https://orcid.org/0009-0006-6957-1073
Figures
Abstract
Background
Early-onset sepsis (EOS) remains a leading cause of mortality and neurodevelopmental injury in preterm infants, yet widely used tools (e.g., Kaiser EOS Calculator) are not designed for <35–37 weeks’ gestation.
Objective
To develop and internally validate a concise, clinically interpretable nomogram for EOS risk stratification in preterm neonates and to evaluate its potential for antibiotic stewardship.
Methods
We performed a single-center retrospective cohort study (July 2023–June 2024) including 1,059 preterm infants admitted within 72 h of birth, randomly split 7:3 into training (n = 742) and validation (n = 317). Forty-five maternal and neonatal candidates were screened (univariable tests, LASSO), followed by multivariable logistic regression to build the final model and nomogram. Discrimination (AUC), calibration (Brier score, calibration curve, Hosmer–Lemeshow), and decision-curve analysis (DCA) were assessed; two biologically plausible interactions were prespecified.
Results
Four routinely available variables—gestational age, birth weight, umbilical cord abnormality, and mechanical ventilation within 72 h—composed the final model. In the validation cohort, AUC was 0.818 (95% CI, 0.767–0.868), Brier score 0.158, and Hosmer–Lemeshow P = 0.71; DCA showed net benefit across 5–65% risk thresholds. Using a ≥ 0.70 treatment threshold, the model identified 88% of EOS cases while recommending antibiotics for ~10% of infants. A culture-proven-only sensitivity analysis yielded comparable discrimination (AUC 0.819) with a Brier score 0.041.
Conclusions
A four-factor nomogram using EMR-available variables accurately stratifies EOS risk in preterm infants and may support risk-based antibiotic decisions while limiting overtreatment. Prospective multicenter external validation is warranted to confirm generalizability and guide implementation.
Citation: Guo L, Zhang Z, Zhao C, Shen C, Zhao H, Chen H (2025) Four-factor nomogram for early-onset sepsis in preterm neonates: Development and internal validation of a stewardship tool. PLoS One 20(10): e0334342. https://doi.org/10.1371/journal.pone.0334342
Editor: Fumihiko Namba, Saitama Medical Center: Saitama Ika Daigaku Sogo Iryo Center, JAPAN
Received: July 19, 2025; Accepted: September 25, 2025; Published: October 9, 2025
Copyright: © 2025 Guo et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data for this study are publicly available from the Zenodo repository (https://doi.org/10.5281/zenodo.17141931).
Funding: This work was supported by the 2024 Hebei Provincial Medical Science Research Program (Grant No. 20241744, project title: “Relationship between hypertensive disorders of pregnancy, gestational diabetes mellitus, and early-onset neonatal sepsis in preterm infants.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Early-onset sepsis (EOS) remains a major cause of death and neurodevelopmental injury in preterm infants (<37 weeks), whose mortality is 3–5 times higher than in term neonates and is further compounded by early empirical antibiotic use [1–3]. Although empiric therapy broadens initial coverage, it disrupts the microbiome, drives resistance, and may have downstream metabolic harms [4]. To curb these effects, AAP and NICE advocate risk-based initiation; however, widely used tools such as the Kaiser EOS Calculator (designed for ≥35 weeks) do not address key vulnerabilities of very preterm infants—immune immaturity, ventilator dependence, and perinatal hypoxia-ischaemia [5,6]. We therefore developed and internally validated a four-factor nomogram tailored to preterm infants (<37 weeks, including <35 weeks), using routinely available perinatal variables, and evaluated its discrimination, calibration, and decision-curve utility to inform antibiotic stewardship.
While machine learning improves late-onset sepsis prediction, its EOS use remains limited and lacks mechanistic clarity [7]. Many models rely on high-frequency data or complex inputs, hindering clinical use and lacking biological transparency [8]. Key risk factors—such as umbilical cord abnormalities, maternal metabolic disorders (e.g., hyperglycemia, insulin resistance), and ventilator-induced lung injury (VILI)—disrupt immune homeostasis but are often omitted in EOS models [9,10].
This study aimed to develop and internally validate a concise, interpretable EOS risk model using data from preterm infants admitted to a tertiary neonatal hospital (2023–2024). We further assessed its value for NICU stratified management and examined biological interactions to inform personalized antibiotic decisions.
Materials and methods
Study design and participants
This single-center retrospective cohort study was conducted in the Neonatal Department of the Fourth Hospital of Shijiazhuang (Hebei, China) from July 2023 to June 2024. Ethics approval was obtained (Approval No. 20230134), and written informed consent was obtained from parents or legal guardians and documented in the medical record. All data were de-identified prior to analysis in accordance with institutional policy.. Based on CDC and AAP criteria [5,11], we included live-born preterm infants (GA < 37 weeks) admitted within 72 hours and with complete maternal and neonatal data (<20% missing). Exclusion criteria were major malformations, palliative care, early death (≤72 h), incomplete culture/treatment data, or duplicated records. We excluded infants who died within 72 h to minimize outcome ascertainment bias and because early treatment responses could not be reliably observed. A total of 1,059 infants were enrolled and randomly split (7:3) into training (n = 742) and validation (n = 317) sets using a fixed seed (set.seed = 3464) (Fig 1). Missing-data handling is detailed in Variable Selection and Missing Data Handling; see S1 Table.
