预测孟加拉国五岁以下儿童死亡率的生存分析
Predicting under-five mortality in Bangladesh through comparative survival analysis using traditional, machine learning, and deep learning models
研究通过传统、机器学习和深度学习模型比较分析孟加拉国五岁以下儿童的死亡率,使用传统Cox比例风险模型、Lasso Cox、Ridge Cox、Elastic Net Cox、随机生存森林和梯度提升生存模型。研究结果显示,深度学习模型在不同随访时间点的AUC值较高,表明其在预测儿童死亡风险方面具有优势。
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Author information
Authors and Affiliations
Bangladesh Institute of Governance and Management, Dhaka, 1207, Bangladesh
Ruhul Amin, Warda Ruheen Bristi, A. S. M. Rafad Asgar & Mahadee Al Mobin
Authors
- Ruhul Amin
- Warda Ruheen Bristi
- A. S. M. Rafad Asgar
- Mahadee Al Mobin
Corresponding author
Correspondence to Mahadee Al Mobin.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix additional information
Appendix additional information
List of variables
See Tables 8, 9, 10, 11, Figs. 7, 8, 9, 10.
Hyperparameter search space and optimal model
Feature importance of the remaining models
Time-dependent AUC values
Time dependent ROC curve of DeepHit model
ROC curve at multiple time points of the DeepHit.
Risk and survival curve
Risk-stratified Kaplan–Meier survival curves at 12 month prediction horizons for the evaluated survival models. Children in the independent test dataset were classified into approximately equal-sized low-, intermediate-, and high-risk groups using rank-based tertiles of predicted cumulative mortality risk at 12 months. P-values are from overall three-group log-rank tests.
Risk-stratified Kaplan–Meier survival curves at 24 month prediction horizons for the evaluated survival models. Children in the independent test dataset were classified into approximately equal-sized low-, intermediate-, and high-risk groups using rank-based tertiles of predicted cumulative mortality risk at 24 months. P-values are from overall three-group log-rank tests.
Risk-stratified Kaplan–Meier survival curves at 36-month prediction horizons for the evaluated survival models. Children in the independent test dataset were classified into approximately equal-sized low-, intermediate-, and high-risk groups using rank-based tertiles of predicted cumulative mortality risk at 36 months. P-values are from overall three-group log-rank tests.
Methodological description
Model details
- 1.
Traditional Survival Models Standard Cox Proportional Hazards (PH) Model The Cox PH model is a widely used semi-parametric method for analyzing time-to-event data, allowing prediction of covariate effects without specifying the baseline hazard. In this study, it was employed as the classical statistical approach for modeling under-five mortality in Bangladesh. The hazard function is defined as:
$$\begin{aligned} h(t \mid X) = h_0(t) \exp \left( \sum _{i=1}^{p} \beta _i X_i\right) \end{aligned}$$
(3)
where \(h_0(t)\) is the baseline hazard and \(X_i\) are time-independent covariates as illustrated in (3). This model quantifies how socio-demographic factors influence the risk of child death over time37.
- 2.
Machine Learning-Based Survival Models
Lasso Cox Proportional Hazards Model The Lasso Cox model extends the Cox PH model by incorporating L1 regularization, enabling variable selection and shrinkage, especially beneficial in high-dimensional data. The model estimates:
$$\begin{aligned} \hat{\beta } = \arg \min _{\beta } \left\{ -\ell (\beta ) + \lambda \sum _{j=1}^{p} |\beta _j| \right\} \end{aligned}$$
(4)
where \(\ell (\beta )\) is the log-partial likelihood and \(\lambda \ge 0\) is the regularization parameter controlling penalty strength as illustrated in (4)34. Larger values of \(\lambda\) drive more coefficients to zero, enhancing model sparsity and interpretability.
Ridge Cox Model The Ridge Cox model applies L2-norm regularization to mitigate multicollinearity and overfitting. It solves:
$$\begin{aligned} \hat{\beta } = \arg \min _{\beta } \left\{ -\ell (\beta ) + \lambda \Vert \beta \Vert _2^2 \right\} \end{aligned}$$
(5)
Unlike Lasso, Ridge regression (as illustrated in (5)) retains all variables by shrinking coefficients instead of eliminating them34.
Elastic Net Cox Model To address limitations of Lasso and Ridge individually, Elastic Net regularization (as illustrated in (6)) combines both penalties. The penalized objective is:
$$\begin{aligned} \mathscr {L}(\beta ) = -\ell (\beta ) + \lambda \left[ \alpha \Vert \beta \Vert _1 + \frac{1 - \alpha }{2} \Vert \beta \Vert _2^2 \right] \end{aligned}$$
(6)
where \(\alpha \in [0,1]\) controls the trade-off between L1 (Lasso) and L2 (Ridge) penalties. This model is particularly useful when the number of predictors exceeds the number of observations45.
