A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction

Victor Wandera Lumumba; Denis Kariuki Muriithi; Amos Kipkorir Langat; Maureen Ambasa Wanyama; Elizabeth Wambui Njoroge; John Kamwele Mutinda; Olivia Waka; Edson Mwebesa

doi:doi:10.11648/j.cbb.20251302.11

Research Article |

| Peer-Reviewed

A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction

Victor Wandera Lumumba^*

, Denis Kariuki Muriithi

, Amos Kipkorir Langat

, Maureen Ambasa Wanyama

, Elizabeth Wambui Njoroge

, John Kamwele Mutinda

, Olivia Waka

, Edson Mwebesa

Published in Computational Biology and Bioinformatics (Volume 13, Issue 2)

Received: 2 September 2025 Accepted: 12 September 2025 Published: 27 October 2025

Views: Downloads:

Download PDF

Share This Article

Twitter
Linked In
Facebook

Abstract

Neonatal mortality remains a critical public health challenge, particularly in low- and middle-income countries (LMICs), where limited healthcare resources and fragmented follow-up systems hinder timely interventions. Accurate prediction of neonatal death is essential for risk stratification, resource allocation, and improving survival outcomes. While traditional survival analysis methods like the Kaplan-Meier estimator and Cox proportional hazards (Cox PH) model are widely used, they face limitations in handling non-linear relationships, high-dimensional data, and violations of proportional hazards assumptions. Random Survival Forests (RSF), a machine learning approach, offers potential advantages but lacks sufficient comparative evaluation in neonatal mortality prediction, especially within LMIC contexts. This study aimed to comparatively evaluate the performance of Kaplan-Meier, Cox PH, and RSF models in predicting neonatal mortality using a synthetic dataset reflecting perinatal epidemiology in Kenya. The research addresses a significant and direct methodological comparisons across these models in neonatal populations, particularly under real-world conditions involving censoring, missing data, and non-proportional hazards. We assessed discrimination (C-index, time-dependent AUC), calibration (Integrated Brier Score, CRPS), and clinical interpretability. The dataset included 2,000 neonates with 17 covariates including but not limited to gestational age, birth weight, maternal health, and socioeconomic status. Results showed that RSF outperformed both Kaplan-Meier and Cox PH in discrimination (C-index: 0.875 vs. 0.868) and maintained strong calibration, particularly at 28 days. Variable importance measures identified gestational age, birth weight, and maternal health score as top predictors. SHAP values enhanced interpretability of RSF outputs. The Cox model provided clinically intuitive hazard ratios but was less flexible in capturing interactions. The study concluded that RSF offers superior predictive accuracy for neonatal mortality and should be integrated into risk prediction tools, especially in data-rich settings. Policy makers should support adoption of advanced analytics in perinatal care systems, while maintaining traditional models for inferential clarity. Combining both paradigms can optimize neonatal survival strategies.

Published in	Computational Biology and Bioinformatics (Volume 13, Issue 2)
DOI	10.11648/j.cbb.20251302.11
Page(s)	42-59
Creative Commons	This is an Open Access article, distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution and reproduction in any medium or format, provided the original work is properly cited.
Copyright	Copyright © The Author(s), 2025. Published by Science Publishing Group

Keywords

Neonatal Mortality, Survival Analysis, Kaplan-Meier Estimator, Cox Proportional Hazards Model, Random Survival Forests

1. Introduction

Neonatal mortality remains a critical public health challenge globally, particularly in low- and middle-income countries (LMICs), where over 98% of the estimated 2.3 million neonatal deaths occurred in 2023

[1]

. Despite significant progress in reducing child mortality over the past three decades, neonatal deaths now account for nearly half of all under-five mortality, underscoring the urgency of refining predictive models to inform timely clinical interventions and health policy

[3, 2]

. Accurate prediction of neonatal mortality is essential for identifying high-risk infants, optimizing resource allocation, and improving neonatal care outcomes

[4]

. Survival analysis methods, including the Kaplan-Meier estimator, Cox proportional hazards (Cox PH) model, and more recently, machine learning-based approaches such as Random Survival Forests (RSF), have been widely applied in clinical epidemiology to model time-to-event data. However, their comparative performance in neonatal mortality prediction remains underexplored, particularly in resource-constrained settings where data complexity, censoring, and non-proportional hazards may limit traditional statistical methods.

The Kaplan-Meier estimator, introduced in 1958, remains a foundational non-parametric method for estimating survival functions from time-to-event data

[5]

. It is widely used in clinical research for its simplicity and interpretability, particularly in generating survival curves and comparing survival distributions across groups using log-rank tests. In neonatal studies, Kaplan-Meier analysis has been employed to estimate survival probabilities in preterm infants

[6]

, evaluate the impact of birth weight on mortality

[7]

and assess outcomes following neonatal intensive care unit (NICU) admission

[8]

. However, its limitation lies in its inability to adjust for multiple covariates simultaneously, rendering it inadequate for multivariable risk prediction in complex clinical scenarios.

To address these limitations, the Cox proportional hazards model has become the standard semi-parametric approach in survival analysis. The Cox PH model allows for the simultaneous assessment of multiple predictors on survival time while estimating hazard ratios (HRs)

[9]

. Its application in neonatal research includes identifying risk factors such as gestational age, birth asphyxia, sepsis, and maternal health indicators

[10]

. The model’s assumption of proportional hazards, that the hazard ratio between groups remains constant over time is a key strength but also a potential limitation. Violations of this assumption, common in neonatal data due to time-varying risk patterns (early vs. late neonatal mortality), can lead to biased estimates

[11]

. Moreover, the Cox model assumes linearity in continuous predictors and may not capture complex interactions without explicit modeling, which limits its predictive accuracy in high-dimensional datasets

[12]

In recent years, machine learning (ML) methods have emerged as powerful alternatives for survival analysis, particularly in handling non-linear relationships, high-dimensional data, and violations of traditional modeling assumptions. Among these, Random Survival Forests (RSF), an extension of Breiman’s random forests to survival data, have gained traction in biomedical research

[13]

. RSF is a non-parametric ensemble method that builds multiple decision trees using bootstrap samples and random feature selection, aggregating predictions to estimate survival probabilities and cumulative hazard functions. Unlike the Cox model, RSF does not assume proportional hazards or linearity and can automatically detect interactions among variables

[14]

. In neonatal research, RSF has been applied to predict mortality in preterm infants using electronic health records (EHRs), demonstrating superior discrimination and calibration compared to traditional models

[15]

Despite the growing use of RSF, direct comparative evaluations of Kaplan-Meier, Cox PH, and RSF in neonatal mortality prediction remain limited. Most studies have focused on adult populations or specific subgroups such as very low birth weight infants

[16]

with few addressing the full spectrum of neonatal mortality in diverse epidemiological contexts. Furthermore, the performance of these models in LMICs, where data quality, missingness, and variable distributions differ significantly from high-income settings, has not been systematically assessed

[17]

. Recent evidence suggests that ML models, including RSF, may outperform traditional methods in predicting neonatal outcomes when trained on large, heterogeneous datasets

[18, 19]

, but concerns remain regarding interpretability, overfitting, and generalizability

[20]

While Kaplan-Meier curves provide intuitive visualizations of survival probabilities, and Cox PH models yield hazard ratios that are familiar to clinicians, RSF outputs such as variable importance measures and survival ensembles are less straightforward to interpret. Efforts to enhance transparency in ML models, such as SHAP (SHapley Additive exPlanations) values and partial dependence plots, are being increasingly integrated into survival analysis

[21, 22]

, but their adoption in neonatal research is still nascent.

Moreover, the dynamic nature of neonatal risk factors such as evolving sepsis markers, respiratory support needs, and feeding tolerance poses challenges for static models. Time-dependent covariates can be incorporated into Cox models and RSF, but their implementation requires careful data structuring and may increase computational complexity

[23]

. In contrast, Kaplan-Meier analysis is inherently limited in this regard, as it does not accommodate time-varying predictors.

Recent advances in data linkage and digital health systems have enabled richer datasets for neonatal research, including longitudinal EHRs, maternal health records, and socioeconomic indicators

[24]

. These data sources offer unprecedented opportunities to improve predictive accuracy but also demand robust analytical frameworks capable of handling complexity. Comparative studies in other clinical domains such as cardiovascular disease and cancer have shown that ensemble ML models often outperform traditional methods in predictive performance, especially with large sample sizes and high-dimensional covariates

[25]

. However, whether these advantages translate to neonatal populations, where event rates are lower and clinical pathways more heterogeneous, remains an open question.

This study presents a comparative evaluation of the Kaplan-Meier estimator, Cox proportional hazards model, and Random Survival Forests for predicting neonatal mortality using a synthetically generated dataset designed to emulate the demographic, clinical, and socioeconomic characteristics of maternal-neonatal populations in Kenya. The synthetic data were simulated based on published distributions and associations observed in nationally representative surveys, including the Kenya Demographic and Health Survey (KDHS 2022) and the Nigeria Demographic and Health Survey (NDHS 2018), as well as multicenter perinatal studies

[26]

. All variables were constructed to reflect biologically plausible relationships and known risk factors for neonatal mortality.