(A) Coefficient trajectories of 36 candidate variables across log(λ) values. As λ increases, uninformative variables shrink toward zero. (B) Tenfold cross-validation curve identifying the optimal penalty (λ_min) that minimizes binomial deviance. A total of seven predictors were retained for multivariable modeling based on this selection.
Definition of outcomes
EOS was defined as either: (1) positive culture of pathogenic bacteria from blood or other sterile fluids within 72 hours of birth, or (2) negative culture with typical infection signs (e.g., respiratory distress, temperature instability, reduced consciousness) plus at least one elevated inflammatory marker (CRP ≥ 10 mg/L, PCT ≥ 2 ng/mL, or IL-6 ≥ 50 pg/mL). Clinical improvement after ≥5 days of antibiotics was required. In cases of death occurring during the treatment period, these infants were excluded from the clinical improvement criteria, but their EOS diagnosis was retained if they met the culture or clinical criteria. The definition was based on AAP and CDC guidelines for neonatal infection and GBS screening [5,11].
Variable selection and missing data handling
We extracted 45 candidate variables, encompassing maternal demographics, antenatal complications, perinatal events, delivery/resuscitation details, early interventions (e.g., corticosteroids, surfactant, mechanical ventilation within 72 h), and initial lab results. To ensure the integrity and accuracy of the analysis, missing data were handled using multiple imputation with the MICE package in R. Continuous variables were imputed using linear regression, and categorical variables with logistic regression. Variables with missingness ≥5% were excluded. Five imputed datasets were generated, and sensitivity analysis was conducted to evaluate the model’s performance across these datasets. The performance metrics (AUC, accuracy, sensitivity, specificity) exhibited negligible variation, indicating high consistency. Given the minimal differences observed, Dataset 4 (denoted with an asterisk) was randomly selected for further analysis (S1 Table). Ultimately, 36 variables with complete data were included. To ensure consistency in binary definitions (e.g., maternal occupation, delivery mode, cord abnormalities), standard criteria are detailed in S2 Table. Nine variables were excluded due to high missingness, ambiguity, or confounding (see S3 Table).
Statistical analysis
Baseline comparisons were conducted using chi-square/Fisher’s exact tests (categorical) and t-test/Wilcoxon tests (continuous, α = 0.05, two-sided). Missing data handling was performed through multiple imputation, generating five imputed datasets. To ensure the robustness of the model, sensitivity analysis was conducted across these datasets. Variable selection was initially performed with univariate logistic regression (P < 0.05), followed by LASSO (glmnet, α = 1, 20-fold CV, λ_min). Overlapping variables were then entered into a multivariate logistic regression model with stepwise selection (min AIC). A nomogram was constructed using the rms package [12]. Model discrimination was assessed using AUC (95% CI), sensitivity, specificity, and the Youden index. Calibration used 1,000 bootstrap resamples, Brier score, intercept/slope, HL test, and E/O ratio. DCA over a 0.05–0.65 threshold range (rmda) evaluated clinical utility. EOS risk was stratified as low (<0.30), moderate (0.30–0.70), and high (>0.70), with chi-square and trend tests used to compare groups. Two interaction terms (cord × gestational age; ventilation × weight) were pre-specified and visualized (interactions). Analyses used R 4.3.2 with glmnet, rms, pROC, ResourceSelection, rmda, interactions, and mice.
Sensitivity analyses
To benchmark the composite outcome against culture-proven incidence, we conducted a culture-proven–only analysis in the validation cohort. Predicted probabilities were generated without refitting from the model trained in the development data, while infants without or with indeterminate culture results were excluded a priori. Performance metrics included AUC (DeLong 95% CI), Brier score, Hosmer–Lemeshow statistics, a calibration curve, and decision-curve analysis across thresholds 0.05–0.65.
Sample-size considerations and EPV
Sample-size considerations: We did not perform an a priori sample-size calculation tailored for multivariable prediction (anticipated incidence, number of candidate predictors, target shrinkage). To mitigate overfitting, we used LASSO penalization and internal validation (pre-specified split with fixed seed and bootstrap), and we provide full model specification and code.
Candidate-level EPV: We initially screened 45 candidate predictors; with 285 EOS events in the training set, the candidate-level EPV ≈ 6.3. We therefore relied on penalization and internal validation rather than post-hoc EPV to support model adequacy.
Results
Study population
A total of 1,059 preterm infants were enrolled (Table 1) and randomly assigned to training (n = 742) and validation (n = 317) sets. EOS incidence was comparable between the two sets (38.4% vs. 38.5%), confirming balanced outcome distribution.
Baseline characteristics comparison
No significant differences were found in maternal or sociodemographic factors between sepsis and non-sepsis groups in either set (P > 0.05), suggesting balanced baseline characteristics (Table 1).
