Predictors of Mortality among Neonates in Lusaka, Zambia: A Comparative Analysis of Machine Learning and Traditional Survival Analysis Techniques

Tshepiso Mokoena^1,2*Moses Mukosha³M Zunza²I Maposa²

• ¹Private, Pretoria, South Africa
• ²Stellenbosch University, Stellenbosch, Western Cape, South Africa
• ³Department of Pharmacy, School of Health Sciences, University of Zambia, Lusaka, Zambia

Neonatal mortality remains a critical global health issue, with 2.3 million deaths in 2022. Sub-Saharan Africa bears 57% of under 5 deaths despite only 30% of global births, with Zambia ranking fourth highest in terms of neonatal mortality among neighbouring countries despite progress. While traditional survival analysis has identified neonatal mortality risk factors, the use of machine learning for prediction remains underexplored. This study aimed to determine factors associated with neonatal mortality as well as compare the predictive performance of traditional survival analysis and machine learning models in predicting mortality among neonates in Lusaka, Zambia (January 2018–Septem-ber 2019). Demographic and clinical data from 1018 neonates were analysed using seven models: Weibull, Lasso, Ridge, Elastic Net (regularized cox), Random Survival Forest, DeepSurv Neural Networks and Gradient Boosting Machine. Model performance was assessed using nested cross validation with hyperparameter tuning in the inner folds and unbiased performance estimation in the outer folds. Evaluation metrics included the concordance index , time dependent area under the curve at 7, 14, and 28 days, brier scores, and calibration plots. Kaplan-Meier plots illustrated survival probabilities over time. Of 1,018 neonates, 757 (74.3%) died. Hypoxic ischemic encephalopathy (TR = 0.71, 95% CI: 0.63–0.81) was strongly associated with reduced survival, while higher birth weight was protective (TR = 1.88, 95% CI: 1.60–2.20). Sepsis demonstrated a paradoxical association with longer survival (TR = 1.16, 95% CI: 1.04–1.30), which persisted after sensitivity analyses. Among predictive models, the random survival forest achieved the highest discrimination (C index = 0.731; AUC = 0.759, 0.665, and 0.728 at 7, 14, and 28 days, respectively; brier = 0.164–0.254), outperforming Weibull (C-index = 0.622) and penalized Cox mod-els (0.620). Elastic Net demonstrated the most stable calibration across time points, while DeepSurv underperformed (C-index = 0.553). Feature importance analysis from random survival forest identified birth weight as the dominant predictor, followed by sex, sepsis, and necrotizing enterocolitis. While the traditional Weibull model remains valuable for interpretability, the integration of machine learning approaches offers enhanced predictive accuracy. Hybrid modelling strategies may improve early risk identification and inform neonatal care in resource-limited settings.