We assessed model performance in terms of discrimination (C-index), calibration (calibration plots, Hosmer-Lemeshow-type tests), and clinical interpretability, while addressing methodological challenges such as censoring, missing data, and violations of proportional hazards. By benchmarking these models against real-world clinical outcomes, we aim to inform best practices in neonatal survival analysis and support the integration of advanced analytics into perinatal care systems.

2. Materials and Methods

2.1. Research Design

This study employed a methodological comparative design to evaluate the performance of three survival analysis techniques—Kaplan-Meier estimation, Cox proportional hazards regression, and Random Survival Forests—in predicting neonatal mortality. The primary objective was to assess model discrimination, calibration, and predictive accuracy using a synthetic dataset structured to reflect real-world perinatal epidemiology in low-resource settings. The time-to-event outcome was neonatal mortality (death within 28 days of life), with follow-up time measured in days. Models were compared under conditions of proportional and non-proportional hazards, with emphasis on clinical interpretability and robustness to data complexity.

2.2. Data Collection

A synthetic cohort of 2,000 mother-newborn dyads was generated using Monte Carlo simulation techniques based on joint distributions derived from empirical studies. Variables including gestational age, birth weight, maternal age, and socioeconomic status were simulated using multivariate normal and categorical distributions calibrated to match prevalence and correlation structures reported in KDHS (2022) and other Demographic and Health Surveys (ICF, 2023). Neonatal mortality was simulated with a baseline hazard informed by Kenyan mortality rates

[1]

and time-to-event was generated using a Weibull distribution with covariate-dependent hazard functions reflecting established risk factors (e.g., low birth weight, lack of skilled birth attendance). Censoring was introduced to reflect loss to follow-up (12%), consistent with field studies.

2.3. Data Analysis

Survival analysis was conducted using three complementary approaches. First, the Kaplan-Meier estimator was used to generate non-parametric survival curves, with log-rank tests comparing survival across categorical groups (delivery method, multiple birth). Second, a semi-parametric Cox proportional hazards model was fitted to estimate hazard ratios (HRs) and assess the independent effect of covariates on neonatal mortality. Proportional hazards assumptions were tested using Schoenfeld residuals and Grambsch-Therneau tests. Third, a Random Survival Forest (RSF) model was implemented using the `randomForestSRC` package in R, with 1,000 trees and 5-fold cross-validation to optimize prediction accuracy. Model performance was evaluated using the concordance index (C-index), integrated Brier score (IBS), and calibration plots. Variable importance was assessed via minimal depth and permutation-based measures. All analyses accounted for right-censoring and were stratified by early (0-6 days) and late (7-28 days) neonatal periods to reflect distinct etiological pathways. Statistical significance was set at α = 0.05.

2.4. Models Fitting and Estimation

To enable a comprehensive comparative evaluation, we present the formal statistical and computational foundations of the three survival analysis methods employed: the Kaplan-Meier estimator, the Cox proportional hazards (Cox PH) model, and Random Survival Forests (RSF). Each method is described in terms of its underlying assumptions, likelihood structure, estimation procedure, and implementation with respect to the neonatal mortality outcome and the covariate space defined by the synthetic dataset.

2.4.1. Kaplan-Meier Estimator

Let

T_{i}

denote the time-to-event (in days) for the

i

-th newborn, where the event is neonatal death within 28 days of life, and

C_{i}

be the corresponding censoring time. The observed time is

t_{i} = \min (T_{i}, C_{i})

, and

δ_{i} = I (T_{i} \leq C_{i})

is the event indicator. The survival function is defined as

S (t) = P (T > t)

The Kaplan-Meier (KM) estimator

[5]

provides a non-parametric estimate of

S (t)

\hat{S} (t) = \prod_{t_{j} \leq t} (1 - \frac{d_{j}}{n_{j}})

(1)

where

t_{j}

are the distinct event times,

d_{j}

is the number of deaths at

t_{j}

, and

n_{j}

is the number of infants at risk just before

t_{j}

In this study, KM curves were stratified by key categorical variables such as Delivery Method (vaginal vs. cesarean), Multiple Birth (singleton vs. twin/multiple), NICU Admission (yes/no), and Skilled Birth Attendant (present/absent). The log-rank test was used to assess differences in survival distributions:

χ^{2} = \frac{{(\sum (d_{j} - e_{j}))}^{2}}{\sum V_{j}}

(2)

where

e_{j}

and

V_{j}

are the expected events and variance under the null hypothesis of equal survival across groups.

Due to its non-parametric nature, the KM estimator does not adjust for confounders such as Gestational Age weeks or Socioeconomic Status, limiting its use to univariate analysis.

2.4.2. Cox Proportional Hazards Model

The Cox PH model

[9]

extends survival analysis to multivariable settings by modeling the hazard function:

h (t | x_{i}) = h_{0} (t) \exp (\sum_{k = 1}^{p} β_{k} x_{ik})

(3)

where: -

h (t | x_{i})

is the hazard for infant

i

at time

t

, -

h_{0} (t)

is the baseline hazard (non-parametric), -

x_{i} = (x_{i 1}, \dots, x_{ip})

is the vector of covariates, -

β_{k}

are regression coefficients.

In this study, the covariate vector includes:

- Continuous: Maternal Age, Birth Weight (kg), Gestational Age weeks, Maternal Health Score, Apgar Score (5min), Distance to Health Facility (km)

- Binary: Smoking During Pregnancy, Alcohol Use During Pregnancy, NICU Admission, Multiple Birth

- Categorical: Delivery Method (reference: vaginal), Socioeconomic Status (low, medium, high), Maternal Education Level, Type of Health Facility

The partial likelihood for parameter estimation is:

L (β) = \prod_{i : δ_{i} = 1} \frac{\exp (x_{i}^{⊤} β)}{\sum_{j \in R (t_{i})} \exp (x_{j}^{⊤} β)}

(4)

where

R (t_{i})

is the risk set at time

t_{i}

Maximization of the log-partial likelihood yields coefficient estimates

\hat{β}

, from which hazard ratios (HRs) are computed as

{HR}_{k} = \exp ({\hat{β}}_{k})

. The proportional hazards assumption was tested using scaled Schoenfeld residuals

[27]

with time-transformed covariates in a regression of residuals on time. For variables violating this assumption (Apgar Score (1min) with time-varying effect), stratified Cox models or time-dependent covariates were considered:

h (t | x_{i}) = h_{0} (t) \exp (β^{⊤} x_{i} + γt \cdot z_{i})

(5)

where

z_{i}

is a time-interacting covariate.

Model selection used backward elimination based on Akaike Information Criterion (AIC), and goodness-of-fit was assessed via scaled residuals and martingale residuals.

2.4.3. Random Survival Forests (RSF)

Random Survival Forests

[13]

is a non-parametric ensemble machine learning method that generalizes Breiman’s random forests to right-censored survival data. RSF does not assume proportional hazards or linearity and can capture complex interactions (between Maternal Nutrition Score and Gestational Age weeks).

Each tree in the forest is grown on a bootstrap sample of the data. At each node, a random subset of

m

variables (out of

p

) is selected for splitting. The optimal split is chosen to maximize survival difference between daughter nodes using a log-rank splitting rule:

L (ψ) = \frac{\sum_{j \in ψ_{L}} (d_{j} - e_{j})}{\sqrt{\sum_{j \in ψ_{L}} V_{j}}}

(6)

where

ψ_{L}

is the left daughter node, and

d_{j}, e_{j}, V_{j}

are observed, expected, and variance terms from the log-rank test.

The survival function for an individual

i

is estimated by averaging over all trees:

{\hat{S}}_{i} (t) = \frac{1}{M} \sum_{m = 1}^{M} {\hat{S}}_{i, m} (t)

(7)

where

{\hat{S}}_{i, m} (t)

is the Nelson-Aalen-based survival estimate in the terminal node of tree

m

Key parameters: - Number of trees:

M = 1000

- Number of variables randomly selected at each split:

m = ⌊ \sqrt{p} ⌋ = 5

(with

p = 25

) - Node size: 3 (minimum number of events per terminal node)

Variable importance (VIMP) is computed via permutation: the increase in prediction error when a variable is randomly shuffled. High VIMP indicates strong predictive contribution. Minimal depth analysis identifies variables closest to tree roots, indicating early decision importance.

Prediction performance is evaluated using the out-of-bag (OOB) C-index:

C-index = P ({\hat{T}}_{i} < {\hat{T}}_{j} ∣ T_{i} < T_{j})

(8)

and the integrated Brier score (IBS):

IBS = \int_{0}^{τ} BS (t) dt

(9)

Where;

BS (t) = \frac{1}{n} \sum_{i = 1}^{n} {(\hat{S} (t | x_{i}) - I (T_{i} > t))}^{2} \cdot \frac{δ_{i}}{\hat{G} (T_{i})}

(10)

and

\hat{G}

is the Kaplan-Meier estimate of censoring survival.

All models were implemented in R 4.5.1 using the following packages: - survival for KM and Cox PH

[28]

, - randomForestSRC for RSF

[14]

. Hyperparameter tuning for RSF used 5-fold cross-validation on the OOB error. Missing data (introduced at 5% MCAR) were handled using multivariate imputation by chained equations (MICE) prior to modeling.