Infants in the sepsis group had significantly lower gestational age, birth weight, and Apgar scores, and higher rates of mechanical ventilation and surfactant use (all P < 0.001), indicating more severe illness. Umbilical cord abnormalities (P = 0.009) and shorter time to NICU admission (P < 0.001) were also more common in the sepsis group. WBC counts were marginally higher in the validation set (P = 0.026); other lab values showed no significant differences.
Screening of candidate variables
Among 36 candidate variables analyzed by univariate logistic regression, 11 were significantly associated with EOS (P < 0.05), with gestational age, birth weight, 1-minute Apgar score, umbilical cord abnormalities, and mechanical ventilation showing the strongest associations(S4 Table). LASSO regression identified seven non-zero predictors at the optimal λ value (Fig 1A and 1B). Six variables overlapped between the two methods—gestational age, birth weight, birth length, 1-minute Apgar score, umbilical cord abnormalities, and mechanical ventilation—and were retained for multivariate modeling.
Multivariate logistic regression model
In multivariate logistic regression, four variables remained independently associated with EOS and were retained in the final model: umbilical cord abnormalities (OR = 1.74, P = 0.010), mechanical ventilation within 72 hours (OR = 0.45, P < 0.001), gestational age (OR = 0.80 per week, P = 0.001), and birth weight (β = –0.001 per 100 g, P = 0.035). Although the odds ratio for ventilation within 72 h is < 1 due to inverse coding, its presence reflects higher illness severity and is associated with increased EOS risk (see Table 2 footnote). Birth length and Apgar score were excluded after adjustment. These four predictors were used to construct the final model and nomogram(Table 2). For transparency and robustness, the full candidate list and univariable screening are provided in S3 and S4 Tables, and comparisons with vs without the two pre-specified interactions and across three selection strategies are shown in S5 and S6 Tables, where validation AUCs differed by ≤0.02 with similar calibration and Brier.
Model construction and nomogram visualization
The final multivariable model incorporated four predictors: umbilical cord abnormalities, mechanical ventilation within 72 hours, gestational age, and birth weight. A nomogram visualized each factor’s contribution for individualized EOS risk estimation(Fig 2). The model’s performance was further assessed in terms of discrimination, calibration, and clinical utility.
The nomogram incorporates four independent predictors: umbilical cord abnormalities, mechanical ventilation within 72 hours, birth weight, and gestational age. Each variable corresponds to a point scale, which is summed to estimate the probability of EOS at the bottom axis. Density plots indicate the distribution of each variable in the training set. The example shown yields a total score of 120 points, corresponding to an EOS probability of 18.2%.
Discrimination performance
The model showed good discriminative ability in the validation set (AUC = 0.818, 95% CI: 0.767–0.868; Fig 3A). Using the optimal cutoff (0.354) derived from the training set, sensitivity, specificity, PPV, and NPV were 72.1%, 75.4%, 64.7%, and 81.2%, respectively (Youden index = 0.475). In the training set, the AUC was 0.740 (95% CI: 0.702–0.778; S1A Fig), with slightly lower sensitivity (66.3%) and specificity (74.8%). Results indicate stable performance without overfitting. In a prespecified culture-proven–only sensitivity analysis of the validation cohort, after excluding infants without/indeterminate culture results (311 analyzed; 15 events, 4.8%), discrimination remained comparable (AUC 0.819, 95% CI 0.725–0.914), and the Brier score was 0.041 (S7 Table).
(A) Receiver operating characteristic (ROC) curve showing the discrimination of the model in the validation set, with an area under the curve (AUC) of 0.818. (B) Calibration plot using 1,000 bootstrap resamples, demonstrating good agreement between predicted and observed probabilities. (C) Decision curve analysis (DCA) indicating a net clinical benefit of the model across a wide range of risk thresholds (5–65%) compared to “treat all” or “treat none” strategies.
Calibration performance
The model showed excellent calibration in the validation set (Brier score = 0.158, HL χ² = 5.40, P = 0.71; Fig 3B). Calibration in the training set was also acceptable (Brier score = 0.193, P = 0.089; E/O = 1.00; S1B Fig). Given the small number of culture-proven events, calibration for that sensitivity analysis is summarized by the calibration curve and Brier statistics, with HL results shown in S7 Table and S2 Fig.
Decision curve analysis
In the validation set, DCA demonstrated greater net benefit than “treat-all” or “treat-none” strategies within the 5%–65% risk threshold range (Fig 3C). Similar trends were noted in the training set, where the model outperformed only outside the < 5% or >65% thresholds (S1C Fig). These findings support using the model within this range to reduce 6%–30% of potential overtreatment while preserving clinical benefit. Under the culture-proven–only outcome, decision-curve analysis also indicated positive net benefit versus treat-all and treat-none at approximately 0.05–0.30 risk thresholds (S2 Fig).
Risk stratification
For clinical application, predicted probabilities were stratified into low (<0.30), moderate (0.30–0.70), and high (>0.70) risk groups, with EOS rates of 27.9%, 66.4%, and 87.5%, respectively (Pearson χ² = 191.07; trend test P < 0.001) (S3A–S3C Fig). Using >0.70 as the intervention threshold could detect 88% of EOS while treating only 10% of infants. Conversely, 70% of low-risk infants may benefit from a 6-hour dynamic reassessment strategy.
