Development and internal validation of an exploratory clinical prediction model for MRI-defined white matter injury in moderate-to-late preterm infants: a single-center retrospective study

Abstract

Background

Moderate-to-late preterm (MLPT) infants (32–36⁺⁶ weeks of gestation) account for approximately 84% of preterm births. Although white matter injury (WMI) may lack obvious early neurological manifestations, it has been associated with adverse neurodevelopmental outcomes. In this cohort, magnetic resonance imaging (MRI)-detected WMI occurred in 20.9% of MLPT infants. Brain MRI is important for WMI detection, but its use for early screening is limited by accessibility and examination timing. Previous WMI prediction models have mainly focused on extremely preterm infants, with limitations in population heterogeneity, sample size, outcome definitions, calibration, and validation. This study aimed to identify clinical predictors of MRI-defined WMI and develop and internally validate a clinical risk prediction model for early risk stratification in MLPT infants.

Methods

This single-center retrospective study included infants born at 32–36⁺⁶ weeks who underwent 3.0-T brain MRI at 7–14 days of life between June 2023 and May 2025 at Jinjiang Municipal Hospital (Shanghai Sixth People’s Hospital, Fujian). Of 208 eligible infants, 7 were excluded because of missing key predictors, resulting in 201 infants. MRI-defined binary WMI was assessed independently by two blinded radiologists according to the Miller grading system. Maternal, delivery-related, and neonatal variables available before MRI were evaluated as candidate predictors. Synthetic Minority Over-sampling Technique (SMOTE) was used to address class imbalance, and recursive feature elimination with cross-validation (RFECV) was applied for feature selection. Multivariable logistic regression and classification and regression tree (CART) models were developed and internally validated using 1,000 bootstrap resampling iterations with optimism correction. Model performance was assessed using discrimination, calibration, Brier score, and classification metrics.

Results

Of the 201 infants included in the final analysis, 42 (20.9%) were diagnosed with WMI, and 159 (79.1%) were classified as non-WMI. RFECV selected 14 candidate predictors for model construction. Following bootstrap optimism correction, the logistic regression model demonstrated an area under the receiver operating characteristic curve (AUC) of 0.62 (95% CI: 0.53–0.70). The model achieved an accuracy of 0.74 (95% CI: 0.68–0.80), sensitivity of 0.00 (95% CI: 0.00–0.09), specificity of 0.94 (95% CI: 0.89–0.98), PPV of 0.09 (95% CI: 0.00–0.41), NPV of 0.78 (95% CI: 0.72–0.83), and a Brier score of 0.18 (95% CI: 0.15–0.22). Calibration assessment showed a calibration intercept of 0.00 (95% CI: −0.42 to 0.44) and a calibration slope of 0.51 (95% CI: 0.27–0.81).

Conclusions

In this study, we developed and internally validated an exploratory prediction model for MRI-defined WMI among MLPT infants using a complete-case cohort from a single center. The logistic regression model showed modest discrimination. However, because the model has only been internally validated within the derivation cohort and has not yet been evaluated in independent external cohorts, it should not currently be used as a stand-alone clinical screening tool. Future studies incorporating larger sample sizes, multicenter external validation, and further optimization of prediction thresholds are warranted to determine its potential clinical utility.

Highlight box.

Key findings

• In a single-center retrospective cohort of 201 moderate-to-late preterm infants, magnetic resonance imaging (MRI)-defined white matter injury (WMI) was identified in 42 (20.9%), indicating that subtle brain abnormalities may occur despite relatively mild clinical presentation.

• An exploratory prediction model was developed using routinely available maternal, perinatal, and early neonatal variables collected before MRI.

• After bootstrap optimism correction, the logistic regression model showed limited discrimination [area under the curve (AUC) 0.62, 95% confidence interval (CI): 0.53–0.70], suggesting that conventional clinical variables alone are insufficient for accurate prediction of MRI-defined WMI.

What is known and what is new?

• Moderate-to-late preterm infants account for the majority of preterm births and are generally considered to have lower neurological risk than extremely preterm infants; however, emerging evidence indicates that subtle white matter abnormalities may still occur during early brain development.

• Existing WMI prediction models have mainly focused on extremely preterm or very low birth weight infants and frequently incorporated severe clinical conditions, limiting their applicability to moderate-to-late preterm populations.

• MRI remains an important tool for identifying subtle brain structural abnormalities that may not be apparent from routine clinical assessment.

What is the implication, and what should change now?

• Clinical variables alone showed insufficient predictive performance, highlighting the need to incorporate additional biomarkers, imaging features, and multicenter data to improve WMI risk stratification in this population.

Introduction

Moderate-to-late preterm (MLPT) infants, typically defined as infants born at 32–36⁺⁶ weeks of gestation, represent approximately 84% of all preterm births worldwide (1). Given their relatively advanced gestational age (GA), MLPT infants often present with fewer acute clinical complications than extremely preterm infants (EPIs), and neonatal care has traditionally emphasized respiratory adaptation, feeding establishment, hyperbilirubinemia, and short-term morbidity. Nevertheless, accumulating evidence indicates that neurodevelopmental outcomes in MLPT infants are not comparable to those of term-born infants, as a subset of these children remains vulnerable to adverse outcomes, including motor impairment, cognitive deficits, language delay, and behavioral problems (2-4). Consequently, early identification of infants at risk for neurological injury within this population is essential for improving neonatal assessment, guiding individualized management, and facilitating timely follow-up interventions.