3. Results and Discussion

3.1. Descriptive Statistics and Cohort Characteristics

Table 1. Baseline Continuous Neonatal Characteristics by Neonatal Mortality Status.

Characteristic	Overall	95% CI	(Survived =0)	95% CI	(Died =1)	95% CI	p-value²
Characteristic	N = 2,000¹	95% CI	N = 1,400¹	95% CI	N = 600¹	95% CI	p-value²
Maternal Age	30 (23, 37)	30, 30	29 (22, 37)	29, 30	32 (25, 39)	31, 32	<0.001
Prenatal Visits	6 (4, 8)	5.9, 6.1	6 (4, 8)	6.1, 6.4	5 (4, 7)	5.3, 5.6	<0.001
Birth Weight kg	2.99 (2.61, 3.39)	3.0, 3.0	3.03 (2.66, 3.44)	3.0, 3.1	2.89 (2.48, 3.30)	2.8, 2.9	<0.001
Gestational Age weeks	38 (36, 40)	38, 38	38 (36, 40)	38, 39	37 (35, 39)	37, 37	<0.001
Maternal Health Score	5.35 (3.20, 7.85)	5.3, 5.6	5.96 (3.66, 8.25)	5.7, 6.0	4.22 (2.53, 6.08)	4.3, 4.7	<0.001
Maternal Nutrition Score	5.47 (3.28, 7.83)	5.4, 5.6	5.92 (3.58, 8.11)	5.6, 5.9	4.63 (2.69, 7.08)	4.7, 5.1	<0.001
Apgar Score (5min)	4.90 (2.50, 7.30)	4.8, 5.0	5.90 (3.30, 8.10)	5.5, 5.8	2.90 (1.30, 4.85)	3.1, 3.4	<0.001
Distance to Health Facility km	7 (3, 14)	9.4, 10	6 (3, 12)	8.4, 9.3	9 (4, 17)	11, 13	<0.001

Abbreviation: CI = Confidence Interval

¹ Median (Q1, Q3)

² Welch Two Sample t-test

Descriptive analysis (Table 1) revealed significant differences in maternal and neonatal characteristics between survivors and non-survivors of neonatal mortality (N = 2,000). Specifically, birth weight was significantly lower among non-survivors (Mdn = 2.89 kg, Q1-Q3 = 2.48-3.30) compared to survivors (Mdn = 3.03 kg, Q1-Q3 = 2.66-3.44), p <.001. This aligns with well-established evidence that low birth weight is a major risk factor for neonatal death, with infants weighing <2.5 kg facing up to a 20-fold increased risk

[2, 1]

Similarly, gestational age was lower in non-survivors (Mdn = 37 weeks) than survivors (Mdn = 38 weeks), p <.001, underscoring the critical role of prematurity in neonatal mortality. Globally, complications of preterm birth account for approximately 35% of neonatal deaths

[29]

The 5-minute Apgar score, a key indicator of neonatal transition was markedly lower among non-survivors (Mdn = 2.90) compared to survivors (Mdn = 5.90), p <.001. Apgar scores ≤3 at 5 minutes is strongly associated with perinatal asphyxia and mortality

[30]

. Maternal factors also differed significantly. Mothers of non-surviving infants had lower maternal health (Mdn = 4.22 vs. 5.96, p <.001) and nutrition scores (Mdn = 4.63 vs. 5.92, p <.001), consistent with evidence linking maternal undernutrition and comorbidities to adverse neonatal outcomes

[31]

. Additionally, distance to health facility was greater for non-survivors (Mdn = 9 km) than survivors (Mdn = 6 km), p <.001, highlighting the role of health system access. In low-resource settings, delays in reaching care contribute significantly to neonatal mortality

[32]

Table 2. Baseline Categorical Neonatal Characteristics by Neonatal Mortality Status.

Characteristic	Overall	95% CI	(Survived =0)	95% CI	(Died =1)	95% CI	p-value²
Characteristic	N = 2,000¹	95% CI	N = 1,400¹	95% CI	N = 600¹	95% CI	p-value²
Socioeconomic Status							<0.001
Low	608 (30%)	28%, 32%	407 (29%)	27%, 32%	201 (34%)	30%, 37%
Medium	620 (31%)	29%, 33%	408 (29%)	27%, 32%	212 (35%)	32%, 39%
High	493 (25%)	23%, 27%	360 (26%)	23%, 28%	133 (22%)	19%, 26%
Very High	279 (14%)	12%, 16%	225 (16%)	14%, 18%	54 (9.0%)	6.9%, 12%
Delivery Method							0.035
Vaginal	1,380 (69%)	67%, 71%	986 (70%)	68%, 73%	394 (66%)	62%, 69%
Cesarean	620 (31%)	29%, 33%	414 (30%)	27%, 32%	206 (34%)	31%, 38%
Multiple Birth							0.2
Singleton	1,900 (95%)	94%, 96%	1,336 (95%)	94%, 96%	564 (94%)	92%, 96%
Multiple	100 (5.0%)	4.1%, 6.1%	64 (4.6%)	3.6%, 5.8%	36 (6.0%)	4.3%, 8.3%
Skilled Birth Attendant							<0.001
Absent	294 (15%)	13%, 16%	180 (13%)	11%, 15%	114 (19%)	16%, 22%
Present	1,706 (85%)	84%, 87%	1,220 (87%)	85%, 89%	486 (81%)	78%, 84%
Smoking During Pregnancy							<0.001
Did not Smoke	1,806 (90%)	89%, 92%	1,292 (92%)	91%, 94%	514 (86%)	83%, 88%
Smoked	194 (9.7%)	8.5%, 11%	108 (7.7%)	6.4%, 9.3%	86 (14%)	12%, 17%
Alcohol Use During Pregnancy							0.012
Did not Use	1,840 (92%)	91%, 93%	1,302 (93%)	92%, 94%	538 (90%)	87%, 92%
Use	160 (8.0%)	6.9%, 9.3%	98 (7.0%)	5.7%, 8.5%	62 (10%)	8.1%, 13%
NICU Admission							<0.001
Not admitted	1,612 (81%)	79%, 82%	1,171 (84%)	82%, 86%	441 (74%)	70%, 77%
Admitted	388 (19%)	18%, 21%	229 (16%)	14%, 18%	159 (27%)	23%, 30%
Type of Health Facility							<0.001
Public	1,240 (62%)	60%, 64%	872 (62%)	60%, 65%	368 (61%)	57%, 65%
Private	475 (24%)	22%, 26%	353 (25%)	23%, 28%	122 (20%)	17%, 24%
Rural Clinic	285 (14%)	13%, 16%	175 (13%)	11%, 14%	110 (18%)	15%, 22%
Maternal Chronic Conditions							<0.001
No Chronic Condition	1,569 (78%)	77%, 80%	1,130 (81%)	79%, 83%	439 (73%)	69%, 77%
Have Chronic Condition	431 (22%)	20%, 23%	270 (19%)	17%, 21%	161 (27%)	23%, 31%
Maternal Education Level							0.068
None	304 (15%)	14%, 17%	200 (14%)	13%, 16%	104 (17%)	14%, 21%
Primary	698 (35%)	33%, 37%	477 (34%)	32%, 37%	221 (37%)	33%, 41%
Secondary	682 (34%)	32%, 36%	488 (35%)	32%, 37%	194 (32%)	29%, 36%
Tertiary	316 (16%)	14%, 17%	235 (17%)	15%, 19%	81 (14%)	11%, 17%
Environmental Exposure							0.001
Not Exposed	1,695 (85%)	83%, 86%	1,210 (86%)	84%, 88%	485 (81%)	77%, 84%
Exposed	305 (15%)	14%, 17%	190 (14%)	12%, 16%	115 (19%)	16%, 23%
History of Pregnancy Complications							0.006
No complication	1,593 (80%)	78%, 81%	1,138 (81%)	79%, 83%	455 (76%)	72%, 79%
Had complication	407 (20%)	19%, 22%	262 (19%)	17%, 21%	145 (24%)	21%, 28%

Abbreviation: CI = Confidence Interval

¹ n (%)

² Pearson’s Chi-squared tests

Non-surviving neonates were more likely to be born to mothers with lower socioeconomic status (SES), with 34% of non-survivors from low SES versus 29% of survivors (p <.001) (Table 2). A significantly higher proportion of non-survivors were born via cesarean (34% vs. 30%, p =.035), likely reflecting complications during delivery. The absence of a skilled birth attendant was more common among non-survivors (19% vs. 13%, p <.001), underscoring the protective role of professional care during childbirth

[33]

. NICU admission was markedly higher among non-survivors (27% vs. 16%, p <.001), indicating severity of condition, though timely access remains a barrier

[34]

. Maternal chronic conditions were present in 27% of non-survivor cases versus 19% of survivors (p <.001), aligning with evidence linking hypertension and diabetes to adverse neonatal outcomes

[35]

. Environmental exposure was more prevalent among non-survivors (19% vs. 14%, p =.001), suggesting environmental risk factors such as air pollution or unsafe water contribute to mortality

[36]

. The study found that a higher proportion of non-survivors had a history of pregnancy complications (24% vs. 19%, p =.006), indicating prior obstetric risk. These findings highlight the interplay of clinical, socioeconomic, and health system factors in neonatal mortality. Strengthening antenatal care, expanding skilled birth attendance, and improving access to emergency neonatal services are critical for reducing mortality, particularly in vulnerable populations.

