Machine Learning Matches Classic Models at Predicting Child Mortality in Zambia

Machine learning models designed to predict which children are most at risk of dying before their fifth birthday performed no better than traditional statistical survival models in a new analysis of data from Zambia, according to a study based on the 2024 Zambia Demographic and Health Survey. The work, led by Kassaye Getaneh Arge, compared four competing modeling frameworks on a cohort of more than 9,000 live births and found their predictive accuracy was statistically indistinguishable, even as a more sophisticated biological model revealed new insight into why some children face elevated risk.

The researchers began with 35,610 live births recorded in the survey and, after applying exclusion criteria related to maternal age and data completeness, arrived at a final analytic sample of 9,362 singleton births. Within that sample, 1,236 children died before age five, corresponding to a mortality rate of 13.2 percent over the follow-up period. The team then pitted the Cox Proportional Hazards model, a mainstay of survival analysis, against three increasingly flexible alternatives: Lasso Cox regression, Random Survival Forests, and Gradient Boosting Machines. Each framework was evaluated with nested cross-validation, using outer folds grouped at the mother level to prevent information from siblings in the same family from leaking between training and test data, and performance was scored with time-dependent Concordance Indices and Area Under the Curve metrics across a 48-month follow-up window.

The results were strikingly even. The C-index, which measures how well a model ranks children by risk, ranged only from 0.745 for the Cox model to 0.750 for the Gradient Boosting Machine, while Integrated Brier Scores, which capture both discrimination and calibration of survival predictions, fell between 0.095 and 0.096 across all four approaches. Formal pairwise comparisons using DeLong’s test found no statistically significant differences, with all p-values exceeding 0.05. In other words, the added complexity of the machine learning methods bought no measurable gain in predictive discrimination or calibration for this dataset.

But the study was not solely a horse race between algorithms. Alongside the predictive comparison, the researchers fitted a Bayesian Joint Model that links a longitudinal process, the length of birth intervals, directly to the hazard of a child’s death. This model revealed a significant association between the birth interval process and mortality risk, with an association parameter of 0.035 (95 percent confidence interval: 0.023 to 0.048, p < 0.001). The authors interpret this finding as support for the maternal depletion hypothesis, the idea that closely spaced pregnancies deplete a mother’s physiological reserves in ways that compromise the survival chances of subsequent children.

Variable importance analysis across the models pointed to birth interval and maternal age as the two most influential predictors of child survival, followed by whether the mother had previously lost a child. The stratified Cox model, used after Schoenfeld residual tests flagged time-varying effects, quantified some of these relationships directly. Mothers with secondary education had a hazard ratio of 0.426 (95 percent CI: 0.318 to 0.570) compared with the reference group, indicating substantially lower mortality risk, while a history of previous child deaths was associated with a hazard ratio of 1.648 (95 percent CI: 1.529 to 1.777).

The study’s authors conclude that advanced joint modeling techniques offer deeper biological insight into the dynamics of under-five mortality, even though machine learning and traditional survival models delivered equivalent predictive performance in this analysis. They suggest the findings support targeted public health interventions centered on birth spacing, maternal age, and enhanced monitoring of high-risk families as critical strategies for reducing under-five mortality in Zambia, a country that continues to record some of the highest child mortality rates in sub-Saharan Africa.

Subject of Research: Predictive survival modeling of under-five child mortality in Zambia using demographic survey data

Article Title: Predicting under-five child mortality in Zambia: Machine learning versus parametric survival models

Article References: Arge, K. G. (2026). Predicting under-five child mortality in Zambia: Machine learning versus parametric survival models. PLOS One, 21(10), e0361197. https://doi.org/10.1371/journal.pone.0361197

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361197

Keywords: under-five mortality, Zambia, machine learning, survival analysis, Cox proportional hazards, random survival forests, gradient boosting, Bayesian joint model, birth interval, maternal depletion, demographic and health survey, risk prediction