An interpretable machine learning model for predicting NICU admission in preterm infants: a single-center retrospective cohort study

Introduction

Globally, preterm infants face multiple health risks due to physiological immaturity, leading to high mortality rates and long-term health complications, which impose a significant economic burden on both families and healthcare systems (1,2), particularly in developing countries (3). According to the World Health Organization, approximately 15 million infants are born preterm each year, and this issue continues to worsen worldwide (4). Current management strategies for preterm infants primarily involve clinical monitoring, pharmacologic interventions, and nutritional support; however, limitations persist in the early identification of high-risk infants and the implementation of timely interventions (5). Therefore, developing effective risk prediction tools to improve clinical outcomes in preterm infants is of critical importance.

Previous studies have identified several risk factors associated with adverse outcomes in preterm infants. For instance, low birth weight and short gestational age have been shown to be significantly correlated with poor neonatal outcomes (6-8). These findings have provided valuable insights for clinical practice and highlighted key aspects of preterm infant management, such as prenatal checkups, which allow for the assessment of both maternal and fetal conditions (9). However, most existing studies have focused on the effects of individual factors, lacking a systematic analysis of the combined influence of multiple perinatal and delivery-related variables.

To address this research gap, the present study aimed to comprehensively analyze the combined impact of antenatal and intrapartum variables on the risk of neonatal intensive care unit (NICU) admission among preterm infants. These variables—including birth weight, gestational age, and prenatal checkup status—have been identified as key determinants of neonatal outcomes (10-13). Therefore, to bridge this gap, we conducted a single-center retrospective cohort study with the primary aim of developing and validating an interpretable machine learning (ML) model that integrates readily available prenatal and intrapartum factors to predict NICU admission risk in preterm infants. We employed Shapley Additive exPlanations (SHAP) analysis not only to identify global feature importance but also to provide patient-specific explanations, thereby enhancing clinical trust and utility. To our knowledge, this is the first study to apply the random forest (RF) model and SHAP to interpret NICU admission prediction within a strictly preterm infant cohort in China. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-1-0121/rc).

Methods

Data collection

A total of 2,610 preterm infants born at Gansu Provincial Hospital between January 2020 and December 2023 with a gestational age of <37 weeks were retrospectively enrolled. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Gansu Provincial Hospital (approval No. 2022-007). Informed consent was waived in this retrospective study. The sample size was determined according to the principle that the number of cases should be 10–20 times the number of predictive variables to ensure stable parameter estimation (14,15).

Study population

The inclusion criteria were as follows: (I) live-born infants with a gestational age of ≥28⁺⁰ weeks and <37⁺⁰ weeks; (II) birth weight ≥500 g; (III) completion of basic prenatal checkup information within 72 hours before delivery; (IV) complete records of key prenatal and immediate perinatal variables, including gestational age, birth weight, Apgar scores, mode of delivery, and prenatal checkup status; and (V) delivery at our hospital with follow-up until the first postnatal transfer decision (NICU admission or rooming-in), allowing for a clearly defined NICU admission label.

The exclusion criteria were: (I) major congenital malformations or chromosomal abnormalities; (II) requirement for emergency surgery or intubation immediately after birth; (III) pre-determined immediate postnatal transfer; (IV) >20% missing key variables; and (V) death within 10 minutes after birth. A total of 20 infants were excluded, resulting in a final cohort of 2,610 preterm infants.

Data extraction and processing

Parental perinatal data included: maternal age (years), maternal ethnicity, paternal age (years), paternal ethnicity, number of fetuses, prenatal checkup, maternal height (m), pre-delivery weight (kg), pre-delivery body mass index (BMI) (kg/m²), infectious diseases [hepatitis B, syphilis, acquired immune deficiency syndrome (AIDS), hepatitis C, etc.], concurrent diabetes, anemia, thyroid diseases (hyperthyroidism, hypothyroidism, etc.), concurrent hypertension, autoimmune diseases (systemic lupus erythematosus, rheumatoid arthritis, etc.), glucocorticoid administration, fever, premature rupture of membranes (PROM), duration of PROM (days), gravida, and parity.

Delivery-room immediate indicators included: gender, gestational age (weeks), mode of delivery, birth order, birth weight (kg), twin-twin transfusion syndrome (TTTS), 1-minute Apgar score, 5-minute score, and 10-minute Apgar score, and meconium-stained amniotic fluid (MSAF).