Previous studies have demonstrated that the overall incidence of white matter injury (WMI) in MLPT infants ranges from approximately 21% to 24% (5). Consistent with previous findings, our single-center cohort identified magnetic resonance imaging (MRI)-detected WMI in 20.9% of MLPT infants, highlighting that occult brain injury remains a clinically relevant concern in this population (6). Notably, the term “preterm brain injury” represents a broad spectrum of pathological conditions, including intraventricular hemorrhage (IVH), hypoxic–ischemic encephalopathy, periventricular white matter abnormalities, and WMI. Importantly, isolated punctate white matter hyperintensities on MRI should be regarded as white matter signal abnormalities, whereas WMI specifically refers to structural white matter lesions defined according to the Miller MRI grading criteria, including focal, multifocal, or patchy pathological changes. These distinct entities may carry different risks for long-term neurodevelopmental impairment (7). To date, research on WMI in preterm infants has predominantly focused on EPIs, extremely low birth weight (BW) infants, or critically ill preterm infants, while evidence specifically addressing MLPT infants remains insufficient (8).

The development of WMI is closely associated with the unique characteristics of cerebral white matter maturation in preterm infants and may also be influenced by multiple factors, including perinatal inflammatory responses, oxidative stress, hypoxic–ischemic exposure, respiratory and circulatory instability, and immature cerebral blood flow regulation. Compared with severe brain injuries, WMI may present with more subtle clinical manifestations. Some affected infants may lack apparent neurological symptoms during the neonatal period, making early detection through routine physical examination challenging (6). MRI is capable of identifying punctate, multifocal, or patchy abnormal white matter signal changes (7) and serves as an important imaging modality for detecting early structural abnormalities of cerebral white matter. Therefore, risk assessment based on MRI-defined WMI may provide a more accurate understanding of the risk of brain injury in MLPT infants. Previous studies have suggested that GA, BW, perinatal infection, respiratory disorders, hypoxic-ischemic exposure, and maternal pregnancy-related factors may be associated with white matter abnormalities in preterm infants (9,10). However, considerable heterogeneity exists among studies regarding study populations, imaging diagnostic criteria, selection of candidate variables, and statistical approaches. More importantly, the development of WMI is rarely attributable to a single factor; rather, it represents the cumulative effect of multiple interacting factors, including maternal background characteristics, delivery-related events, neonatal adaptation after birth, and early postnatal disease burden. Therefore, reliance solely on individual clinical indicators or conventional univariate analyses may be insufficient for accurate risk stratification and early identification of WMI among MLPT infants.

Existing WMI prediction models have several notable limitations related to population bias, as the majority of these models were developed in EPIs (GA <32 weeks) or extremely low BW infants (10,11). These models typically incorporate high-risk clinical indicators associated with severe illness, such as severe asphyxia, severe neonatal respiratory distress syndrome (NRDS), and multiple organ dysfunction syndrome (10). Although these cohorts generally provide sufficient numbers of outcome events for model development, the variable profiles derived from such critically ill populations may not be applicable to MLPT infants, who usually experience relatively milder complications (12). A limited number of small-sample prediction studies focusing on MLPT infants share several common methodological limitations. First, limitations in outcome ascertainment: many studies have relied primarily on cranial ultrasound (CUS) for WMI diagnosis. However, the sensitivity of CUS for detecting non-cystic and diffuse WMI is substantially lower than that of MRI (7,13). The lack of a unified MRI-based gold standard definition of WMI may therefore result in systematic underdiagnosis and outcome misclassification bias. Second, limitations in variable selection strategies: most studies have selected predictors based on univariate screening followed by stepwise logistic regression. Such approaches are inadequate for identifying potential interactions and nonlinear relationships among variables. Third, incomplete validation strategies: many models have relied solely on simple random splitting of the dataset into training and testing sets for internal validation, without applying bootstrap correction to address model overfitting. Furthermore, essential performance metrics, including calibration curves and Brier scores, have often not been fully reported, leaving the external generalizability of these models uncertain.

In clinical practice, information regarding maternal pregnancy characteristics, delivery-related events, neonatal adaptation after birth, and early postnatal complications can usually be obtained before an MRI examination. These data are routinely collected from medical records and demonstrate good clinical accessibility. If these variables can be systematically integrated to develop a prediction model based on pre-MRI clinical information, such a model may facilitate early identification of WMI risk, optimization of neuroimaging assessment strategies, and subsequent neurodevelopmental follow-up studies in MLPT infants. Furthermore, the reliability of a prediction model depends not only on its discriminatory ability but also requires comprehensive evaluation incorporating calibration performance, overall prediction error, threshold-based classification outcomes, and internal validation results.