Random Survival Forest (RSF) RSF is a nonparametric ensemble method designed for right-censored data62. It extends the random forest algorithm by using survival trees to model the cumulative hazard function (CHF). For a terminal node h, the Nelson–Aalen estimator (as illustrated in (7)) is:
$$\begin{aligned} \hat{H}_{l,h}(t) = \sum _{t_{l,h} \le t} \frac{d_{l,h}}{Y_{l,h}} \end{aligned}$$
(7)
where \(d_{l,h}\) is the number of deaths and \(Y_{l,h}\) is the number at risk. The ensemble CHF (as illustrated in (8)) for individual i is:
$$\begin{aligned} \hat{e}(t \mid x_i) = \frac{1}{B} \sum _{b=1}^{B} \hat{H}^{(b)}(t \mid x_i) \end{aligned}$$
(8)
RSF does not assume proportional hazards and captures nonlinearities and interactions in high-dimensional data. prediction performance was evaluated using Harrell’s C-index, with error defined as \(1 - \text {C-index}\)33.
Gradient Boosting Survival Model (GBM-Cox) This model applies boosting techniques to survival analysis by optimizing a Cox-based loss (as illustrated in (9)):
$$\begin{aligned} \Phi (y, F) = -\sum _{i=1}^{n} \delta _i \left[ F(x_i) - \log \left( \sum _{j: t_j \ge t_i} e^{F(x_j)} \right) \right] \end{aligned}$$
(9)
The model updates iteratively:
$$\begin{aligned} F_m(x) = F_{m-1}(x) + \nu \rho _m f_m(x) \end{aligned}$$
(10)
where \(f_m(x)\) is the new base learner, \(\nu\) is the learning rate, and \(\rho _m\) is the optimal step size as illustrated in (10)68.
- 3.
Deep Learning Survival Models
DeepSurv Model DeepSurv generalizes the Cox PH model using deep neural networks to model the log-risk function \(\hat{h}_\theta (x)\)55. The loss function (as illustrated in (11)) is a penalized negative log-partial likelihood:
$$\begin{aligned} L(\theta ) = -\frac{1}{N_E} \sum _{i:E_i=1} \left[ \hat{h}_\theta (x_i) - \log \left( \sum _{j \in R(T_i)} e^{\hat{h}_\theta (x_j)} \right) \right] + \lambda \Vert \theta \Vert _2^2 \end{aligned}$$
(11)
This model allows learning complex nonlinear interactions in covariates, improving predictive power in high-dimensional, heterogeneous data. In this study, DeepSurv was implemented as a fully connected feed-forward neural network. Hyperparameters, including the number of hidden layers, neurons per layer, dropout rate, learning rate, batch size, optimizer settings, and early stopping criteria, were optimized through five-fold cross-validation. The final architecture and optimized hyperparameters are presented in Table 9.
DeepHit Model DeepHit is a neural framework for modeling the joint distribution of time-to-event and event type under competing risks60. It predicts (as illustrated in (12)):
$$\begin{aligned} \hat{y}^{(k,s)} = P(T = s, \text {event} = k \mid x) \end{aligned}$$
(12)
The cumulative incidence function (CIF) (as illustrated in (13)) is estimated as:
$$\begin{aligned} \hat{F}^k(t \mid x) = \sum _{s=1}^{t} \hat{y}^{(k,s)} \end{aligned}$$
(13)
DeepHit minimizes a composite loss:
$$\begin{aligned} L_{\text {Total}} = L_1 + L_2 \end{aligned}$$
(14)
where \(L_1\) is a log-likelihood loss accounting for censoring, and \(L_2\) is a ranking loss encouraging proper risk ordering as illustrated in (14). This model is particularly well-suited for data with non-proportional hazards and intricate dependencies. DeepHit was implemented using a multi-layer neural network for discrete-time survival prediction. Hyperparameters including network architecture, learning rate, batch size, dropout, alpha, sigma, and the number of discrete time intervals were optimized using the same cross-validation strategy as DeepSurv. The final architecture and selected hyperparameters are summarized in Table 9.
Model comparison metrics
Concordance Index (C-index) The Concordance Index (C-index) evaluates how well a survival model ranks individuals according to their predicted risk. It measures the proportion of all usable subject pairs where the predictions and outcomes are concordant. In practical terms, the C-index tests whether the model assigns a higher predicted risk to the individual who dies earlier—as one would expect in a well-performing survival model. For Cox-based models, the C-index is defined as:
$$\begin{aligned} \hat{c} = \frac{1}{\text {num}} \sum _{i:\, \delta _i = 1} \sum _{j:\, y_i < y_j} \mathbb {I} \left[ X_i \hat{\beta }> X_j \hat{\beta } \right] \end{aligned}$$
(15)
where \(\delta _i = 1\) indicates that the event (e.g., death) was observed for individual i, and \(\mathbb {I}[\cdot ]\) is the indicator function as illustrated in (15). The denominator, num, represents the total number of comparable pairs. In this study, we used the C-index to evaluate the ability of each model to correctly rank children by their predicted risk of dying before age five33,34.