3.2. Neonatal Mortality Risk Over Time

Table 3. Neonatal Mortality Rates Overall and by Early (0-6 days) and Late (7-28 days).

Period	Deaths	Infants at Risk	Mortality Rate (%)	95% CI
Early Neonatal (0-6 days)	257	2000	12.85	[11.4, -14.4]
Late Neonatal (7-28 days)	223	1743	12.79	[11.3, -14.5]
Post-Neonatal (>28 days)	120	1520	7.89	[6.6, -9.4]
Overall	600	2000	30	[28, -32.1]

The results in Table 3 shows that the neonatal mortality in this cohort was substantial, with an overall rate of 30.0% (95% CI: 28.0-32.1%), reflecting a high-risk population consistent with settings experiencing weak health systems and limited access to perinatal care

[1]

. The early neonatal period (0-6 days) accounted for 257 deaths, yielding a mortality rate of 12.85% (95% CI: 11.4-14.4%). An additional 223 deaths occurred in the late neonatal period (7-28 days), resulting in a conditional mortality rate of 12.79% (95% CI: 11.3-14.5%) among infants who survived the first week. This near-parity in early and late neonatal mortality rates underscores that risk remains elevated throughout the neonatal period, challenging the assumption that most deaths are confined to the immediate postnatal phase

[2]

Beyond 28 days, 120 post-neonatal deaths occurred among 1,520 survivors, yielding a post-neonatal mortality rate of 7.89% (95% CI: 6.6-9.4%). This indicates that while the highest absolute risk is in the neonatal period, significant mortality persists into infancy, likely driven by infectious diseases, malnutrition, and inadequate follow-up care

[37]

. The findings highlight distinct epidemiological patterns: early deaths are often linked to intrapartum complications, preterm birth, and birth asphyxia, whereas late neonatal and post-neonatal deaths are more commonly associated with sepsis, poor feeding, and delayed care-seeking

[38]

. The high late neonatal mortality rate suggests that survival beyond the first week does not equate to safety, emphasizing the need for sustained postnatal monitoring and community-based interventions.

These results support the integration of extended postnatal care into maternal and child health programs, particularly in low-resource settings where follow-up is often fragmented. Accurate period-specific mortality estimates enable targeted public health strategies across the early life course.

3.3. Kaplan-Meier Survival Analysis

Download: Download full-size image

Figure 1. Overall Neonatal Survival Probability.

Figure 1 shows the Kaplan-Meier survival curve showing the overall neonatal survival probability over time. Initially, there's a steep decline in survival, with the probability dropping from 1.00 to around 0.75 within the first 14 days. After this period, the curve flattens significantly, indicating a much slower rate of neonatal deaths. The shaded pink area represents the 95% confidence interval, which widens considerably over time as the number of neonates at risk decreases.

Download: Download full-size image

Figure 2. Neonates Survival Probabilities Determined by Various Categorical Predictors.

Figure 2 displays Kaplan-Meier survival curves for neonates, categorized by five key factors. Survival is consistently lower among neonates who were admitted to the NICU (p<0.0001), born via cesarean section (p=0.021), and those born to mothers with low socioeconomic status (p<0.0001). Additionally, survival is significantly lower when a skilled birth attendant is absent (p=0.00051). The survival difference between singleton and multiple births is not statistically significant (p=0.078). Each graph highlights the impact of a specific factor on neonatal survival probability over time.

Table 4. Log Rank Test for the Neonates Survival Probabilities.

Key Categorical Predictors	Groups Compared	N (per group)	Observed deaths	Expected deaths	Chi-square	df	p-value
NICU Admission	Not admitted vs Admitted	1612 / 388	441 / 159	490 / 110	26.9	1	<0.001
Delivery Method	Vaginal vs Cesarean	1380 / 620	394 / 206	420 / 180	5.3	1	0.02
Socioeconomic Status	Low / Medium / High / Very High	608 / 620 / 493 / 279	201 / 212 / 133 / 54	177.0 / 181.2 / 154.3 / 87.5	24.4	3	<0.001
Skilled Birth Attendant	Absent vs Present	294 / 1706	114 / 486	84.5 / 515.5	12.1	1	<0.001
Multiple Birth	Singleton vs Multiple	1900 / 100	564 / 36	573 / 27	3.1	1	0.08

Table 4 results indicates that the log-rank test revealed statistically significant differences in neonatal survival across key categorical predictors. Infants admitted to NICU experienced significantly lower mortality than non-admitted infants (χ² = 26.9, df = 1, p <.001), with observed deaths of 159 versus 441 despite a smaller sample size. Cesarean deliveries had higher observed mortality than vaginal births (χ² = 5.3, p =.02). Socioeconomic status significantly influenced survival (χ² = 24.4, df = 3, p <.001), with the lowest survival in low and medium groups. Absence of skilled birth attendance was associated with higher mortality (χ² = 12.1, p <.001). Multiple births showed a trend toward increased risk (χ² = 3.1, p =.08), though not statistically significant.

Download: Download full-size image

Figure 3. Survival Probabilities for Neonates Between 14 and 28 days of the Study.

Figure 3 above shows neonates survival probability for NICU admission between 14 and 28 days. The survival probability for neonates not admitted to the NICU is consistently and significantly higher throughout the 63-day period. This difference is statistically significant, as indicated by the p<0.0001 value. The curves diverge most notably within the first 28 days, after which the survival probability plateaus for both groups, although the gap remains. The table below the graph shows the decreasing number of neonates at risk over time for each group.

Table 5. Log Rank Test for Neonates NICU Admission Between 14 and 28 Days.

NICU Admission	Survival for 14 Days	Survival for 28 Days
NICU Admission = No	0.822	0.780
NICU Admission = Yes	0.763	0.678
Log-rank Test (NICU Admission)	Chi-sq = 26.9, p = 2e-07

From Table 5 above, it is evident that neonates admitted to NICU exhibited lower survival probabilities beyond 14 days compared to non-admitted infants, reflecting higher baseline illness severity. By day 14, survival was 82.2% (no admission) versus 76.3% (admission); by day 28, it was 78.0% and 67.8%, respectively. The log-rank test showed a statistically significant difference (χ² = 26.9, p = 2 × 10⁻⁷), indicating that NICU admission status strongly predicts neonatal survival. This suggests that while NICU care targets high-risk infants, timely access and quality of care remain critical for improving long-term survival outcomes in this vulnerable population.

3.4. Cox-Proportional Hazard Model

Table 6. Cox Proportional Hazards Regression Results.

Variable	Coefficient	Hazard Ratio (HR)	SE	Z-value	95% CI Lower	95% CI Upper	P-value
Maternal Age	0.0243	1.0246	0.0049	4.951	0.0147	0.034	7.00E-07
Prenatal Visits	-0.1317	0.8766	0.0177	-7.464	-0.1663	-0.0972	0.00E+00
Birth Weight (kg)	-0.6353	0.5298	0.0716	-8.869	-0.7757	-0.4949	0.00E+00
Gestational Age weeks	-0.139	0.8702	0.0139	-9.974	-0.1664	-0.1117	0.00E+00
Maternal Health Score	-0.1815	0.834	0.0172	-10.541	-0.2153	-0.1478	0.00E+00
Socioeconomic Status-Medium	-0.0213	0.979	0.1015	-0.21	-0.2201	0.1776	8.34E-01
Socioeconomic Status-High	-0.3145	0.7301	0.1148	-2.739	-0.5396	-0.0895	6.17E-03
Socioeconomic Status-Very High	-0.365	0.6942	0.1553	-2.35	-0.6694	-0.0605	1.88E-02
Delivery Method-Cesarean	0.3091	1.3622	0.0878	3.522	0.1371	0.4812	4.29E-04
Multiple Birth -Multiple	0.5766	1.78	0.1745	3.305	0.2347	0.9186	9.49E-04
Maternal Nutrition Score	-0.1119	0.8941	0.0162	-6.917	-0.1437	-0.0802	0.00E+00
Maternal Chronic Conditions-Yes	0.3564	1.4281	0.0943	3.779	0.1715	0.5412	1.58E-04
Skilled Birth Attendant-Present	-0.3009	0.7402	0.1059	-2.841	-0.5085	-0.0933	4.50E-03
Maternal Education Level-Primary	-0.0827	0.9207	0.1213	-0.682	-0.3204	0.155	4.95E-01
Maternal Education Level-Secondary	-0.3091	0.7341	0.1244	-2.485	-0.5529	-0.0653	1.30E-02
Maternal Education Level-Tertiary	-0.4786	0.6197	0.1513	-3.162	-0.7752	-0.1819	1.57E-03
Smoking During Pregnancy-Yes	0.3681	1.445	0.1202	3.063	0.1325	0.6036	2.19E-03
Alcohol Use During Pregnancy-Yes	0.4027	1.4958	0.1378	2.922	0.1326	0.6728	3.48E-03
Apgar_Score_5min	-0.2743	0.7601	0.0164	-16.69	-0.3065	-0.2421	0.00E+00
NICU Admission-Admitted	0.5132	1.6707	0.094	5.458	0.3289	0.6976	0.00E+00
Distance to Health Facility (km)	0.0347	1.0354	0.004	8.783	0.027	0.0425	0.00E+00

The multivariable Cox proportional hazards regression model identified several significant predictors of neonatal mortality. After adjusting for covariates, lower birth weight (HR = 0.530 per kg, 95% CI: 0.448-0.625, p <.001) and shorter gestational age (HR = 0.870 per week, 95% CI: 0.850-0.891, p <.001) were strongly associated with increased mortality risk, consistent with established perinatal risk factors (Table 6). Poorer maternal health (HR = 0.834 per unit increase, 95% CI: 0.801-0.868, p <.001) and lower maternal nutrition score (HR = 0.894, 95% CI: 0.864-0.925, p <.001) were also independently associated with higher mortality, highlighting the intergenerational impact of maternal well-being.