Based on the above background, this study was conducted using a complete single-center cohort of MLPT infants. MRI-defined WMI was selected as the primary outcome, and maternal characteristics, pregnancy- and delivery-related factors, and early neonatal clinical data available before the MRI examination were incorporated. Machine learning-assisted variable selection methods were applied to identify potential predictors, and two prediction models, including a Logistic regression model and a classification and regression tree (CART) model, were developed. Internal validation was performed using bootstrap resampling. Model performance was systematically evaluated by reporting discrimination, calibration, Brier score, threshold-based classification metrics, and confusion matrices. This study aimed to provide a foundation for early WMI risk stratification in MLPT infants and to support future multicenter external validation studies. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0503/rc).

Methods

Study design and study population

This was a single-center, retrospective prediction model development and internal validation study. The study population consisted of MLPT infants admitted to the Department of Neonatology, Jinjiang Municipal Hospital (Shanghai Sixth People’s Hospital, Fujian), between June 2023 and May 2025. Eligible infants were consecutively enrolled according to a consecutive sampling strategy. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Jinjiang Municipal Hospital (Shanghai Sixth People’s Hospital, Fujian) (No. JJSYYYXLL-2023022). Informed consent was obtained from all patients’ parents or legal guardians.

The inclusion criteria were as follows: (I) GA of 32–36⁺⁶ weeks; (II) admission to the neonatal department within 24 hours after birth; (III) completion of cranial MRI examination during hospitalization; (IV) availability of complete and traceable perinatal and neonatal clinical data.

The exclusion criteria were as follows: (I) congenital structural abnormalities of the brain; (II) inherited metabolic disorders or chromosomal abnormalities; (III) congenital infections; (IV) severe intracranial hemorrhage; (V) severe congenital heart disease or multiple congenital malformations; and (VI) missing key data for candidate predictors.

Data collection

A total of 208 infants were initially screened. Seven infants were excluded due to missing data for candidate predictors, resulting in a final cohort of 201 infants with complete data for analysis. Based on cranial MRI findings during hospitalization, the complete-case cohort was divided into the WMI group (n=42) and the non-WMI group (n=159). All candidate predictors were restricted to variables available before the MRI examination.

MRI acquisition and outcome definition

All enrolled infants underwent cranial MRI examination at 7–14 days after birth once their vital signs were clinically stable. This time window was selected according to the optimal evaluation period recommended by neonatal cranial MRI expert consensus. An MRI performed too early (<7 days) may result in diffusion-weighted imaging (DWI) signals related to acute edema obscuring chronic white matter abnormalities, whereas examinations performed too late (>14 days) may increase the risk of missed diagnosis due to attenuation of subtle punctate lesion signals. A standardized imaging time window was therefore applied to ensure consistency in radiological assessment (14). MRI examinations were performed using a Siemens Lumina 3.0T MRI system equipped with a dedicated neonatal head coil. The imaging protocols included T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), DWI, and fluid-attenuated inversion recovery (FLAIR) sequences. The main imaging parameters were as follows: T1WI: repetition time (TR) =2,000 ms, echo time (TE) =20 ms; T2WI: TR =4,000 ms, TE =100 ms; DWI: b value =1,000 s/mm².

MRI-defined WMI was defined as abnormal signal intensity in the periventricular or deep cerebral white matter on MRI, including punctate white matter abnormalities, multifocal white matter abnormalities, and patchy white matter signal changes. According to the MRI-based grading criteria proposed by Miller et al., mild WMI was defined as focal punctate hyperintensities involving ≤3 lesions or lesions <2 mm in size; moderate WMI was defined as multifocal punctate or patchy hyperintensities involving >3 lesions or lesions >2 mm in size; and severe WMI was defined as lesions involving ≥5% of the periventricular white matter in each cerebral hemisphere. Because no universally accepted MRI grading system for WMI in MLPT infants has been established to date, WMI severity was converted into a binary outcome variable in this study. Infants with mild, moderate, or severe WMI were all classified as WMI-positive, while infants without these MRI abnormalities were classified as WMI-negative.

Two radiologists with more than 5 years of experience in neonatal neuroimaging independently reviewed all MRI scans. During the image interpretation process, the radiologists were blinded to the infants’ clinical predictor variables and study group assignments. A random 30% sample of MRI scans was selected for repeat assessment, and interobserver agreement was evaluated using Cohen’s kappa coefficient. The interobserver agreement for MRI interpretation was excellent (Cohen’s kappa =0.84), confirming good reproducibility and reliability of the imaging assessment. In cases of disagreement, the final diagnosis was determined through consensus review by a third senior neuroradiology specialist.

Candidate predictors

In this study, candidate predictors were prespecified based on the potential pathophysiological mechanisms underlying WMI in MLPT infants and findings from previous relevant studies. The candidate variables covered multiple domains, including maternal characteristics, pregnancy- and delivery-related factors, perinatal factors, and early neonatal clinical conditions, aiming to capture the comprehensive effects of the intrauterine environment, cerebral maturation status at birth, and early postnatal disease exposure on the risk of WMI development.