A total of 32 variables from the parental perinatal characteristics and delivery-room immediate indicators categories were included for predictive model feature selection.

Outliers in continuous variables were identified using boxplots. Outliers were defined as values exceeding the upper quartile plus 1.5 times the interquartile range (IQR) or falling below the lower quartile minus 1.5 × IQR. Each outlier was replaced by the nearest of these two bounds to bring it closer to the main data distribution.

Using R version 4.4.2, missing data (<20%) were imputed using linear and logistic regression models. Final imputed values were obtained via nearest neighbor matching within the predictive mean matching (PMM) framework. While this approach is generally robust, it operates under the assumption that the phenotypic patterns present in the training data are representative of the broader target population. This assumption may not hold in all clinical scenarios. Specifically, for neonates presenting with atypical symptoms or rare conditions that are poorly represented in the training cohort, the PMM algorithm may impute values based on the "nearest neighbors" that do not reflect the true physiological state of the individual. This could introduce bias and lead to overly confident but inaccurate predictions. Therefore, model predictions for such atypical cases should be interpreted with caution.

Statistical analysis

Data analysis was performed using R 4.4.2 and the online platform https://work.statpai.com/workbench. The dataset was randomly split into a training set (70%, n=1,827) for model development and an independent test set (30%, n=783) for final model evaluation. To optimize hyperparameters and prevent data leakage, the training set was further partitioned using a 70/30 split, creating a sub-training set (n=1,279) and an internal validation set (n=548). Model hyperparameters were tuned using exhaustive grid search with cross-validation on the sub-training set, with model performance on the validation set guiding the selection of the optimal hyperparameters. The final model was then retrained on the entire training set using these optimal parameters and its performance was assessed on the unseen test set.

Continuous variables with a normal distribution were expressed as mean ± standard deviation (SD) and compared using independent-samples t-tests. Non-normally distributed continuous variables were expressed as median (IQR) and compared using the Mann-Whitney U test. Categorical variables were presented as counts (%) and compared using the χ² test or Fisher’s exact test, as appropriate. Feature selection was conducted using the Boruta algorithm. A two-sided P value <0.05 was considered statistically significant.

Feature selection and hyperparameter optimization

To ensure robustness and interpretability, we implemented a rigorous two-stage feature selection process. First, the Boruta algorithm, which utilizes a cross-validated RF classifier (5-fold), was employed for an all-relevant feature assessment and preliminary screening. Subsequently, the least absolute shrinkage and selection operator (LASSO) regression was applied for further refinement. The optimal regularization parameter (λ) for LASSO was determined via 10-fold cross-validation. We searched for λ on a logarithmic scale and selected the value that produced the most parsimonious model within one standard error of the minimum cross-validated error (“1-SE rule”), favoring a sparser and more interpretable coefficient profile.

Furthermore, to guarantee that all comparative ML models [e.g., RF, XGBoost, light gradient boosting machine (LightGBM), decision tree (DT), artificial neural network (ANN)] were evaluated at their peak performance, a systematic hyperparameter optimization was conducted. For each algorithm, a comprehensive search over a predefined parameter space (e.g., number of trees and maximum depth for RF) was performed using 5-fold cross-validation on the development set. The search employed either an exhaustive grid search or an efficient random search strategy. The specific hyperparameter set yielding the best cross-validated performance metric was subsequently used to train the final model for evaluation.

Model construction and validation

Significant variables were selected using the Boruta algorithm and LASSO regression, and the generalized variance inflation factor (GVIF) was calculated in R. Variables with significant collinearity, defined as GVIF^1/(2·Df) ≥3.16, were removed. Predictive models were constructed in the training cohort using five ML algorithms—DT, RF, LightGBM, XGBoost, and ANN—while the remaining 30% of the data were used as the testing set. A random seed of 123 was set. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) in both the validation and testing sets. Accuracy, sensitivity, specificity, and F1 scores of the optimal model were calculated from the validation set confusion matrix. Calibration curves assessed model consistency, and DCA evaluated clinical net benefit.

Model interpretability

The SHAP method was used to quantify each feature’s contribution and potential interactions. Based on the above evaluation metrics, the optimal diagnostic model was selected, and SHAP feature importance plots, beeswarm plots, and waterfall plot of key variables were generated to provide a visual interpretation of the model.