Integrated Brier Score (IBS) IBS assesses the Model discrimination of a model’s predicted survival probabilities over time by calculating the mean squared error between predicted and actual survival status. It adjusts for right-censoring using inverse probability weighting. IBS (as illustrated in (16)) is defined as:
$$\begin{aligned} \text {IBS} = \frac{1}{t_{\text {max}}} \int _0^{t_{\text {max}}} \text {BS}(t) \, dt \end{aligned}$$
(16)
where the time-dependent Brier Score \(\text {BS}(t)\) (as illustrated in (17)) is given by:
$$\begin{aligned} \text {BS}(t) = \frac{1}{N} \sum _{i=1}^{N} \left[ \hat{y}_i(t) - y_i(t) \right] ^2 \end{aligned}$$
(17)
Here, \(\hat{y}_i(t)\) represents the predicted survival probability for individual i at time t, and \(y_i(t)\) denotes the actual survival status. The calculation incorporates the predicted survival function \(S(t \mid X_i)\), the observed event time \(T_i\), and the censoring distribution G(t). IBS values range from 0 to 1, with lower scores indicating superior predictive Model discrimination. In practice, an IBS value below 0.25 is considered indicative of a useful model34,69,70.
Time dependent area under the ROC curve (AUC) The area under the receiver operating characteristic (ROC) curve is a widely used measure to evaluate how well predicted risk scores discriminate between cases and controls in binary outcomes. In survival analysis, the time-to-event outcome is observed during follow-up and varies over time; therefore, prediction performance is evaluated using time-dependent ROC curves and the corresponding area under the curve, AUC(t)51–53. At a given time point t, cumulative cases are defined as individuals who experience the event by t\((T \le t)\), while dynamic controls are those who remain event-free beyond t\((T> t)\). The corresponding time-dependent sensitivity and specificity are defined as:
$$\begin{aligned} & Se^{C}(t, c) = \Pr (M> c \mid T \le t) \end{aligned}$$
(18)
$$\begin{aligned} & Sp^{D}(t, c) = \Pr (M \le c \mid T> t) \end{aligned}$$
(19)
where M represents the predicted risk score and c is a threshold. The time-dependent ROC curve is obtained by plotting \(Se^{C}(t, c)\) against \(1 - Sp^{D}(t, c)\) across all possible thresholds. The cumulative/dynamic time-dependent AUC is defined as:
$$\begin{aligned} AUC(t) = \Pr \big (M_i> M_j \mid T_i \le t,\; T_j> t\big ) \end{aligned}$$
(20)
which quantifies the probability that, at time t, a randomly selected individual who experiences the event before t has a higher predicted risk than a randomly selected individual who survives beyond t71. To address right-censoring, AUC(t) was estimated using inverse probability of censoring weighting (IPCW) at multiple follow-up time points.
Time-dependent Calibration Curve Calibration was assessed to evaluate the agreement between predicted and observed risks over time. While calibration-in-the-large reflects the overall agreement between average predicted and observed risks, it is also important to assess calibration across the full range of predictions. Time-dependent calibration curves were constructed at multiple clinically relevant time points72,73. To avoid arbitrary grouping of individuals, calibration curves were estimated using a pseudo-value-based approach, where observed risks were derived from the Kaplan–Meier estimator and smoothed as a function of predicted risks using a non-parametric method. This allows calibration to be evaluated continuously across the full spectrum of predicted probabilities74. Calibration was further summarized using the calibration intercept and slope. The intercept reflects calibration-in-the-large, where a value of 0 indicates no systematic over- or underestimation of predicted risk. The slope assesses agreement across the range of predictions, where a value of 1 indicates perfect calibration. A slope less than 1 suggests overfitting (overestimation of high predicted risks and underestimation of low predicted risks), whereas a slope greater than 1 indicates underfitting75. Calibration curves, along with intercept and slope, were evaluated at multiple follow-up times (e.g., 1 year, 2nd Year and 3rd year) to assess the temporal stability of model predictions. Models with calibration curves closer to the 45-degree reference line and intercept and slope values closer to 0 and 1, respectively, were considered better calibrated.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Amin, R., Bristi, W.R., Asgar, A.S.M.R. et al. Predicting under-five mortality in Bangladesh through comparative survival analysis using traditional, machine learning, and deep learning models. Sci Rep (2026). https://doi.org/10.1038/s41598-026-80198-6
Received: 05 August 2025
Accepted: 05 November 2026
Published: 18 November 2026
DOI: https://doi.org/10.1038/s41598-026-80198-6
Keywords
来源:Nature 健康护理主题 · nature.com