Infants born via cesarean delivery had a 36.2% higher hazard of mortality (HR = 1.362, 95% CI: 1.147-1.618, p <.001), reflecting underlying complications rather than mode of delivery itself. Multiple births were associated with a 78.0% increased hazard (HR = 1.780, 95% CI: 1.263-2.508, p =.001), emphasizing the vulnerability of twins and higher-order multiples. Low Apgar score at 5 minutes was one of the strongest predictors (HR = 0.760 per point, 95% CI: 0.738-0.783, p <.001), indicating impaired neonatal transition. NICU admission was associated with a 67.1% higher hazard (HR = 1.671, 95% CI: 1.391-2.006, p <.001), likely due to confounding by indication (sicker infants are admitted).

Maternal factors such as chronic conditions (HR = 1.428, p <.001), smoking (HR = 1.445, p =.002), and alcohol use (HR = 1.496, p =.003) significantly increased mortality risk. Conversely, skilled birth attendance was protective (HR = 0.740, 95% CI: 0.608-0.899, p =.005). Greater distance to health facility increased risk (HR = 1.035 per km, 95% CI: 1.027-1.042, p <.001), underscoring access barriers.

Higher socioeconomic status (High: HR = 0.730; Very High: HR = 0.694) and maternal education (Secondary: HR = 0.734; Tertiary: HR = 0.620) were protective, with a clear dose-response relationship. Prenatal visits were also protective (HR = 0.877 per visit, p <.001). The model confirmed maternal age as a risk factor (HR = 1.025 per year, p <.001). All key covariates were adjusted simultaneously, demonstrating the independent effects of clinical, socioeconomic, and health system factors on neonatal survival.

Download: Download full-size image

Figure 4. Cox Proportional Hazard Ratios.

The forest plot (Figure 4) presents the Hazard Ratios (HR) from a Cox proportional hazards model, identifying factors associated with neonatal mortality. A vertical dashed line at an HR of 1.0 serves as the reference point. Factors with an HR greater than 1.0, and a confidence interval that does not cross this line, are associated with an increased risk of death. These include NICU admission (HR=1.57, p<0.001), smoking (HR=1.44, p=0.002), alcohol use (HR=1.50, p=0.002), multiple births (HR=1.26, p<0.001), and cesarean delivery (HR=1.35, p=0.021). Conversely, factors with an HR less than 1.0, such as tertiary education (HR=0.58, p=0.002) and higher birth weight (HR=0.55, p<0.001), are protective. The model's overall performance is indicated by a concordance index of 0.83.

3.5. Random Survival Forests (RSF)

The Random Survival Forest (RSF) model was trained on 2,000 infants with 600 neonatal deaths, using 1,000 trees and a terminal node size of 3 (Table 7). With 17 predictor variables and 4 randomly selected at each split, the model employed log-rank splitting and resampling without replacement. The out-of-bag (OOB) continuous ranked probability score (CRPS) was 42.46 (standardized: 0.168), indicating good predictive accuracy. The OOB error rate was 0.205, reflecting strong performance in capturing survival risk based on maternal, neonatal, and environmental factors.

Table 7. Summary of Random Survival Forest (RSF) Model Results.

Characteristic	Result
Sample size	2000
Number of deaths	600
Number of trees	1000
Forest terminal node size	3
Average number of terminal nodes	272.42
Variables tried at each split	4
Total number of variables	17
Resampling used to grow trees	Without replacement (swor)
Resample size used to grow trees	1264
Analysis	RSF
Family	Survival
Splitting rule	Log-rank (random)
Number of random split points	10
(OOB) Continuous Ranked Probability Score (CRPS)	42.46
(OOB) Standardized CRPS	0.168
(OOB) Requested performance error	0.205

Random Forest Variable Importance

Figure 5 shows the Variable Importance (VIMP) scores derived from a Random Survival Forest (RSF) model, ranking the predictors of neonatal survival. VIMP measures the decrease in prediction accuracy when a variable's values are randomly permuted, with higher scores indicating greater importance.

The plot reveals that the 5-minute Apgar score is, by far, the most important predictor, with a VIMP score exceeding 0.06. This is followed by Maternal Health Score and Gestational Age, both with VIMP scores above 0.015. Other significant predictors include Birth Weight and Distance to Health Facility. Factors like NICU Admission, Maternal Chronic Conditions, Socioeconomic Status, and Skilled Birth Attendant have a very low VIMP, indicating they contribute minimally to the model's predictive power. The least important variables are Alcohol Use During Pregnancy, Delivery Method, and Multiple Birth. This hierarchy of importance suggests that clinical and physiological factors at birth are the most influential determinants of neonatal survival in this model.

Download: Download full-size image

Figure 5. Variables Importance for Random Survival Forest.

3.6. KM Survival Curve Stratification

We derived individual risk scores from both the multivariable Cox proportional hazards model and the Random Survival Forest (RSF) model. For the Cox model, linear predictors were calculated and categorized into tertiles (low, medium, high risk). For the RSF model, subject-specific predicted survival probabilities at 14 days were transformed into risk estimates (1 - survival probability), which were likewise divided into tertiles. Kaplan-Meier curves stratified by risk tertiles demonstrated clear separation of survival trajectories for both approaches. In the Cox-based grouping, patients in the high-risk tertile experienced significantly lower survival compared with the medium- and low-risk groups (log-rank p < 0.001). Similarly, the RSF-based tertiles also showed strong discrimination between groups (log-rank p < 0.001). The RSF model displayed steeper separation between low- and high-risk groups, suggesting enhanced early discrimination of mortality risk compared to the Cox model. These findings confirm that both Cox regression and RSF models can meaningfully stratify risk in this population, with RSF potentially offering additional prognostic value.

Download: Download full-size image

Figure 6. Stratified KM Survival Curves for Cox and RSF.

Figure 6 shows the Kaplan-Meier curves comparing the survival probabilities of neonates stratified into three risk tertiles (low, medium, high) based on scores from a Cox model (left) and a Random Survival Forest (RSF) model (right). Both plots demonstrate excellent risk stratification, with the high-risk group showing significantly lower survival than the medium and low-risk groups ($p < 0.0001$ for both). The RSF model shows a steeper decline in the high-risk group, suggesting it provides better discrimination of early mortality risk compared to the Cox model. Both models successfully identified distinct survival trajectories, confirming their utility in predicting neonatal outcomes.

Table 8 shows the evaluation of the predictive performance of the Cox proportional hazards model and Random Survival Forest (RSF) using time-dependent area under the curve (AUC) at key postnatal intervals. Both models demonstrated excellent discrimination, with AUCs increasing over time, indicating improved predictive accuracy as event information accumulates. At 7 days, RSF outperformed Cox (AUC = 0.835 vs. 0.828), with a consistent advantage maintained at 14 days (0.852 vs. 0.845), 28 days (0.875 vs. 0.868), and 90 days (0.882 vs. 0.876). The superior AUCs of RSF suggest it better captures complex, nonlinear risk patterns and interactions, making it more effective for dynamic neonatal mortality prediction in high-dimensional settings.

Table 8. Area Under the Curve for Cox and RSF.

Time (days)	Cox AUC	RSF AUC
7	0.828	0.835
14	0.845	0.852
28	0.868	0.875
90	0.876	0.882

3.7. Comparison of Predictive Performance

Table 9. Comparison of Predictive Performance.

Metrics	Cox	RSF
Harrell C	0.825	0.953
Integrated Brier score (crps):	0.155	0.156
Reference (CRPS)	0.172	0.172

Table 9 presents a comparative evaluation of predictive performance between the Cox proportional hazards model and Random Survival Forest (RSF). The RSF model demonstrated superior discrimination with a Harrell’s C-index of 0.953, substantially higher than the Cox model’s 0.825, indicating better ability to distinguish high- from low-risk infants. The Integrated Brier Score (IBS), measuring overall prediction accuracy, was nearly identical between models (Cox: 0.155; RSF: 0.156), suggesting comparable calibration. However, both models showed improvement over the reference score (0.172), indicating predictive value beyond a null model. Despite its excellent discrimination, RSF’s slightly higher Brier score may reflect overfitting in high-dimensional regions or sensitivity to event timing. The Cox model, while less discriminatory, offers interpretable hazard ratios and assumes proportional hazards. Overall, RSF excels in capturing complex, nonlinear survival patterns, making it particularly suitable for high-dimensional neonatal risk prediction where traditional assumptions may not hold. These results highlight the trade-off between interpretability and predictive accuracy in survival modeling.