All candidate predictors were extracted from the hospital electronic medical record system, the neonatal intensive care unit (NICU) database, and perinatal clinical records. The timing of variable collection was restricted to information available before the MRI examination to ensure the clinical applicability of predictors at the intended prediction time point and to avoid information leakage caused by obtaining variables after outcome occurrence. (Details are provided in Appendix 1, Figures S1,S2, and Tables S1-S4.)

Maternal and pregnancy-related variables

Maternal perinatal characteristics were collected, including the following variables:

1. Maternal age: defined as the mother’s actual age at the time of delivery (years), reflecting age-related maternal pregnancy risk factors.

2. Maternal height and pre-pregnancy weight: maternal height (cm) and pre-pregnancy weight (kg) were obtained from the initial antenatal care records and were used to evaluate baseline maternal physical status.

3. Gravidity and parity: gravidity was defined as the total number of pregnancies, including previous and current pregnancies. Parity was defined as the number of pregnancies reaching the GA at which delivery was considered viable.

4. Assisted reproductive technology (ART): determined based on maternal medical history records, including assisted reproductive methods such as in vitro fertilization and embryo transfer. ART was coded as a binary variable (yes/no).

5. Premature rupture of membranes (PROM): defined as spontaneous rupture of the fetal membranes before the onset of regular uterine contractions. PROM status was determined based on obstetric medical records and coded as a binary variable (yes/no).

6. Chorioamnionitis: diagnosed based on obstetric records and relevant laboratory findings, including maternal clinical signs of infection, elevated inflammatory markers, or placental pathological evidence of infectious/inflammatory changes. Chorioamnionitis was coded as a binary variable (yes/no).

Delivery and perinatal variables

The following delivery and perinatal characteristics were collected:

1. Mode of delivery: based on delivery records, the mode of delivery was categorized as vaginal delivery or cesarean section and included as a categorical variable in the analysis.

2. Number of previous cesarean deliveries: the number of previous cesarean deliveries was obtained from maternal obstetric history records and analyzed as either a continuous or categorical variable.

3. GA: defined as the completed gestational weeks at birth and expressed in weeks. GA was determined based on the last menstrual period, early pregnancy ultrasonography findings, and obstetric assessment.

4. BW: defined as the infant’s first recorded body weight after birth and expressed in grams (g).

5. Sex: determined from the birth medical records and categorized as male or female.

6. 1-minute Apgar score: obtained from the Apgar score recorded at 1 minute after birth and used to reflect the infant’s early postnatal adaptation status.

Early neonatal clinical variables

All neonatal disease-related variables were defined as clinical conditions diagnosed before the MRI examination. The following variables were collected:

1. NRDS: diagnosed based on a combination of clinical manifestations, chest imaging findings, and the requirement for respiratory support.

2. Neonatal pneumonia: determined according to infection-related clinical manifestations, laboratory indicators [such as C-reactive protein (CRP) and procalcitonin (PCT)], imaging findings, and clinical diagnosis.

3. Mechanical ventilation: recorded as whether the infant received endotracheal intubation and mechanical ventilation from birth until the time of MRI examination.

4. Neonatal infection-related indicators: infection-associated clinical data available before the MRI examination were collected, including CRP, PCT, and blood culture results, and were analyzed according to clinical records.

Statistical analysis

All statistical analyses were performed using R software (version 4.4.1). Data preprocessing, model development, internal validation, and visualization were primarily performed using the R packages pROC, rpart, rpart.plot, broom, splines, and tidyverse.

The study data were initially subjected to data cleaning, variable verification, and assessment of missingness. Among the original screened cohort of 208 infants, 7 were excluded due to missing data for candidate predictors, resulting in 201 infants included for model development and internal validation. In the complete-case analysis cohort, 42 infants were diagnosed with MRI-defined WMI, whereas 159 infants were classified as non-WMI. The primary outcome was MRI-defined WMI, treated as a binary outcome according to the presence or absence of WMI. All candidate predictors consisted of clinical variables available before the MRI examination, including infant birth and perinatal characteristics, neonatal clinical indicators, and maternal pregnancy- and delivery-related factors. Missing data were handled using complete-case analysis, and no multiple imputation was performed.

Baseline characteristics were described according to WMI status. Continuous variables were presented as mean ± standard deviation or median (interquartile range), depending on their distribution, while categorical variables were presented as numbers and percentages. The normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed variables were compared between groups using the independent-samples t-test, whereas non-normally distributed variables were analyzed using the Mann-Whitney U test. Categorical variables were compared using the Pearson chi-square test or Fisher’s exact test, as appropriate. In addition to P values, standardized mean differences (SMDs) were calculated to evaluate differences in variable distributions between groups. All statistical tests were two-sided, and a P value <0.05 was considered statistically significant.

During the preliminary exploratory variable selection stage, near-zero variance variables were first removed. Considering the imbalance in the distribution between the WMI and non-WMI groups, the Synthetic Minority Over-sampling Technique (SMOTE) was applied only during the variable selection process to address class imbalance. Subsequently, recursive feature elimination with cross-validation (RFECV) based on random forest algorithms was performed for predictor selection, and the retained variables were determined using cross-validation. It should be noted that SMOTE and RFECV were used exclusively for exploratory variable selection and were not applied for final model parameter estimation, model performance evaluation, or bootstrap internal validation. In the revised analysis, the final predictor set was prespecified and fixed before model construction, and variable selection was not repeated during bootstrap resampling for internal validation.