Interaction effects analysis
Interaction models revealed potential synergy between key perinatal factors (S4 Fig). For umbilical cord abnormalities and gestational age, the interaction term was marginally significant (β = 0.24, P = 0.098; AIC = 1177.7), showing a 20% EOS risk increase at 25–28 weeks, which attenuated after 34 weeks (S4A Fig). The interaction between mechanical ventilation and birth weight was significant (β = 0.00130, P < 0.001; AIC = 1165.9). Mechanical ventilation increased EOS risk markedly in infants <1500 g (predicted probability >60%), but not in those ≥3000 g (S4B Fig).
These interactions indicate that adverse maternal-fetal conditions may amplify infection risk in extremely preterm or low-birth-weight infants, highlighting the need for individualized antibiotic strategies.
Discussion
Main findings
We developed a nomogram-based EOS prediction model for preterm infants (<37 weeks), using four readily available perinatal variables: gestational age, birth weight, umbilical cord abnormalities, and mechanical ventilation within 72 hours. The model showed strong discrimination in the validation cohort (AUC 0.818; 95% CI: 0.767–0.868), outperforming both the training set (AUC 0.740) and the Kaiser EOS Calculator (AUC 0.65–0.75) [13]. Calibration appeared acceptable based on the calibration curve, Brier score (0.158), and HL statistics. DCA demonstrated consistent clinical benefit across a 5–65% risk range, in line with the 2023 AAP antibiotic guidelines [14]. Using >0.70 as the treatment threshold, the model identified 88% of EOS cases while limiting antibiotic use to 10%, potentially avoiding >70% overtreatment. Stratification and interaction terms supported risk refinement in infants at gestational or weight extremes, consistent with DCA-based neonatal sepsis research [15], and laid the basis for downstream mechanistic discussion.
The comparatively high EOS incidence in our development cohort reflects our composite outcome definition (culture-proven or clinically suspected requiring ≥5 days of antibiotics), which was chosen to prioritize sensitivity in a stewardship context. To benchmark against published culture-proven rates, we conducted a culture-proven–only sensitivity analysis in the validation cohort: after excluding infants without/indeterminate cultures, 311 were analyzed with 15 culture-positive events (4.8%). Discrimination remained comparable to the primary analysis (AUC 0.819, 95% CI 0.725–0.914) with a Brier score of 0.041, indicating that our main conclusions are robust to outcome definition.
Interpretation of results and mechanistic insights
This study developed an EOS prediction model for preterm infants based on four perinatal clinical variables. Their independent and interactive effects are interpreted as follows:
Gestational Age and Birth Weight: Foundations of Innate Immune Maturity: Each additional week of gestation and every 100g increase in birth weight reduced EOS risk by ~20% and 1%, respectively. These findings align with the known maturation of innate immunity. Flow cytometry and single-cell analyses have shown that neonates <32 weeks exhibit impaired innate immune cell counts and function in the first 72h, consistent with “immune developmental delay” [16]. A 2023 cohort study reported that infants born at 23–27 weeks had over fivefold higher infection-related hospitalization within the first year than term infants [17]. These two predictors reflect biologically relevant gradients of prematurity.
Umbilical Cord Abnormalities: Inflammatory Amplification via Ischemia–Reperfusion: Cord abnormalities—such as knots or abnormal insertion—can impair placental perfusion and provoke sterile inflammation, with elevated IL-6 and CRP in umbilical blood [15,18]. In infants born at 25–28 weeks, this factor increased EOS risk by ~20 percentage points, likely amplifying immune vulnerability.
Mechanical Ventilation: Concealed Risks Behind Apparent Protection: Infants receiving ventilation within 72h had higher EOS risk (ORs: 0.33 and 0.45). Ventilation likely reflects illness severity and immune compromise, rather than protection. Prior studies show that VILI increases alveolar-capillary permeability and microbial translocation [19,20]. The interaction with birth weight < 1500g further magnifies this risk, emphasizing tailored respiratory strategies (note: variable coded inversely, hence OR < 1 reflects higher risk).
Interaction Effects: A Biological Basis for Precision Antibiotic Strategies: Two interaction terms—gestational age × cord abnormalities and birth weight × ventilation—highlight synergistic risk patterns: “hypoxia with immune immaturity” and “VILI-induced bacteremia.” These terms improved model fit (lower AIC) and support individualized clinical decisions. Notably, the gestational age × cord abnormality interaction had a P-value of 0.098, suggesting borderline significance, but was retained due to mechanistic plausibility. Sensitivity analysis showed that excluding this term only slightly reduced AUC (from 0.821 to 0.818) and had minimal impact on Brier score, calibration, and Hosmer–Lemeshow test (P from 0.044 to 0.069) (see S5 Table), supporting its inclusion.
Comparison with previous models
The Kaiser Permanente EOS Calculator (KP-EOS) is widely used for neonatal EOS risk assessment, but it was developed for term infants (≥35 weeks) and excludes preterm-specific factors. A 2021–2023 validation in >100,000 term infants showed an AUC of 0.79, with limited accuracy in those <34 weeks. Hence, international guidelines advise its use only in term or near-term infants [14,21].