4. Conclusions

This study presents a comprehensive comparative evaluation of three survival analysis methods—Kaplan-Meier estimator, Cox proportional hazards (Cox PH) model, and Random Survival Forests (RSF)—in predicting neonatal mortality using a synthetic dataset designed to reflect the epidemiological and clinical complexity of maternal-neonatal populations in low-resource settings. The primary objective was to assess the performance of these models in terms of discrimination, calibration, and clinical interpretability, with the ultimate goal of informing best practices in neonatal risk prediction and guiding the integration of advanced analytics into perinatal care systems.

The results demonstrate that each model offers unique strengths and limitations. The Kaplan-Meier estimator provided intuitive, non-parametric survival curves and facilitated univariate comparisons across key risk factors such as NICU admission and delivery method, confirming their significant impact on neonatal survival. The Cox PH model extended this analysis by enabling multivariable inference, identifying critical predictors including low birth weight (HR = 0.530), preterm birth (HR = 0.870), absence of skilled birth attendance (HR = 0.740), and maternal chronic conditions (HR = 1.428). However, the model’s reliance on the proportional hazard’s assumption was challenged, particularly for time-varying effects, limiting its robustness in dynamic clinical environments.

In contrast, the Random Survival Forest (RSF) model demonstrated superior predictive performance, achieving a Harrell’s C-index of 0.953 and outperforming the Cox model (C-index: 0.825) in both discrimination and time-dependent AUC. RSF effectively captured non-linear relationships and high-order interactions without requiring parametric assumptions, as evidenced by variable importance and minimal depth analyses, which highlighted gestational age, Apgar score, and maternal health as top predictors. Despite its reduced interpretability compared to Cox, the use of SHAP values and VIMP enhanced transparency in feature contributions. From the study, the high predictive accuracy of RSF supports its adoption in risk stratification tools for early identification of high-risk neonates, particularly in data-rich settings. While traditional survival models remain essential for inference and transparency, machine learning approaches like RSF offer significant advantages in predictive accuracy and adaptability to complex data structures. Integrating both paradigms can optimize neonatal mortality prediction, ultimately supporting timely interventions and improving survival outcomes in vulnerable populations.

Abbreviations

KM	Kaplan-Meier
Cox PH	Cox Proportional Hazards
RSF	Random Survival Forest
HR	Hazard Ratio
CI	Confidence Interval
AUC	Area Under the Curve
C-index	Concordance Index
IBS	Integrated Brier Score
CRPS	Continuous Ranked Probability Score
NICU	Neonatal Intensive Care Unit
APGAR	Appearance, Pulse, Grimace, Activity, Respiration
OOB	Out-of-Bag

Acknowledgments

The authors gratefully acknowledge the Department of Physical Sciences and the Center for Data Analytics and Modelling at Chuka University for providing institutional support and computational resources. Besides, while the dataset was carefully constructed to reflect biologically plausible relationships and epidemiological patterns observed in the Kenya Demographic and Health Survey (KDHS 2022) and other nationally representative sources, authors acknowledged the limitations inherent in simulation-based research. The reliance on synthetic data limits direct clinical generalizability, Thus, future validation using real-world cohorts will be essential to confirm the external applicability of these findings.

Author Contributions

Victor Wandera Lumumba: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

Denis Kariuki Muriithi: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing

Amos Kipkorir Langat: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources

Maureen Ambasa Wanyama: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Resources, Validation, Visualization, Writing – original draft, Writing – review & editing

Elizabeth Wambui Njoroge: Conceptualization, Investigation, Methodology, Resources, Supervision

John Kamwele Mutinda: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision

Olivia Waka: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Resources, Validation, Visualization, Writing – original draft, Writing – review & editing

Edson Mwebesa: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision

Funding

This work is not supported by any external funding.

Data Availability Statement

The data is available from the corresponding author upon reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest.