Given the limited number of WMI events in this study, a random split into training and testing datasets may have resulted in an insufficient number of outcome events in the validation subset. Therefore, in the revised analysis, the previous 7:3 training/testing dataset split was no longer applied. Instead, all 201 complete cases were used for final model development, and internal validation was performed using bootstrap resampling. The final model incorporated 14 predictors that were retained through the preliminary RFECV procedure and prespecified before the reanalysis. Continuous variables were entered into the model using their original measurement scales whenever appropriate. Because the association between maternal height and the outcome did not satisfy the linearity assumption, maternal height was modeled as a quadratic term, expressed as [height (m)]².

Before model construction, the functional form of continuous variables and potential multicollinearity were assessed. For continuous predictors, the goodness of fit between linear terms and natural spline functions was compared. The presence of significant nonlinear associations was evaluated using the Akaike information criterion (AIC) and likelihood ratio tests. Multicollinearity was assessed using the variance inflation factor (VIF) and tolerance, where VIF =1/(1 − R²) and tolerance =1/VIF.

Based on the prespecified predictor set, two multivariable prediction models were developed: a multivariable Logistic regression model and a CART model. The Logistic regression model was used to estimate the individual predicted probability of WMI occurrence. The complete regression equation, intercept, regression coefficients, variable coding schemes, and formula for calculating individual predicted probabilities were reported. The CART model was developed using the Gini impurity index as the splitting criterion. Ten-fold cross-validation was applied to determine the optimal pruning parameter, and the final tree structure was selected according to the one-standard-error rule.

The classification threshold of the models was prespecified as 0.50 before analysis and was not optimized post hoc based on the study data. A predicted probability ≥0.50 was classified as predicted WMI, whereas a predicted probability <0.50 was classified as predicted non-WMI. Model performance was evaluated from four aspects: discrimination, calibration, overall prediction error, and threshold-based classification performance. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Threshold-based classification metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the confusion matrix. Calibration performance was evaluated using calibration curves, calibration intercepts, and calibration slopes. Overall prediction error was assessed using the Brier score. All performance metrics and confusion matrix results were reported with 95% confidence intervals (CIs).

Internal validation of the prediction models was performed using 1,000 bootstrap resampling iterations. In each bootstrap iteration, the Logistic regression model and CART model were refitted using the bootstrap sample, and model performance was evaluated both on the bootstrap sample and on the original cohort of 201 infants. Optimism bias was defined as the difference between the model performance estimated in the bootstrap sample and the performance of the corresponding bootstrap-fitted model when applied to the original dataset. The optimism-corrected performance estimates were calculated by subtracting the mean optimism bias from the apparent performance estimates. The bootstrap percentile method was used to estimate the 95% CIs for all performance metrics and confusion matrix counts. The final model results were presented using the complete prediction equation, coefficient table, calibration curves, confusion matrices, and bootstrap optimism-corrected performance estimates.

Results

Basic characteristics of the study subjects

The baseline characteristics of all study participants are presented in Table 1. During the study period, a total of 208 MLPT infants were screened. Seven infants were excluded due to missing candidate predictor data, resulting in a final cohort of 201 infants with complete data. Among them, 42 infants (20.9%) were classified into the WMI group, and 159 infants (79.1%) were classified into the non-WMI group. No statistically significant differences were observed between the two groups in baseline variables, including maternal age, maternal height, maternal weight, GA, BW, sex, mode of delivery, and ART (P>0.05). SMDs were additionally reported to assess differences in variable distributions between the two groups.

Table 1. Baseline maternal and neonatal characteristics according to WMI status.

Variable	Non-WMI group (n=159)	WMI group (n=42)	P value	SMD
Maternal characteristics
Maternal age (years)	30.20 (5.18)	29.98 (5.75)	0.82	0.04
Maternal height (m)	1.59 (0.06)	1.58 (0.06)	0.41	0.15
Maternal weight (kg)	67.47 (11.21)	68.52 (13.28)	0.64	0.09
Mode of delivery			0.11	0.30
Elective cesarean section	117 (73.6)	25 (59.5)
Vaginal delivery	42 (26.4)	17 (40.5)
Assisted reproductive technology			0.77	0.12
No	147 (92.5)	40 (95.2)
Yes	12 (7.5)	2 (4.8)
Neonatal characteristics
Gestational age (weeks)	34.59 (1.41)	34.41 (1.38)	0.47	0.12
Birth weight (g)	2,315.00 (461.53)	2,372.26 (405.76)	0.43	0.13
Sex			>0.99	0.02
Male	96 (60.4)	25 (59.5)
Female	63 (39.6)	17 (40.5)

Data are presented as mean (SD) for continuous variables and n (%) for categorical variables. P values indicate comparisons between the non-WMI and WMI groups. Absolute SMD values are reported. SD, standard deviation; SMD, standardized mean difference; WMI, white matter injury

Feature selection results

Recursive feature elimination (RFE) was applied to select candidate predictors. The results showed that the overall model accuracy gradually increased with the inclusion of additional variables and reached the highest level when 14 variables were retained (Figure 1). The variable importance ranking indicated that mode of delivery was the most influential predictor. In contrast, maternal height and neonatal pneumonia demonstrated relatively lower predictive importance.