Direct contrast with KP-EOS. Population: KP-EOS is intended for ≥35 weeks, whereas our model targets preterm infants <37 weeks (including <35 weeks). Predictors: KP-EOS focuses on intrapartum maternal factors and newborn status at birth; our four-factor model additionally incorporates an early NICU intervention (mechanical ventilation within 72 h). Workflow: KP-EOS is largely manual; our nomogram can be embedded in the EHR for automated scoring. Evidence: KP-EOS has limited validation in very preterm infants, while we report internal validation in a preterm cohort.
Few models exist for preterm EOS prediction; fewer have external validation. For instance, a 2024 BMJ Paediatrics Open study modified KP-EOS but included only 94 infants <34 weeks and showed calibration drift [13]. Machine learning (ML) models have reported high internal AUCs (>0.90) using continuous EHR data but lack interpretability and external validation, limiting real-world application [7,22].
This study offers three improvements: Robust sample and modeling: The cohort included 1,059 preterm infants (<37 weeks), and internal validation demonstrated AUC 0.818 with a Brier score 0.158, exceeding prior reports (AUC 0.65–0.75).. Clinical interpretability: Two interaction terms—cord abnormalities × gestational age and ventilation × birth weight—enabled dynamic risk adjustment, improving interpretability. Decision-making utility: DCA quantified net clinical benefit across the 5–65% risk threshold, supporting implementation. Unlike KP-EOS and most ML models, this model integrates DCA, as recommended by TRIPOD-C and PROGRESS guidelines.
In summary, this model advances existing tools through broader applicability, statistical rigor, and clinical integration, providing a reliable EOS risk reference for preterm infants.
Clinical utility and practical implications
Antibiotic Stewardship: Although <2% of neonates develop EOS, over half receive antibiotics [23]. Using a 0.70 threshold, our model achieved 88% sensitivity while reducing immediate antibiotic use to 10%, potentially lowering annual antibiotic consumption by 30–40% per bed, comparable to recent multicenter findings using DCA to guide therapy [24].
Integration into Clinical Pathways and Resource Allocation: The model’s three-tiered risk stratification aligns with the AAP’s “6-hour reassessment” strategy: high-risk infants receive cultures and treatment; moderate-risk infants undergo reassessment; low-risk infants are monitored clinically [5]. For a NICU with ~800 annual preterm admissions, this approach may reduce blood cultures, cut testing costs, and streamline clinician workload.
Digital Implementation and Scalability: The four variables are available in EMRs and can be auto-integrated via HL7. Unlike the manually operated KP-EOS (≥35 weeks only), this model supports real-time risk alerts and treatment guidance, fitting emerging trends in digital pediatric ICU tools [25].
Broad Applicability and Health Economic Value: A JAMA Network Open study showed that a 10% antibiotic reduction in NICUs saves ~ ¥1,240 per infant [26]. Applied regionally, our model could yield multi-million yuan savings and mitigate antimicrobial resistance, supporting WHO AWaRe and national stewardship goals.
Strengths and innovations
A Robust Modeling Process: The study enrolled 1,059 cases and 285 EOS events; modeling followed the 2022 TRIPOD and PROGRESS “C-type” workflow [27], incorporating multiple imputation, LASSO selection, predefined interactions, and stratified internal validation. Code and seeds are fully disclosed to ensure reproducibility.
Enhanced Interpretability via Interaction Terms:Two biologically plausible interaction terms—gestational age × cord abnormality and birth weight × ventilation—were included to enhance clinical understanding. Risk surface visualizations clarified their synergistic effects, allowing frontline clinicians to interpret outputs without statistical expertise.
Quantifying Net Benefit via DCA: DCA demonstrated net clinical benefit across a 5–65% threshold range, supporting antibiotic stewardship. Despite growing recognition by TRIPOD-C, DCA remains underused in neonatal sepsis modeling [28], highlighting this study’s translational value.
Nomogram + Three-Tiered Stratification” Enables One-Click Deployment: The model requires only four EMR variables and delivers risk-tiered results suitable for real-time integration into HIS–CDSS platforms. Simulated deployment showed alert response times <0.2 seconds, aligning with Level 3 of the Digital Health Maturity Model.
Together, these features highlight the model’s methodological transparency, clinical usability, and deployment efficiency, supporting its potential to improve EOS risk stratification in preterm infants.