Subgroup analysis in late preterm infants (34–37 weeks)

To evaluate the model’s performance in a clinically relevant population with greater admission uncertainty, we conducted a subgroup analysis focusing on late preterm infants (gestational age 34–37 weeks). Using the same feature set identified from the full cohort analysis, we re-optimized all five ML models specifically for this subgroup. Hyperparameter tuning was performed using 5-fold cross-validation on the training data, and the best-performing model was selected based on area under the curve (AUC). Model performance was evaluated in both the validation and test sets using the same metrics as the main analysis, including AUC, accuracy, sensitivity, specificity, precision, F1 score, calibration curves, and DCA.

Results

Baseline characteristics

A total of 2,610 preterm infants (<37 weeks’ gestational age) were included, with an NICU admission rate of 51.3%. Infants admitted to the NICU had lower birth weight, shorter gestational age, higher proportions of no prenatal checkups, maternal hypertension, PROM, and low Apgar scores (P<0.05). Maternal age, ethnicity, comorbid diabetes/anemia/thyroid disease, and neonatal sex did not differ significantly between groups (P>0.05) (Table S1).

Feature selection and model evaluation

Using Boruta and LASSO regression (Figure 1) via https://work.statpai.com/workbench, 10 significant variables were identified, and their relationships to Boruta were visualized in a network diagram (Figure 2). Variance inflation factors were calculated in R (Table 1), and variables with GVIF^1/(2·Df) ≥3.16 were excluded due to multicollinearity. All GVIF^1/(2·Df) values were below the recommended threshold of √10, indicating acceptable multicollinearity among the final selected features. The final 10 variables included (Table 1): gestational age, birth weight, 1-minute Apgar score, 5-minute Apgar score, 10-minute Apgar score, paternal ethnicity, prenatal checkup, maternal height (m), mode of delivery, and paternal age (years).

Figure 1 Feature selection with Boruta and LASSO regression. (A) The boxplots show the minimum, average, and max shadow score. (B) The correlation between L1 norm and different coefficients in LASSO regression. L1 norm is the regularisation term for LASSO. (C) The correlation between lambda with binomial deviance. There are two dashed lines in the graph. The left dashed line indicates the minimum mean squared error while the right one indicates one standard error away from the minimum mean squared error. BMI, body mass index; LASSO, least absolute shrinkage and selection operator; MSAF, meconium-stained amniotic fluid; PROM, premature rupture of membranes; TTTS, twin-twin transfusion syndrome.

Figure 2 Network diagram. BMI, body mass index; LASSO, least absolute shrinkage and selection operator; PROM, premature rupture of membranes.

Five ML models—DT, RF, LightGBM, XGBoost, and ANN—were used for prediction. The testing set comprised 30% of the data, with a random seed set to 123. Based on the ROC curves, calibration curves, and DCA from both the validation and testing sets, the RF model was identified as the optimal model. The optimal parameters were identified through automatic adjustment: n_estimators = 300 and max_features = 3. Among the five ML models, the RF model performed the best. Its predicted probabilities exhibited the smallest deviation from actual probabilities on the calibration curve. The AUC for the validation and testing sets were 0.861 [95% confidence interval (CI): 0.830–0.891] and 0.869 (95% CI: 0.841–0.897), respectively. The RF model demonstrated strong predictive performance in both datasets. In the validation set, it achieved an accuracy of 82.3%, sensitivity of 83.0%, precision of 82.7%, and F1 score of 82.8%. In the testing set, the model maintained comparable performance with an accuracy of 82.2%, sensitivity of 79.8%, precision of 82.2%, and F1 score of 82.1%, indicating good discriminative ability for NICU admission risk among preterm infants and suitability for preliminary clinical screening.

DCA further supported its clinical utility, showing that using the RF model to predict NICU admission would provide a positive net benefit across a wide range of clinically reasonable probability thresholds (approximately 0.13 to 0.90).

The confusion matrices for the RF model on both datasets are shown in Figure 3. Based on the testing set matrix (Figure 3B), the model yielded an accuracy of 82.2% (644/783), a sensitivity of 79.8% (320/401), and a specificity of 84.5% (324/383), further confirming its balanced classification performance. This balanced discriminative power was consistent with the high AUC observed in the testing set (0.869, 95% CI: 0.841–0.897). The comprehensive performance evaluation, including ROC curves, calibration plots, and DCA for all five compared models, is consolidated in Figure 4.