Figure 1.

Model performance during recursive feature elimination and importance ranking of selected variables. The upper panel illustrates model accuracy across different numbers of variables during RFE. The lower panel shows the relative importance ranking of variables retained in the final feature set. RFE, recursive feature elimination.

Comparison of model performance

The apparent receiver operating characteristic (ROC) analysis based on the entire cohort demonstrated that the Logistic regression model achieved an AUC of 0.73, whereas the CART model showed an AUC of 0.50. The ROC curve of the Logistic regression model remained above the diagonal reference line across multiple threshold ranges, indicating a certain degree of discriminative ability. In contrast, the ROC curve of the CART model was largely overlapped with the diagonal reference line, suggesting limited discriminatory performance. Considering both model interpretability and apparent discrimination, the Logistic regression model was selected as the primary prediction model, while the CART model was retained as an exploratory comparative model (Figure 2).

Figure 2.

Full-sample apparent ROC curves for the logistic regression and decision tree models. The ROC curves were generated using the full analysis cohort of 201 complete cases. The orange line represents the logistic regression model, and the blue line represents the decision tree model. The dashed diagonal line indicates the reference line for no discrimination. AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic.

Figure 3 summarizes the bootstrap-corrected performance, calibration, and classification results of the logistic regression model. After optimism correction using bootstrap internal validation, the logistic regression model achieved an AUC of 0.62 (95% CI: 0.53–0.70) and an accuracy of 0.74 (95% CI: 0.68–0.80). At the prespecified classification threshold of 0.50, the sensitivity was 0.00 (95% CI: 0.00–0.09), specificity was 0.94 (95% CI: 0.89–0.98), PPV was 0.09 (95% CI: 0.00–0.41), NPV was 0.78 (95% CI: 0.72–0.83), and the Brier score was 0.18 (95% CI: 0.15–0.22). The calibration intercept was 0.00 (95% CI: −0.42–0.44), and the calibration slope was 0.51 (95% CI: 0.27–0.81). The confusion matrix showed 2 true-positive cases (95% CI: 1–14), 40 false-negative cases (95% CI: 28–41), 5 false-positive cases (95% CI: 2–21), and 154 true-negative cases (95% CI: 138–157). For the pruned CART model, the optimism-corrected AUC was 0.48 (95% CI: 0.33–0.50), with an accuracy of 0.78 (95% CI: 0.69–0.84). The sensitivity was 0.00 (95% CI: 0.00–0.00), indicating that the CART model failed to establish a stable positive classification structure at the prespecified threshold.

Figure 3.

Bootstrap-corrected performance and confusion matrix of the logistic regression model. (A) Model performance after correction for overfitting using bootstrap internal validation, including AUC, sensitivity, specificity, PPV, NPV, and Brier score. Points indicate estimates, and horizontal lines indicate 95% CIs; the dashed vertical line denotes the prespecified classification threshold of 0.50. (B) Calibration estimates of the model, including calibration intercept and calibration slope. Points indicate estimates, and horizontal lines indicate 95% CIs. (C) Confusion matrix at the 0.50 classification threshold. Values in each cell indicate the count and bootstrap 95% CI. AUC, area under the receiver operating characteristic curve; CI, confidence interval; FN, false negative; FP, false positive; NPV, negative predictive value; PPV, positive predictive value; TN, true negative; TP, true positive; WMI, white matter injury.

Figure 4 presents the estimated odds ratios (ORs) and 95% CIs for the variables included in the multivariable logistic regression model. Multivariable logistic regression analysis showed that gravidity was associated with an increased probability of WMI occurrence (OR =1.60, 95% CI: 1.05–2.49), whereas parity was associated with a lower probability of WMI occurrence (OR =0.41, 95% CI: 0.17–0.93). Vaginal delivery (OR =2.73, 95% CI: 0.78–9.83) and NRDS (OR =2.40, 95% CI: 0.98–5.96) showed a trend toward increased risk; however, their 95% CIs crossed 1, indicating no statistically significant associations. The 95% CIs of the 1-minute Apgar score, female sex, neonatal pneumonia, number of previous cesarean deliveries, maternal weight, BW, and maternal age also crossed 1. In contrast, GA, PROM, and maternal height squared term showed ORs below 1, but none reached statistical significance.

Figure 4.

Logistic regression coefficients for predictors of WMI. Odds ratios are shown on a logarithmic scale. Points indicate odds ratio estimates, and horizontal lines indicate 95% CIs. The dashed vertical line represents an odds ratio of 1.0. Values above 20 were clipped for display. CI, confidence interval; NRDS, neonatal respiratory distress syndrome; PROM, premature rupture of membranes; WMI, white matter injury.