References

[1]	Arora, A. (2024, March 12). Levels and trends in child mortality - UNICEF DATA. UNICEF DATA. https://data.unicef.org/resources/levels-and-trends-in-child-mortality-2023/
[2]	Liu, L., Oza, S., Hogan, D., Chu, Y., Perin, J., Zhu, J., Lawn, J. E., Cousens, S., Mathers, C., & Black, R. E. (2016). Global, regional, and national causes of under-5 mortality in 2000-15: an updated systematic analysis with implications for the Sustainable Development Goals. The Lancet, 388(10063), 3027-3035. https://doi.org/10.1016/s0140-6736(16)31593-8
[3]	WHO. (2024, March 14). Newborn mortality. https://www.who.int/news-room/fact-sheets/detail/newborn-mortality
[4]	Yan, T., Mullany, L. C., Subedi, S., Hazel, E. A., Khatry, S. K., Mohan, D., Zeger, S., Tielsch, J. M., LeClerq, S. C., & Katz, J. (2023). Risk factors for neonatal mortality: an observational cohort study in Sarlahi district of rural southern Nepal. BMJ Open, 13(9), e066931-e066931. https://doi.org/10.1136/bmjopen-2022-066931
[5]	Kaplan, E. L., & Meier, P. (1958). Nonparametric Estimation from Incomplete Observations. Journal of the American Statistical Association, 53(282), 457-481. https://doi.org/10.2307/2281868
[6]	Getaneh, T., Homaira, N., Kasaye, H., Jane, S., Chughtai, A. A., & Lui, K. (2025). Global inequities in the survival of extremely preterm infants: a systematic review and meta-analysis. BMC Pediatrics, 25(1). https://doi.org/10.1186/s12887-025-05933-w
[7]	Odongkara, B., Nankabirwa, V., Achora, V., Arach, A. A., Napyo, A., Musaba, M., Mukunya, D., Ndeezi, G., Tylleskär, T., & Tumwine, J. K. (2024). Low birthweight increased the risk of neonatal death twenty-folds in Northern Uganda: a community-based cohort study. MedRxiv (Cold Spring Harbor Laboratory). https://doi.org/10.1101/2024.04.25.24306373
[8]	Orsido, T. T., Asseffa, N. A., & Berheto, T. M. (2019). Predictors of Neonatal mortality in Neonatal intensive care unit at referral Hospital in Southern Ethiopia: a retrospective cohort study. BMC Pregnancy and Childbirth, 19(1). https://doi.org/10.1186/s12884-019-2227-5
[9]	Cox, D. R. (1972). Regression Models and Life-Tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187-202. https://doi.org/10.1111/j.2517-6161.1972.tb00899.x
[10]	Ezeh, O. K., Agho, K. E., Dibley, M. J., Hall, J., & Page, A. N. (2014). Determinants of neonatal mortality in Nigeria: evidence from the 2008 demographic and health survey. BMC Public Health, 14(1). https://doi.org/10.1186/1471-2458-14-521
[11]	Orsini, N. (2013). Review of Flexible Parametric Survival Analysis Using Stata: Beyond the Cox Model by Patrick Royston and Paul C. Lambert. The Stata Journal: Promoting Communications on Statistics and Stata, 13(1), 212-216. https://doi.org/10.1177/1536867x1301300115
[12]	Wang, P., Li, Y., & Reddy, C. K. (2019). Machine Learning for Survival Analysis. ACM Computing Surveys, 51(6), 1-36. https://doi.org/10.1145/3214306
[13]	Ishwaran, H., Kogalur, U. B., Blackstone, E. H., & Lauer, M. S. (2008). Random survival forests. The Annals of Applied Statistics, 2(3), 841-860. https://doi.org/10.1214/08-aoas169
[14]	Berkowitz, M., Altman, R. M., & Loughin, T. M. (2024). Random forests for survival data: which methods work best and under what conditions? The International Journal of Biostatistics. https://doi.org/10.1515/ijb-2023-0056
[15]	MASHRAFI, S. S. A., Tafakori, L., & Abdollahian, M. (2025). Predicting Early Neonatal Mortality using Machine Learning Models. https://doi.org/10.21203/rs.3.rs-7016070/v1
[16]	Chatziioannidis, I., Mitsiakos, G., Karagianni, P., Tsakalidis, C., & Dimopoulou, A. (2020). Predictors of early neonatal mortality in extremely low birth weight infants in a Neonatal Intensive Care Unit over a 10-year period. DOAJ (DOAJ: Directory of Open Access Journals). https://doi.org/10.7363/090115
[17]	Shukla, V., & Carlo, W. A. (2019). Technology-driven Neonatal Health Care in Low-resource Settings: Expectations and Reality. EClinicalMedicine, 12, 2-3. https://doi.org/10.1016/j.eclinm.2019.06.005
[18]	Mangold, C., Zoretic, S., Thallapureddy, K., Moreira, A., Chorath, K., & Moreira, A. (2021). Machine Learning Models for Predicting Neonatal Mortality: A Systematic Review. Neonatology, 118(4), 394-405. https://doi.org/10.1159/000516891
[19]	O’Sullivan, C., Hsiang-Te Tsai, D., Chang-Yen Wu, I., Boselli, E., Hughes, C., Padmanabhan, D., & Hsia, Y. (2023). Machine learning applications on neonatal sepsis treatment: a scoping review. 23(1). https://doi.org/10.1186/s12879-023-08409-3
[20]	Collins, G. S., Dhiman, P., Ma, J., Schlussel, M. M., Archer, L., Calster, B. V., Harrell, F. E., Martin, G. P., Moons, K. G. M., Smeden, M. van, Sperrin, M., Bullock, G. S., & Riley, R. D. (2024). Evaluation of clinical prediction models (part 1): from development to external validation. BMJ, 384, e074819. https://doi.org/10.1136/bmj-2023-074819
[21]	Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., & Lee, S.-I. (2020). From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2(1), 56-67. https://doi.org/10.1038/s42256-019-0138-9
[22]	Weissler, E. H., Naumann, T., Andersson, T., Ranganath, R., Elemento, O., Luo, Y., Freitag, D. F., Benoit, J., Hughes, M. C., Khan, F., Slater, P., Shameer, K., Roe, M., Hutchison, E., Kollins, S. H., Broedl, U., Meng, Z., Wong, J. L., Curtis, L., & Huang, E. (2021). The role of machine learning in clinical research: transforming the future of evidence generation. Trials, 22(1). https://doi.org/10.1186/s13063-021-05489-x
[23]	Hendry, D. J. (2013). Data generation for the Cox proportional hazards model with time-dependent covariates: a method for medical researchers. Statistics in Medicine, 33(3), 436-454. https://doi.org/10.1002/sim.5945
[24]	Carvajal-Aguirre, L., Amouzou, A., Mehra, V., Ziqi, M., Zaka, N., & Newby, H. (n.d.). Gap between contact and content in maternal and newborn care: An analysis of data from 20 countries in sub-Saharan Africa. Journal of Global Health, 7(2). https://doi.org/10.7189/jogh.07.020501
[25]	Suliman, A., Masud, M., Serhani, M. A., Abdullahi, A. S., & Oulhaj, A. (2024). Predictive performance of machine learning compared to statistical methods in time-to-event analysis of cardiovascular disease: a systematic review protocol. BMJ Open, 14(4), e082654-e082654. https://doi.org/10.1136/bmjopen-2023-082654
[26]	Feng, J., Lee, J., Vesoulis, Z. A., & Li, F. (2021). Predicting mortality risk for preterm infants using deep learning models with time-series vital sign data. Npj Digital Medicine, 4(1). https://doi.org/10.1038/s41746-021-00479-4
[27]	Grambsch, P. M., & Therneau, T. M. (1994). Proportional Hazards Tests and Diagnostics Based on Weighted Residuals. Biometrika, 81(3), 515-515. https://doi.org/10.2307/2337123
[28]	Spytek, M., Krzyziński, M., Langbein, S. H., Baniecki, H., Wright, M. N., & Biecek, P. (2023). survex: an R package for explaining machine learning survival models. Bioinformatics, 39(12). https://doi.org/10.1093/bioinformatics/btad723
[29]	Blencowe, H., Cousens, S., Chou, D., Oestergaard, M., Say, L., Moller, A.-B., Kinney, M., & Lawn, J. (2013). Born Too Soon: The global epidemiology of 15 million preterm births. Reproductive Health, 10(Suppl 1), S2. https://doi.org/10.1186/1742-4755-10-s1-s2
[30]	Casey, B. M., McIntire, D. D., & Leveno, K. J. (2021). The Continuing Value of the Apgar Score for the Assessment of Newborn Infants. New England Journal of Medicine, 344(7), 467-471. https://doi.org/10.1056/nejm200102153440701
[31]	Darnton-Hill, I., & Mkparu, U. (2015). Micronutrients in Pregnancy in Low- and Middle-Income Countries. Nutrients, 7(3), 1744-1768. https://doi.org/10.3390/nu7031744
[32]	Cooper, P. A. (2013). The Challenge of Reducing Neonatal Mortality in Low- and Middle-Income Countries. Pediatrics, 133(1), 4-6. https://doi.org/10.1542/peds.2013-2579
[33]	Nakua, E. K., Sevugu, J. T., Dzomeku, V. M., Otupiri, E., Lipkovich, H. R., & Owusu-Dabo, E. (2015). Home birth without skilled attendants despite millennium villages project intervention in Ghana: insight from a survey of women’s perceptions of skilled obstetric care. BMC Pregnancy and Childbirth, 15(1). https://doi.org/10.1186/s12884-015-0674-1
[34]	McIntire, D. D., & Leveno, K. J. (2008). Neonatal Mortality and Morbidity Rates in Late Preterm Births Compared With Births at Term. Obstetrics & Gynecology, 111(1), 35-41. https://doi.org/10.1097/01.aog.0000297311.33046.73
[35]	Sokou, R., Lianou, A., Lampridou, M., Panagiotounakou, P., Kafalidis, G., Paliatsiou, S., Volaki, P., Tsantes, A. G., Boutsikou, T., Iliodromiti, Z., & Iacovidou, N. (2025). Neonates at Risk: Understanding the Impact of High-Risk Pregnancies on Neonatal Health. Medicina, 61(6), 1077-1077. https://doi.org/10.3390/medicina61061077
[36]	Liu, Q., Deng, J., Yan, W., Qin, C., Du, M., Wang, Y., Zhang, S., Liu, M., & Liu, J. (2024). Burden and trends of infectious disease mortality attributed to air pollution, unsafe water, sanitation, and hygiene, and non-optimal temperature globally and in different socio-demographic index regions. Global Health Research and Policy, 9(1). https://doi.org/10.1186/s41256-024-00366-x
[37]	Nabila, M., Baidani, A., Mourajid, Y., Chebabe, M., & Abderraouf, H. (2024). Analysis of Risk Determinants of Neonatal Mortality in the Last Decade: A Systematic Literature Review (2013-2023). Pediatric Reports, 16(3), 696-716. https://doi.org/10.3390/pediatric16030059
[38]	Mangu, C. D., Rumisha, S. F., Lyimo, E. P., Mremi, I. R., Massawe, I. S., Bwana, V. M., Chiduo, M. G., & Mboera, L. E. G. (2020). Trends, patterns and cause-specific neonatal mortality in Tanzania: a hospital-based retrospective survey. International Health, 13(4), 334-343. https://doi.org/10.1093/inthealth/ihaa070

Cite This Article

Plain Text BibTeX RIS

APA Style

Lumumba, V. W., Muriithi, D. K., Langat, A. K., Wanyama, M. A., Njoroge, E. W., et al. (2025). A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Computational Biology and Bioinformatics, 13(2), 42-59. https://doi.org/10.11648/j.cbb.20251302.11

Copy | Download

ACS Style

Lumumba, V. W.; Muriithi, D. K.; Langat, A. K.; Wanyama, M. A.; Njoroge, E. W., et al. A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Comput. Biol. Bioinform. 2025, 13(2), 42-59. doi: 10.11648/j.cbb.20251302.11

Copy | Download

AMA Style

Lumumba VW, Muriithi DK, Langat AK, Wanyama MA, Njoroge EW, et al. A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Comput Biol Bioinform. 2025;13(2):42-59. doi: 10.11648/j.cbb.20251302.11

Copy | Download

@article{10.11648/j.cbb.20251302.11,
  author = {Victor Wandera Lumumba and Denis Kariuki Muriithi and Amos Kipkorir Langat and Maureen Ambasa Wanyama and Elizabeth Wambui Njoroge and John Kamwele Mutinda and Olivia Waka and Edson Mwebesa},
  title = {A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction
},
  journal = {Computational Biology and Bioinformatics},
  volume = {13},
  number = {2},
  pages = {42-59},
  doi = {10.11648/j.cbb.20251302.11},
  url = {https://doi.org/10.11648/j.cbb.20251302.11},
  eprint = {https://article.sciencepublishinggroup.com/pdf/10.11648.j.cbb.20251302.11},
  abstract = {Neonatal mortality remains a critical public health challenge, particularly in low- and middle-income countries (LMICs), where limited healthcare resources and fragmented follow-up systems hinder timely interventions. Accurate prediction of neonatal death is essential for risk stratification, resource allocation, and improving survival outcomes. While traditional survival analysis methods like the Kaplan-Meier estimator and Cox proportional hazards (Cox PH) model are widely used, they face limitations in handling non-linear relationships, high-dimensional data, and violations of proportional hazards assumptions. Random Survival Forests (RSF), a machine learning approach, offers potential advantages but lacks sufficient comparative evaluation in neonatal mortality prediction, especially within LMIC contexts. This study aimed to comparatively evaluate the performance of Kaplan-Meier, Cox PH, and RSF models in predicting neonatal mortality using a synthetic dataset reflecting perinatal epidemiology in Kenya. The research addresses a significant and direct methodological comparisons across these models in neonatal populations, particularly under real-world conditions involving censoring, missing data, and non-proportional hazards. We assessed discrimination (C-index, time-dependent AUC), calibration (Integrated Brier Score, CRPS), and clinical interpretability. The dataset included 2,000 neonates with 17 covariates including but not limited to gestational age, birth weight, maternal health, and socioeconomic status. Results showed that RSF outperformed both Kaplan-Meier and Cox PH in discrimination (C-index: 0.875 vs. 0.868) and maintained strong calibration, particularly at 28 days. Variable importance measures identified gestational age, birth weight, and maternal health score as top predictors. SHAP values enhanced interpretability of RSF outputs. The Cox model provided clinically intuitive hazard ratios but was less flexible in capturing interactions. The study concluded that RSF offers superior predictive accuracy for neonatal mortality and should be integrated into risk prediction tools, especially in data-rich settings. Policy makers should support adoption of advanced analytics in perinatal care systems, while maintaining traditional models for inferential clarity. Combining both paradigms can optimize neonatal survival strategies.
},
 year = {2025}
}