Limitations
Despite adherence to TRIPOD/PROGRESS guidelines and favorable validation, seven key limitations merit attention:
Single-center, retrospective design: The dataset originated from a single institution (July 2023–Jun 2024), possibly limiting generalizability due to local microbiology and care practices. Retrospective design may also entail data loss. Prior studies suggest ~70% of neonatal prediction models overestimate performance without external validation [29]. Sample size and potential overfitting. We did not perform an a priori sample-size calculation tailored for multivariable prediction (anticipated incidence, number of candidate predictors, target shrinkage). Although LASSO penalization and internal validation were used, some residual overfitting cannot be excluded. For transparency, the candidate-level EPV is reported in Methods (285 events for 45 initial candidates; EPV ≈ 6.3), and full model specification and code are provided to facilitate appraisal and future external validation. Outcome classification bias: EOS was defined to include culture-negative cases with elevated inflammation markers and ≥5 days of antibiotics. While enhancing sensitivity, this may include fungal or viral infections, especially in antibiotic-overuse settings [30,31]. Future work should use datasets with complete pathogen data. Limited range of candidate variables: The model excluded metabolic indicators (e.g., glucose, lipids) and serial infection markers (e.g., CRP, PCT). Incorporating multimodal data—metabolomics, microbiomics, and real-time vitals—could enhance prediction and interpretability.The Risk Threshold Was Empirically Set: The thresholds (0.30 and 0.70) were determined based on the DCA curve, the 0.05–0.65 interval recommended by the AAP-endorsed KP-EOS Calculator [5], widely adopted thresholds in international EOS prediction tools [32,33], No formal cost–benefit algorithm was applied, and future adjustments may be needed to account for regional pathogen profiles and resource constraints. To verify variable selection robustness, we developed two alternatives: a 7-variable “LASSO-only” model and an 11-variable univariate logistic regression model. Both showed AUCs within 0.02 of the final model in the validation set (S6 Table), supporting selection stability. Finally, in the culture-proven–only sensitivity analysis the number of events was small (15 in the validation cohort; 311 analyzed after excluding 6 without/indeterminate cultures), which limits the precision of calibration parameters. We therefore emphasize the calibration curve, Brier score, and HL statistics and defer full reporting of calibration intercept and slope to future external validations.
The model still awaits prospective multicenter validation to confirm performance across diverse NICUs.To enhance generalizability, future work will follow TRIPOD-CLUSTER guidelines for temporal and geographic validation, with prospective multicenter studies as needed.
Future directions for research
To support refinement and implementation, future research will follow four directions:
Temporal Generalizability Validation: New cases from 2025–2026 at our center will be used for time-based validation, followed by geographic validation in NICUs with distinct bacterial profiles and care practices, in line with TRIPOD-CLUSTER guidelines [34]. Multimodal Data Integration: Incorporating maternal metabolic markers (e.g., fasting glucose, HbA1c), placental/umbilical cord metabolomics, and neonatal CRP/PCT trends into the EMR may improve model performance. A “segmented LASSO + ensemble learning” upgrade is planned, as prior studies show such integration can raise predictive AUC by 0.05–0.08 [15]. Digital Implementation and Dynamic Recalibration: Real-time integration via the HIS-CDSS platform is underway. Model parameters are recalibrated quarterly using sliding-window algorithms to mitigate drift and maintain DCA alignment, consistent with TRIPOD-AI’s “continuous learning” framework [35]. Impact on Decision-Making and Economic Evaluation: A crossover cohort may compare model- vs empirically guided antibiotic use and costs. Prior estimates suggest stewardship-driven 10% antibiotic reduction saves ~ ¥1,240 per neonate and cuts hospitalization costs by 8–12% [36].
In summary, this closed-loop strategy—spanning validation, optimization, recalibration, and impact evaluation—may enhance the model’s utility and promote antimicrobial stewardship in NICUs.
Supporting information
S1 Fig. Model performance in the training cohort.
(A) ROC curve showing the discrimination of the model in the training set (AUC = 0.740). (B) Calibration plot using 1,000 bootstrap resamples. (C) Decision curve analysis demonstrating clinical net benefit of the model across various risk thresholds.
https://doi.org/10.1371/journal.pone.0334342.s001
(TIF)
S2 Fig. Decision-curve analysis under a culture-proven-only outcome (validation cohort).
Predicted probabilities were obtained from the training-set model without refitting. Net benefit is shown across thresholds 0.05–0.65 with 200 bootstrap resamples for confidence intervals (policy: opt-in). The model yielded higher net benefit than “treat-all” and “treat-none” at low–to-moderate thresholds, consistent with the low prevalence of culture-proven EOS. Validation cohort: n = 311 analyzed after excluding 6 infants with no/indeterminate cultures; 15 culture-positive events (4.8%).
https://doi.org/10.1371/journal.pone.0334342.s002
(TIF)
S3 Fig. Risk stratification of early-onset sepsis based on predicted probability.
(A) Distribution of sepsis and non-sepsis cases across low, moderate, and high predicted risk groups. (B) Sepsis rates with 95% confidence intervals in each group. (C) Increasing trend in sepsis rate with higher predicted risk level.
https://doi.org/10.1371/journal.pone.0334342.s003
(TIF)
S4 Fig. Interaction effects between key predictors on sepsis risk.