Discussion

Main findings of the study

This study developed and internally validated a prediction model for MRI-defined WMI in MLPT infants based on 201 complete cases. The study incorporated maternal, pregnancy- and delivery-related, and early neonatal clinical variables that were available before the MRI examination. Two prediction models were developed: a Logistic regression model and a CART model. Model stability was evaluated using bootstrap resampling. In this cohort, 42 infants were diagnosed with MRI-defined WMI, accounting for 20.9% of the study population. This incidence was comparable to the previously reported prevalence of approximately 21–24% among MLPT infants, suggesting that although MLPT infants have GAs relatively close to term, their white matter maturation may still be vulnerable to perinatal environmental influences (14,15), resulting in structural abnormalities detectable by MRI. In recent years, with improved survival rates among EPIs, the pattern of preterm brain injury has gradually shifted from the traditional severe cystic periventricular leukomalacia (cPVL) to predominantly non-cystic white matter abnormalities. These mild or occult white matter abnormalities may still reflect disrupted early brain development and are associated with subsequent neurodevelopmental risks (9,16).

It should be emphasized that the WMI investigated in this study represents a specific imaging-defined type of WMI and should not be considered equivalent to the broad concept of “preterm brain injury” or all types of white matter signal abnormalities detected by MRI. Because early neurological manifestations in MLPT infants are often nonspecific, routine clinical observation alone may be insufficient to identify infants with potential MRI abnormalities. Therefore, exploring risk assessment approaches based on clinical information available before MRI examination has important research implications.

However, the present study also demonstrated that predicting MLPT WMI solely based on routine perinatal clinical variables remains substantially limited. After optimism correction using bootstrap internal validation, the Logistic regression model achieved an AUC of 0.62 (95% CI: 0.53–0.70), indicating limited discriminative ability. At the prespecified classification threshold of 0.50, the model showed a sensitivity of 0.00 (95% CI: 0.00–0.09). Although the model achieved relatively high specificity (0.94, 95% CI: 0.89–0.98), it failed to effectively identify infants with WMI. Therefore, the current model does not meet the requirements for use as an independent clinical screening tool. Instead, it should be considered an exploratory prediction model designed to evaluate the feasibility of integrating routinely available clinical information for predicting MRI-defined WMI risk and to provide a foundation for future multicenter validation studies and further model optimization.

Interpretation of predictive factors

In this study, RFECV was applied to select predictive variables from the candidate predictors, and the final prediction models were developed based on the prespecified fixed predictor set. The variables retained in the final models covered maternal pregnancy characteristics, fetal maturity, delivery-related factors, and early postnatal clinical status, suggesting that MLPT WMI may result from the combined effects of multiple factors occurring across the prenatal, perinatal, and early neonatal periods.

It should be emphasized that the inclusion of variables in the prediction model does not necessarily indicate that these variables have definite causal effects. The objective of this study was to explore which combinations of clinical information may contribute to risk identification, rather than to establish causal relationships. Therefore, the variable importance ranking derived from RFECV and the relative contribution of individual predictors within the model should not be directly interpreted as independent risk factors or as measures of causal effect magnitude.

In the final Logistic regression model, variables including gravidity, parity, mode of delivery, NRDS, and 1-minute Apgar score were incorporated into the prediction framework. These indicators were derived from routine clinical records and represent information available in the clinical setting before the MRI examination, demonstrating good clinical accessibility and interpretability. Among these variables, NRDS may reflect early postnatal respiratory instability, whereas the Apgar score reflects immediate neonatal adaptation status. GA and BW represent fetal maturity and intrauterine growth status, respectively.

From a pathophysiological perspective, the development of WMI is closely associated with the specific vulnerability of the developing preterm brain white matter. Preterm birth occurs during a critical period characterized by rapid brain growth, establishment of neural connections, and active myelination (17). During this stage, white matter structures remain immature and are particularly susceptible to external insults, including hypoxic–ischemic injury, inflammatory responses, and oxidative stress (18). Previous studies have demonstrated that preterm-related WMI involves multiple mechanisms, including hypoxic-ischemic events (19), inflammatory activation (20,21), oxidative stress, and impaired cerebral blood flow regulation. Among these mechanisms, inflammatory responses and hypoxic-ischemic processes may interact synergistically, jointly affecting the development of immature brain tissue (20,22,23).

In the present study, NRDS was incorporated into the prediction model, which may reflect a potential association between early postnatal cardiorespiratory instability and WMI. Infants with NRDS often require oxygen therapy or mechanical ventilation. Fluctuations in oxygenation, changes in carbon dioxide levels, and oxygen exposure may contribute to the development of WMI by affecting cerebral blood flow stability and oxidative stress status (24-26). A lower 1-minute Apgar score may indicate impaired early postnatal adaptation and could be associated with cerebral oxygen delivery and circulatory stability.