Figure 3 Confusion matrices for the RF model. Performance on the validation set (A). Performance on the internal testing set (B). The matrices compare actual versus predicted NICU admission status. NICU, neonatal intensive care unit; RF, random forest.

Figure 4 Performance evaluation of all machine learning models. ROC curve analysis (A,B), calibration curve analysis (C,D), and DCA curves (E,F). ANN, artificial neural network; AUC, area under the curve; DCA, decision curve analysis; DT, decision tree; RF, random forest; ROC, receiver operating characteristic.

Explaining RF models using the SHAP algorithm

SHAP values were used for global interpretation of the RF model, including 50 preterm infants and 10 prenatal and immediate peripartum variables. Figure 5A shows the top 10 features ranked by mean |SHAP| values, with birth weight (mean |SHAP| = 0.17), prenatal checkup (0.13), and gestational age (0.09) contributing most to model output. Birth weight had a substantially higher SHAP contribution than other variables, while Figure 5B shows markedly increased SHAP values in infants without prenatal checkups. Gestational age was ranked third, and all infants below 32 weeks gestation corresponded to high positive SHAP values. Notably, although the 5-minute Apgar score was statistically significant in univariate analysis, it ranked only fifth in the SHAP importance plot, suggesting that the RF model partially captures the physiological information reflected by the Apgar score through the non-linear interaction of birth weight and gestational age—both of which are clinically known to influence Apgar components such as muscle tone and respiratory effort (16,17). Variables such as paternal ethnicity and maternal height had SHAP values close to zero. Figure 5B (beeswarm plot) indicates that lower birth weight is associated with higher SHAP values, highlighting low birth weight as a core driver of NICU admission. Infants without prenatal checkups had SHAP values concentrated above 0.15, suggesting a significantly higher risk of NICU admission. All samples with gestational age <32 weeks were located on the positive SHAP axis, aligning with clinical thresholds (18). Figure 5C presents a representative individual waterfall explanation (baseline risk 51.2%): for this infant, paternal ethnicity was Han, birth weight was 1,195 g, and prenatal checkup was absent; these three factors cumulatively contributed approximately +0.40 as the main positive drivers, while other variables such as maternal height 160 cm had negative contributions. The final model output was consistent with the actual transfer outcome of the infant.

Figure 5 SHAP analysis for the random forest model. (A) Feature importance bar plot showing the mean absolute SHAP values for each predictor. Birth weight (+0.17), prenatal checkup status (+0.13), and gestational age (+0.09) are the top three most important features. (B) SHAP beeswarm plot displaying the distribution of SHAP values for each feature across all patients. Each point represents a single patient prediction. The color indicates the feature value (red: high, blue: low). Features are ranked in descending order of importance. (C) SHAP waterfall plot for an individual patient (f(x) = 0.8), illustrating the contribution of each feature to the model’s output. The base value (E[f(x)] = 0.512) represents the average model prediction over the dataset. Features in red push the prediction higher, while features in blue push the prediction lower. SHAP, Shapley Additive exPlanations.

Subgroup analysis in late preterm infants (34–37 weeks)

Using the feature set derived from the full cohort, the RF model re-optimized for the late preterm subgroup (n_estimators =50, max_features =3) demonstrated robust performance. In the validation set, it achieved an AUC of 0.842 (95% CI: 0.806–0.879), accuracy of 82.8%, sensitivity of 64.4%, precision of 86.2%, and F1 score of 73.7%, with excellent calibration and positive net benefit on DCA (Figure 6A-6C). In the test set, performance remained stable: AUC 0.838 (95% CI: 0.807–0.868), accuracy 81.8%, sensitivity 64.9%, precision 82.8%, and F1 score 72.8%, with good calibration and favorable net benefit (Figure 6D-6F). These findings support the model’s reliability in predicting NICU admission among late preterm infants.

Figure 6 Model performance for the late preterm infant subgroup (34–37 weeks gestation). (A,D) ROC curves for the validation set (A) and test set (D), showing AUC values of 0.842 and 0.838, respectively. (B,E) Calibration curves demonstrating good agreement between predicted and observed probabilities in both the validation (B) and test (E) sets. (C,F) Decision curve analysis showing positive net benefit across a range of threshold probabilities in the validation (C) and test (F) sets, supporting the model’s clinical utility. ANN, artificial neural network; AUC, area under the curve; DCA, decision curve analysis; DT, decision tree; RF, random forest; ROC, receiver operating characteristic.

Discussion

The NICU is a core resource in perinatal medicine, and its bed turnover efficiency directly affects preterm infant survival and quality of life (19). In recent years, with increasing maternal age and the widespread use of assisted reproductive technologies, the absolute number of preterm infants has continued to rise (20,21). Traditionally, NICU admission decisions rely on the on-site judgment of obstetric and neonatal physicians, with decision variables scattered and lacking quantitative basis, leading to potential over-transfer of mild preterm infants and missed treatment opportunities for critically ill infants (22). Therefore, constructing an intelligent predictive model based on prenatal and immediate peripartum information from electronic medical records to achieve “assessment at birth, stratification upon assessment” represents an important breakthrough for improving perinatal service efficiency and equity. This study, focusing on pure preterm infants (28⁺⁰–37⁺⁰ weeks) in a large single-center retrospective cohort, is the first in China to introduce interpretable ML (RF-SHAP) into the NICU admission decision-making process, aiming to maintain high discrimination while providing transparent and actionable clinical decision pathways. More importantly, by leveraging the SHAP framework, our model moves beyond a “black box” prediction, offering transparent and individualized risk assessments that align with clinical reasoning.

This study is the first to apply SHAP analysis to NICU admission decisions in a pure preterm cohort. Results showed that birth weight had the highest SHAP contribution, consistent with studies indicating that extremely low birth weight infants have a higher probability of requiring parenteral nutrition and respiratory support (23,24). Prenatal checkup ranked second, likely because regular checkups allow timely detection of preeclampsia, fetal growth restriction, and other complications, reducing the incidence of emergency delivery and severe prematurity (25,26). Figure 5B demonstrates markedly higher SHAP values for infants without prenatal checkups, suggesting that this threshold could serve as a target for outpatient quality improvement. Gestational age ranked third, and all infants <32 weeks gestation corresponded to high positive SHAP values. Notably, although the 5-minute Apgar score was statistically significant in univariate analysis, it ranked only fifth in the SHAP ranking, indicating that the RF model captures part of the Apgar information through the non-linear interaction of birth weight and gestational age, reflecting the model’s capacity to compress redundant information (27,28). Furthermore, variables such as paternal ethnicity and maternal height had SHAP values close to zero, suggesting minimal influence on NICU admission within preterm infants and potential exclusion in simplified models to improve deployment efficiency.

This subgroup analysis focused on late preterm infants (34–37 weeks), where NICU admission decisions are often uncertain (29). Using the feature set from the full cohort and re-optimizing the RF model for this subgroup, we observed stable performance with a test set AUC of 0.838. The combination of moderate sensitivity (64.9%) and high precision (82.8%) offers a clinically meaningful balance: while some high-risk infants may be missed, positive predictions are highly reliable, supporting resource-limited decision-making. Consistent performance across validation and test sets, along with good calibration and positive net benefit on DCA, supports its clinical potential. These findings highlight the value of population-specific model optimization and confirm the predictive relevance of the selected features even in a more homogeneous subgroup.

The model demonstrates robust predictive performance for NICU admission across the entire preterm cohort, with particular value in the late preterm subgroup (34–37 weeks), where clinical decisions are most uncertain. Its high precision helps minimize unnecessary admissions and conserve neonatal intensive care resources, while its moderate sensitivity underscores the importance of integrating clinical judgment for low-risk predictions. As a decision-support tool, the model provides an objective risk stratification framework to guide individualized triage and resource allocation, especially in settings where admission thresholds are variable.