Copy | Download

TY - JOUR
T1 - A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction

AU - Victor Wandera Lumumba
AU - Denis Kariuki Muriithi
AU - Amos Kipkorir Langat
AU - Maureen Ambasa Wanyama
AU - Elizabeth Wambui Njoroge
AU - John Kamwele Mutinda
AU - Olivia Waka
AU - Edson Mwebesa
Y1 - 2025/10/27
PY - 2025
N1 - https://doi.org/10.11648/j.cbb.20251302.11
DO - 10.11648/j.cbb.20251302.11
T2 - Computational Biology and Bioinformatics
JF - Computational Biology and Bioinformatics
JO - Computational Biology and Bioinformatics
SP - 42
EP - 59
PB - Science Publishing Group
SN - 2330-8281
UR - https://doi.org/10.11648/j.cbb.20251302.11
AB - Neonatal mortality remains a critical public health challenge, particularly in low- and middle-income countries (LMICs), where limited healthcare resources and fragmented follow-up systems hinder timely interventions. Accurate prediction of neonatal death is essential for risk stratification, resource allocation, and improving survival outcomes. While traditional survival analysis methods like the Kaplan-Meier estimator and Cox proportional hazards (Cox PH) model are widely used, they face limitations in handling non-linear relationships, high-dimensional data, and violations of proportional hazards assumptions. Random Survival Forests (RSF), a machine learning approach, offers potential advantages but lacks sufficient comparative evaluation in neonatal mortality prediction, especially within LMIC contexts. This study aimed to comparatively evaluate the performance of Kaplan-Meier, Cox PH, and RSF models in predicting neonatal mortality using a synthetic dataset reflecting perinatal epidemiology in Kenya. The research addresses a significant and direct methodological comparisons across these models in neonatal populations, particularly under real-world conditions involving censoring, missing data, and non-proportional hazards. We assessed discrimination (C-index, time-dependent AUC), calibration (Integrated Brier Score, CRPS), and clinical interpretability. The dataset included 2,000 neonates with 17 covariates including but not limited to gestational age, birth weight, maternal health, and socioeconomic status. Results showed that RSF outperformed both Kaplan-Meier and Cox PH in discrimination (C-index: 0.875 vs. 0.868) and maintained strong calibration, particularly at 28 days. Variable importance measures identified gestational age, birth weight, and maternal health score as top predictors. SHAP values enhanced interpretability of RSF outputs. The Cox model provided clinically intuitive hazard ratios but was less flexible in capturing interactions. The study concluded that RSF offers superior predictive accuracy for neonatal mortality and should be integrated into risk prediction tools, especially in data-rich settings. Policy makers should support adoption of advanced analytics in perinatal care systems, while maintaining traditional models for inferential clarity. Combining both paradigms can optimize neonatal survival strategies.

VL - 13
IS - 2
ER -

Copy | Download

Author Information

Victor Wandera Lumumba

Department of Physical Sciences, Chuka University, Chuka, Kenya

Contact Email

http://orcid.org/0009-0000-2840-8364
Denis Kariuki Muriithi

Department of Physical Sciences, Chuka University, Chuka, Kenya

Contact Email

http://orcid.org/0000-0002-3210-0925
Amos Kipkorir Langat

Department of Statistics and Actuarial Sciences, Jomo Kenyatta University of Agriculture and Technology, Nairobi, Kenya

Contact Email

http://orcid.org/0000-0002-7813-6835
Maureen Ambasa Wanyama

Department of Public Health, Medwise Solutions Consultancy Limited, Nairobi, Kenya

Contact Email

http://orcid.org/0009-0003-7008-350X
Elizabeth Wambui Njoroge

Department of Physical Sciences, Chuka University, Chuka, Kenya

Contact Email

http://orcid.org/0009-0009-4732-7354
John Kamwele Mutinda

Department of Statistics and Finance, University of Science and Technology of China, Hefei, China

Contact Email

http://orcid.org/0009-0004-7919-2064
Olivia Waka

Department of Physical Sciences, Chuka University, Chuka, Kenya

Contact Email

http://orcid.org/0009-0000-1444-2844
Edson Mwebesa

Department of Mathematics, Physics and Computing, Moi University, Eldoret, Kenya

Contact Email

http://orcid.org/0000-0003-4421-5356

Download PDF

Submit an Article

Plain Text BibTeX RIS

APA Style

Lumumba, V. W., Muriithi, D. K., Langat, A. K., Wanyama, M. A., Njoroge, E. W., et al. (2025). A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Computational Biology and Bioinformatics, 13(2), 42-59. https://doi.org/10.11648/j.cbb.20251302.11

Copy | Download

ACS Style

Lumumba, V. W.; Muriithi, D. K.; Langat, A. K.; Wanyama, M. A.; Njoroge, E. W., et al. A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Comput. Biol. Bioinform. 2025, 13(2), 42-59. doi: 10.11648/j.cbb.20251302.11

Copy | Download

AMA Style

Lumumba VW, Muriithi DK, Langat AK, Wanyama MA, Njoroge EW, et al. A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction. Comput Biol Bioinform. 2025;13(2):42-59. doi: 10.11648/j.cbb.20251302.11

Copy | Download

@article{10.11648/j.cbb.20251302.11,
  author = {Victor Wandera Lumumba and Denis Kariuki Muriithi and Amos Kipkorir Langat and Maureen Ambasa Wanyama and Elizabeth Wambui Njoroge and John Kamwele Mutinda and Olivia Waka and Edson Mwebesa},
  title = {A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction
},
  journal = {Computational Biology and Bioinformatics},
  volume = {13},
  number = {2},
  pages = {42-59},
  doi = {10.11648/j.cbb.20251302.11},
  url = {https://doi.org/10.11648/j.cbb.20251302.11},
  eprint = {https://article.sciencepublishinggroup.com/pdf/10.11648.j.cbb.20251302.11},
  abstract = {Neonatal mortality remains a critical public health challenge, particularly in low- and middle-income countries (LMICs), where limited healthcare resources and fragmented follow-up systems hinder timely interventions. Accurate prediction of neonatal death is essential for risk stratification, resource allocation, and improving survival outcomes. While traditional survival analysis methods like the Kaplan-Meier estimator and Cox proportional hazards (Cox PH) model are widely used, they face limitations in handling non-linear relationships, high-dimensional data, and violations of proportional hazards assumptions. Random Survival Forests (RSF), a machine learning approach, offers potential advantages but lacks sufficient comparative evaluation in neonatal mortality prediction, especially within LMIC contexts. This study aimed to comparatively evaluate the performance of Kaplan-Meier, Cox PH, and RSF models in predicting neonatal mortality using a synthetic dataset reflecting perinatal epidemiology in Kenya. The research addresses a significant and direct methodological comparisons across these models in neonatal populations, particularly under real-world conditions involving censoring, missing data, and non-proportional hazards. We assessed discrimination (C-index, time-dependent AUC), calibration (Integrated Brier Score, CRPS), and clinical interpretability. The dataset included 2,000 neonates with 17 covariates including but not limited to gestational age, birth weight, maternal health, and socioeconomic status. Results showed that RSF outperformed both Kaplan-Meier and Cox PH in discrimination (C-index: 0.875 vs. 0.868) and maintained strong calibration, particularly at 28 days. Variable importance measures identified gestational age, birth weight, and maternal health score as top predictors. SHAP values enhanced interpretability of RSF outputs. The Cox model provided clinically intuitive hazard ratios but was less flexible in capturing interactions. The study concluded that RSF offers superior predictive accuracy for neonatal mortality and should be integrated into risk prediction tools, especially in data-rich settings. Policy makers should support adoption of advanced analytics in perinatal care systems, while maintaining traditional models for inferential clarity. Combining both paradigms can optimize neonatal survival strategies.
},
 year = {2025}
}

Copy | Download

TY - JOUR
T1 - A Comparative Evaluation of Kaplan-Meier, Cox Proportional Hazards, and Random Survival Forests for Neonatal Mortality Prediction

VL - 13
IS - 2
ER -

Copy | Download