(A) Interaction between gestational age and umbilical cord abnormalities. The risk-reducing effect of gestational age is attenuated in infants with cord abnormalities. (B) Interaction between birth weight and mechanical ventilation within 72 hours. The risk-lowering effect of birth weight is offset in ventilated infants. Shaded areas represent 95% confidence intervals.
https://doi.org/10.1371/journal.pone.0334342.s004
(TIF)
S1 Table. Performance comparison of five imputed datasets: AUC, accuracy, sensitivity, and specificity.
https://doi.org/10.1371/journal.pone.0334342.s005
(DOCX)
S2 Table. Definitions of binary variables used in the baseline dataset.
https://doi.org/10.1371/journal.pone.0334342.s006
(DOCX)
S3 Table. Candidate variables excluded prior to modeling.
https://doi.org/10.1371/journal.pone.0334342.s007
(DOCX)
S4 Table. Univariable logistic regression analysis of candidate predictors for early-onset sepsis (EOS) in the training cohort.
https://doi.org/10.1371/journal.pone.0334342.s008
(DOCX)
S5 Table. Comparison of model performance with and without interaction terms in the validation cohort.
https://doi.org/10.1371/journal.pone.0334342.s009
(DOCX)
S6 Table. Comparison of predictive performance among three variable selection strategies in the validation cohort.
https://doi.org/10.1371/journal.pone.0334342.s010
(DOCX)
S7 Table. Culture-proven–only sensitivity analysis in the validation cohort.
https://doi.org/10.1371/journal.pone.0334342.s011
(DOCX)
References
- 1. Shane AL, Sánchez PJ, Stoll BJ. Neonatal sepsis. Lancet. 2017;390(10104):1770–80. pmid:28434651
- 2. Stoll BJ, Hansen NI, Bell EF, Walsh MC, Carlo WA, Shankaran S, et al. Trends in care practices, morbidity, and mortality of extremely preterm neonates, 1993-2012. JAMA. 2015;314(10):1039.
- 3. Stoll BJ, Puopolo KM, Hansen NI, Sánchez PJ, Bell EF, Carlo WA, et al. Early-onset neonatal sepsis 2015 to 2017, the rise of Escherichia coli, and the need for novel prevention strategies. JAMA Pediatr. 2020;174(7):e200593. pmid:32364598
- 4. Fleiss N, Tarun S, Polin RA. Infection prevention for extremely low birth weight infants in the NICU. Semin Fetal Neonatal Med. 2022;27(3):101345. pmid:35550785
- 5. Puopolo KM, Benitz WE, Zaoutis TE, Cummings J, Juul S, Hand I, et al. Management of neonates born at ≥35 0/7 weeks’ gestation with suspected or proven early-onset bacterial sepsis. Pediatrics. 2018;142(6).
- 6.
Health NIf, Excellence C. Neonatal infection: antibiotics for prevention and treatment. National Institute for Health and Care Excellence (NICE); 2021.
- 7. Tan X, Zhang X, Chai J, Ji W, Ru J, Yang C, et al. Constructing a predictive model for early-onset sepsis in neonatal intensive care unit newborns based on SHapley Additive exPlanations explainable machine learning. Transl Pediatr. 2024;13(11):1933–46. pmid:39649648
- 8. Yang M, Peng Z, van Pul C, Andriessen P, Dong K, Silvertand D, et al. Continuous prediction and clinical alarm management of late-onset sepsis in preterm infants using vital signs from a patient monitor. Comput Methods Programs Biomed. 2024;255:108335. pmid:39047574
- 9. Murillo C, Rueda C, Larroya M, Boada D, Grau L, Ponce J, et al. Intra-amniotic infection and/or inflammation is associated with fetal cardiac concentric hypertrophy and diastolic dysfunction in preterm labor and preterm prelabor rupture of membranes. Am J Obstet Gynecol. 2024;230(6):665.e1-665.e30. pmid:38290925
- 10. Yu B, Jia S, Chen Y, Guan R, Chen S, Tang W, et al. CXCL4 deficiency limits M4 macrophage infiltration and attenuates hyperoxia-induced lung injury. Mol Med. 2024;30(1):253. pmid:39707183
- 11. Verani JR, McGee L, Schrag SJ. Prevention of perinatal group B streptococcal disease--revised guidelines from CDC, 2010. MMWR Recomm Rep. 2010;59(Rr-10):1–36.
- 12. Niestroy JC, Moorman JR, Levinson MA, Manir SA, Clark TW, Fairchild KD, et al. Discovery of signatures of fatal neonatal illness in vital signs using highly comparative time-series analysis. NPJ Digit Med. 2022;5(1):6. pmid:35039624
- 13. Alhindi MY, Almalki FG, Al Saif S, Almalahi A, Alsaegh MH, Mustafa A, et al. Evaluating a modified use of the kaiser permanente early-onset sepsis risk calculator to reduce antibiotic exposure: a retrospective study. BMJ Paediatr Open. 2024;8(1):e002597. pmid:38844386
- 14. Kuzniewicz MW, Escobar GJ, Forquer H, Li S, Shu D, Kipnis P, et al. Update to the neonatal early-onset sepsis calculator utilizing a contemporary cohort. Pediatrics. 2024;154(4).