This study has several limitations. First, its single-center, retrospective design may limit the generalizability of our findings and introduces the potential for selection bias. External validation in multi-center, prospective cohorts is essential to confirm the model’s robustness and transportability. Second, while our model leverages routinely collected clinical variables to ensure broad applicability, future studies may explore the integration of novel biomarkers—such as inflammatory or metabolic indicators—to further enhance predictive performance and offer deeper biological insights into neonatal risk stratification (30). Third, the model’s sensitivity of 79.8% in the testing set indicates that some infants requiring NICU care might be missed if used as a sole screening tool; therefore, it should serve as an adjunct to, not a replacement for, clinical judgment. Future work should focus on: (I) integrating real-time physiological data streams; (II) developing a user-friendly clinical decision support interface for point-of-care use; and (III) evaluating the model’s impact on clinical outcomes and resource utilization in an interventional study.

Conclusions

In conclusion, the prenatal interpretable model developed in this study achieved accurate prediction of NICU admission for preterm infants with an AUC of 0.861. Birth weight, prenatal checkup, and gestational age constituted the core of the decision-making process. By visualizing with SHAP, risk stratification and resource allocation can be performed immediately at birth, providing a reliable and actionable intelligent decision-support tool for perinatal quality improvement.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-1-0121/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-1-0121/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-1-0121/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-1-0121/coif).The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Gansu Provincial Hospital (approval No. 2022-007). Informed consent was waived in this retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Wang Y, Wang S, Dou H, et al. Neuroprotective effects of caffeine, erythropoietin, magnesium sulfate, and thyroxine in preterm infants: a network meta-analysis. Transl Pediatr 2025;14:2640-56. [Crossref] [PubMed]
2. Kvaratskhelia N, Rurua N, Vadachkoria SG. Biomedical and Psychosocial Determinants of Early Neurodevelopment After Preterm Birth. Glob Pediatr Health 2023;10:2333794X231160366.
3. Jafarpour L, Galehdar N, Mokhayeri Y, et al. Maternal risk factors associated with the birth of preterm infants in the West of Iran: a matched case-control study. BMC Pregnancy Childbirth 2025;25:278. [Crossref] [PubMed]
4. Locke A, Kanekar S. Imaging of Premature Infants. Clin Perinatol 2022;49:641-55. [Crossref] [PubMed]
5. Meiliana M, Alexander T, Bloomfield FH, et al. Nutrition guidelines for preterm infants: A systematic review. JPEN J Parenter Enteral Nutr 2024;48:11-26. [Crossref] [PubMed]
6. Chan JKN, Lee KCK, Correll CU, et al. Adverse obstetric and neonatal outcomes associated with maternal schizophrenia-spectrum disorders and prenatal antipsychotic use: a meta-analysis of 37,214,330 pregnancy deliveries and propensity-score weighted population-based cohort study assessing confounder dependency of risk estimates. Mol Psychiatry 2025;30:954-67. [Crossref] [PubMed]
7. Etchecopar-Etchart D, Rahmati M, Yon DK, et al. Maternal and neonatal complications of pregnant women with bipolar disorder: a systematic review and meta-analysis. EClinicalMedicine 2025;79:103007. [Crossref] [PubMed]
8. El-Saie A, Varghese NP, Webb MK, et al. Bronchopulmonary dysplasia - associated pulmonary hypertension: An updated review. Semin Perinatol 2023;47:151817. [Crossref] [PubMed]
9. Wilson RD. Antenatal reproductive screening for pregnant people including preconception: Provides the best reproductive opportunity for informed consent, quality, and safety. Best Pract Res Clin Obstet Gynaecol 2024;96:102520. [Crossref] [PubMed]
10. Zhou H, Yang Y, Chi P, et al. The association of Chinese and American antenatal care utilization indices with birth outcomes. Front Public Health 2024;12:1420943. [Crossref] [PubMed]
