Answers are grounded in this paper and cite their sources.
Under-five mortality (U5M) remains a persistent public health challenge in Bangladesh, where 36.6 deaths per 1,000 live births separate the country from the SDG 3.2.1 target of 25. Standard survival models have shaped our understanding of this burden for decades, but their rigid structural assumptions limit how much of the underlying predicted risk architecture they can recover. Using data on 13,119 children from the 2022 Bangladesh Demographic and Health Survey (BDHS), this study systematically compares eight survival models across three methodological paradigms: traditional Cox proportional hazards and its penalized variants (Lasso, Ridge, Elastic Net), ensemble methods (Random Survival Forest (RSF) and Gradient Boosting Survival Analysis (GBSA)), and deep learning architectures (DeepSurv and DeepHit). A two-stage feature selection procedure: univariate Cox screening followed by Random Survival Forest permutation importance identified 15 predictors for model development. Performance was evaluated across discrimination, probabilistic accuracy, time-dependent AUC, and calibration, with stability confirmed across six k-fold configurations ( \(k = 5\) -30) via Friedman’s test ( \(p = 0.223\) ), while external validation was performed using an independent Nepal Demographic and Health Survey cohort processed with an identical analytical pipeline. DeepHit achieved the strongest joint performance: concordance index (C-index) of 0.977, Integrated Brier Score (IBS) of 0.016, mean time-dependent Area Under the Curve (AUC) of 0.987, and calibration slopes of 0.99–1.04 across three follow-up years, the only model in this comparison to approach ideal calibration. RSF led on C-index (0.978) but showed systematically inflated calibration slopes (1.42–1.45), while Cox-family models were competitive on discrimination but severely miscalibrated. Pairwise Wilcoxon signed-rank tests further demonstrated that the performance differences between the best-performing models and conventional survival approaches were statistically significant ( \(p < 0.05\) ) with large effect size. In the Nepal cohort, DeepHit maintained its performance without retraining, with a C-index of 0.9663 and a mean AUC of 0.9811. The relative rankings of the other models also carried over. Number of living children, total children ever born, and antenatal care intensity consistently dominated feature importance, locating the strongest mortality signal in maternal reproductive history rather than socioeconomic status alone.
No takes yet. Share an insight, caveat, or question.
Amin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: