and evaluated post hoc against mortality, described a single high-risk phenotype; a K-means cluster comprising 6.7% of neonates contained 89.3% of all deaths. In this large, heterogeneous cohort, well-regularized ML models provided calibrated, interpretable, and early neonatal mortality risk stratification. The case-only design of the external cohort permits assessment of sensitivity only; external discrimination and calibration remain to be established, and prospective evaluation is warranted before clinical deployment.