11. Aga MA, Woldeamanuel BT. Nutritional and prenatal care factors associated with low birth weight among full-term infants in public hospitals of Addis Ababa, Ethiopia. BMJ Paediatr Open 2025;9:e003732. [Crossref] [PubMed]
12. Song IG. Neurodevelopmental outcomes of preterm infants. Clin Exp Pediatr 2023;66:281-7. [Crossref] [PubMed]
13. Li Q, Zheng Y, Guo X, et al. Extremely Low Birth Weight Infant (Gestational Age of 29 Weeks) With Kabuki Syndrome Type I: Case Report and Literature Review. Mol Genet Genomic Med 2024;12:e70025. [Crossref] [PubMed]
14. Dhiman P, Ma J, Andaur Navarro CL, et al. Methodological conduct of prognostic prediction models developed using machine learning in oncology: a systematic review. BMC Med Res Methodol 2022;22:101. [Crossref] [PubMed]
15. Baeza-Delgado C, Cerdá Alberich L, Carot-Sierra JM, et al. A practical solution to estimate the sample size required for clinical prediction models generated from observational research on data. Eur Radiol Exp 2022;6:22. [Crossref] [PubMed]
16. Qiu MF, Zhang Y, Tang YZ, et al. An investigation into the impact of family integrated care on extrauterine growth restriction at discharge in very low birth weight infants: a multi-centre study. J Glob Health 2025;15:04168. [Crossref] [PubMed]
17. Tamai K, Matsumoto N, Yorifuji T, et al. Delivery room intubation and severe intraventricular hemorrhage in extremely preterm infants without low Apgar scores: A Japanese retrospective cohort study. Sci Rep 2023;13:14990. [Crossref] [PubMed]
18. Tang M, Ibrahim A, Laughon C, et al. Prescribing practices of inhaled corticosteroids for premature infants in the neonatal intensive care unit. J Perinatol 2024;44:953-6. [Crossref] [PubMed]
19. Mominkhan DM, Aldahmashi F, Almudeer AH, et al. Implementing Telemedicine Intervention in Neonatal Intensive Care Units: Augmented Teleconsultation and Real-Time Monitoring Experience. Telemed Rep 2025;6:50-7. [Crossref] [PubMed]
20. Miao H, He H, Nie C, et al. Spatiotemporal Characteristics and Risk Factors for All and Severity-Specific Preterm Births in Southern China, 2014-2021: Large Population-Based Study. JMIR Public Health Surveill 2024;10:e48815. [Crossref] [PubMed]
21. Bu Z, Zhang J, Hu L, et al. Preterm Birth in Assisted Reproductive Technology: An Analysis of More Than 20,000 Singleton Newborns. Front Endocrinol (Lausanne) 2020;11:558819. [Crossref] [PubMed]
22. Saleem S, Tikmani SS, Goudar SS, et al. Neonatal mortality among preterm infants admitted to neonatal intensive care units in India and Pakistan: A prospective study. BJOG 2023;130:68-75. [Crossref] [PubMed]
23. Zhao T, Li A, Wu W, et al. Maternal Racial Differences and Socioeconomic Status in Preterm Infant Neurodevelopment, Feeding, and Growth. Nurs Res 2026;75:88-95. [Crossref] [PubMed]
24. Lee SM, Sie L, Liu J, et al. Evaluation of Trends in Bronchopulmonary Dysplasia and Respiratory Support Practice for Very Low Birth Weight Infants: A Population-Based Cohort Study. J Pediatr 2022;243:47-52.e2. [Crossref] [PubMed]
25. Hasan HM, Thabit MSM, Abdulaziz M. Preeclampsia Among Pregnant Women Admitted to a Tertiary Care Hospital in Taiz City: Insights from a Conflict-Stricken Setting in Southwest Yemen. Int J Womens Health 2026;18:586732. [Crossref] [PubMed]
26. Ybaseta-Medina J, Meza-León N, Munive-Bendezú R, et al. Clinical, social, and occupational determinants of severe preeclampsia: a multifactorial case-control study on maternal health inequities in Peru. BMC Pregnancy Childbirth 2026;26:297. [Crossref] [PubMed]
27. Walker AM, Cliff A, Romero J, et al. Evaluating the performance of random forest and iterative random forest based methods when applied to gene expression data. Comput Struct Biotechnol J 2022;20:3372-86. [Crossref] [PubMed]
28. Fernandez-Lozano C, Hervella P, Mato-Abad V, et al. Random forest-based prediction of stroke outcome. Sci Rep 2021;11:10071. [Crossref] [PubMed]
29. Workalemahu T, Clark EAS, Madsen MJ, et al. Mapping genetic susceptibility to spontaneous preterm birth: analysis of Utah pedigrees to find inherited genetic factors. Am J Obstet Gynecol 2025;232:557.e1-557.e10. [Crossref] [PubMed]
30. Chang AL, Reiss JD, Culos A, et al. Quantitative assessment of neonatal health using dried blood spot metabolite profiles and deep learning. Sci Transl Med 2026;18:eadv4942. [Crossref] [PubMed]