- 15. Wu M, Deng Y, Wang X, He B, Wei F, Zhang Y. Development of risk prediction nomogram for neonatal sepsis in Group B Streptococcus-colonized mothers: a retrospective study. Sci Rep. 2024;14(1):5629. pmid:38453985
- 16. Das A, Ariyakumar G, Gupta N, Kamdar S, Barugahare A, Deveson-Lucas D, et al. Identifying immune signatures of sepsis to increase diagnostic accuracy in very preterm babies. Nat Commun. 2024;15(1):388. pmid:38195661
- 17. Nilsen SM, Valand J, Rogne T, Asheim A, Yin W, Metsälä J, et al. Gestational age at birth and hospitalisations for infections among individuals aged 0–50 years in Norway: a longitudinal, register-based, cohort study. eClinicalMedicine. 2023;62:102108.
- 18. Andersson CB, Klingenberg C, Thellesen L, Johnsen SP, Kesmodel US, Petersen JP. Umbilical cord pH levels and neonatal morbidity and mortality. JAMA Netw Open. 2024;7(8):e2427604. pmid:39141385
- 19. Kalikkot Thekkeveedu R, El-Saie A, Prakash V, Katakam L, Shivanna B. Ventilation-induced lung injury (VILI) in neonates: evidence-based concepts and lung-protective strategies. J Clin Med. 2022;11(3):557. pmid:35160009
- 20. Ding P, Yang R, Li C, Fu H-L, Ren G-L, Wang P, et al. Fibroblast growth factor 21 attenuates ventilator-induced lung injury by inhibiting the NLRP3/caspase-1/GSDMD pyroptotic pathway. Crit Care. 2023;27(1):196. pmid:37218012
- 21. Fleiss N, Schwabenbauer K, Randis TM, Polin RA. What’s new in the management of neonatal early-onset sepsis? Arch Dis Child Fetal Neonatal Ed. 2023;108(1):10–4. pmid:35618407
- 22. Tądel K, Dudek A, Bil-Lula I. AI algorithms for modeling the risk, progression, and treatment of sepsis, including early-onset sepsis-a systematic Review. J Clin Med. 2024;13(19):5959. pmid:39408019
- 23. Mascarenhas D, Ho MSP, Ting J, Shah PS. Antimicrobial stewardship programs in neonates: a meta-analysis. Pediatrics. 2024;153(6).
- 24. Lloyd LG, van Weissenbruch MM, Bekker A, Ferreyra C, Gleeson B, Dramowski A. Theoretical impact of a bedside decision-making tool on antibiotic use for suspected neonatal healthcare-associated infection: an observational study. BMC Pediatr. 2025;25(1):52. pmid:39838378
- 25. Achten NB, Dorigo-Zetsma JW, van der Linden PD, van Brakel M, Plötz FB. Sepsis calculator implementation reduces empiric antibiotics for suspected early-onset sepsis. Eur J Pediatr. 2018;177(5):741–6.
- 26. Giannoni E, Dimopoulou V, Klingenberg C, Navér L, Nordberg V, Berardi A, et al. Analysis of antibiotic exposure and early-onset neonatal sepsis in Europe, North America, and Australia. JAMA Netw Open. 2022;5(11):e2243691. pmid:36416819
- 27. Snell KIE, Levis B, Damen JAA, Dhiman P, Debray TPA, Hooft L, et al. Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses (TRIPOD-SRMA). BMJ. 2023;381:e073538. pmid:37137496
- 28. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–74. pmid:17099194
- 29. Feller D, Wingbermuhle R, Berg B, Vigdal ØN, Innocenti T, Grotle M, et al. Improvements are needed in the adherence to the TRIPOD statement for clinical prediction models for patients with spinal pain or osteoarthritis: a metaresearch study. J Pain. 2024;25(11):104624. pmid:39002741
- 30. Dimopoulou V, Klingenberg C, Navér L, Nordberg V, Berardi A, El Helou S, et al. Antibiotic exposure for culture-negative early-onset sepsis in late-preterm and term newborns: an international study. Pediatr Res. 2025;97(5):1629–35. pmid:39289592
- 31. Kariniotaki C, Thomou C, Gkentzi D, Panteris E, Dimitriou G, Hatzidaki E. Neonatal sepsis: a comprehensive review. Antibiotics (Basel). 2024;14(1):6. pmid:39858292
- 32. Escobar GJ, Puopolo KM, Wi S, Turk BJ, Kuzniewicz MW, Walsh EM, et al. Stratification of risk of early-onset sepsis in newborns ≥ 34 weeks’ gestation. Pediatrics. 2014;133(1):30–6. pmid:24366992
- 33. Kuzniewicz MW, Puopolo KM, Fischer A, Walsh EM, Li S, Newman TB, et al. A quantitative, risk-based approach to the management of neonatal early-onset sepsis. JAMA Pediatr. 2017;171(4):365–71. pmid:28241253
- 34. Debray TPA, Collins GS, Riley RD, Snell KIE, Van Calster B, Reitsma JB, et al. Transparent reporting of multivariable prediction models developed or validated using clustered data: TRIPOD-Cluster checklist. BMJ. 2023;380:e071018. pmid:36750242
- 35. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:q902. pmid:38636956
- 36. Zay Ya K, Win PTN, Bielicki J, Lambiris M, Fink G. Association between antimicrobial stewardship programs and antibiotic use globally: a systematic review and meta-analysis. JAMA Netw Open. 2023;6(2):e2253806. pmid:36757700