Furthermore, GA and BW demonstrated relatively high contributions within the prediction model. Infants with lower GA and lower BW have immature cerebral autoregulation, incomplete oligodendrocyte maturation, and limited antioxidant capacity, making them more vulnerable to perinatal environmental fluctuations and insults (27-29). The inclusion of maternal age, PROM, and chorioamnionitis in the model further suggests that intrauterine conditions and inflammatory exposure may be associated with the risk of WMI. Previous studies have shown that maternal infection and fetal inflammatory responses may influence white matter maturation through multiple pathways, including the release of inflammatory mediators, microglial activation, disruption of blood-brain barrier integrity, and injury to oligodendrocyte precursor cells (OPCs) (30,31).

Overall, this study aimed to develop a prediction model for MRI-defined WMI rather than to establish a causal inference model. The RFECV-based variable importance ranking in the preliminary analysis was primarily used to explore potential predictive factors. These results should not be interpreted as representing the independent effect size of each variable or its etiological contribution. Therefore, this study emphasizes the potential value of integrating multidimensional clinical information for identifying the risk of MRI-defined WMI, rather than considering any single indicator as a decisive factor.

From a clinical perspective, the model findings suggest that risk assessment for WMI in MLPT infants should incorporate multiple aspects, including maternal pregnancy background, fetal maturity, delivery-related factors, and early postnatal clinical conditions, rather than relying on a single clinical indicator for decision-making.

Methodological refinement and internal validation

In this study, the model development process was further standardized and refined. In the preliminary analysis, SMOTE and RFECV were primarily used for exploratory variable selection and handling class imbalance. The final prediction models were developed based on the 201 complete cases, and model parameters were not re-estimated using synthetic samples generated by SMOTE. This strategy minimized the potential influence of artificially generated data on predicted probabilities, calibration performance, and classification metrics, thereby improving the authenticity, interpretability, and clinical relevance of the model results.

Considering the limited number of WMI events in this study, a simple random split into training and testing datasets may have resulted in an insufficient number of events in the validation set, thereby increasing the instability of model performance estimates. Therefore, we developed the models using the entire cohort of complete cases and performed 1,000 bootstrap resampling iterations for internal validation to estimate model optimism bias.

Bootstrap internal validation can improve the stability of model performance evaluation under limited sample conditions; however, it remains an internal validation approach and cannot replace independent external validation. Therefore, future studies should conduct external validation in larger cohorts across different regions and healthcare institutions to further assess the generalizability and clinical applicability of the proposed model.

Model evaluation and clinical application boundaries

In this study, the classification threshold was prespecified as 0.50. This threshold was primarily selected to ensure transparency and reproducibility of the evaluation process and should not be interpreted as a universal clinical decision cutoff. For WMI risk assessment, the selection of an optimal threshold should be determined according to the specific clinical purpose. In the future, if similar models are applied to guide decisions regarding additional MRI assessment or intensified neurodevelopmental follow-up, the balance between the risks of missed diagnosis, unnecessary examinations, parental anxiety, and healthcare resource utilization should be carefully considered. Therefore, a single fixed threshold is unlikely to be applicable across all clinical scenarios. The classification results reported at this threshold in the present study were mainly intended for methodological demonstration and reproducible comparison of model performance.

This study comprehensively evaluated model performance by reporting discrimination, calibration, Brier score, threshold-based classification metrics, and confusion matrices. From a clinical perspective, model evaluation should not rely solely on a single metric such as the AUC; instead, it should incorporate calibration of predicted probabilities, types of misclassification, and potential clinical application scenarios to achieve a more comprehensive interpretation. Decision curve analysis (DCA) may serve as a supplementary approach to describe the potential net benefit of the model across different risk thresholds. However, the results of DCA should still be interpreted cautiously in conjunction with the overall predictive performance of the model.

Based on the current findings, this model cannot replace cranial MRI, neurological examination, or comprehensive clinical assessment. At its present stage, the value of this model lies in systematically integrating perinatal information available before the MRI examination, thereby providing a research foundation for identifying MLPT infants who may require closer monitoring, standardized neuroimaging evaluation, and long-term neurodevelopmental follow-up. If further optimized and validated in larger, multicenter cohorts, such models may provide additional evidence to support risk-stratified management of WMI in MLPT infants in the future.

Strengths and limitations of the study

This study focused on the relatively understudied population of MLPT infants and developed a WMI prediction model based on real-world clinical data. The study clearly defined variable definitions, missing data handling strategies, model development procedures, prespecified classification thresholds, and the complete model equations. In the revised analysis, additional performance metrics, including bootstrap optimism-corrected discrimination, calibration, Brier score, predictive values, confusion matrices, and their corresponding 95% CIs, were reported, providing a more comprehensive and transparent evaluation of model performance.

Several limitations should also be acknowledged. First, this was a single-center retrospective study, and the study population consisted of MLPT infants who underwent MRI examination, which may have introduced selection bias. Second, the number of WMI events in the complete-case cohort was limited, and the number of outcome events per model parameter was relatively low, resulting in potential instability of parameter estimates and an increased risk of overfitting. Third, complete-case analysis excluded seven participants with missing data, which may have introduced additional bias. Fourth, the models underwent only internal validation and require further evaluation through independent validation cohorts, including temporal validation, geographic external validation, or multicenter validation. Fifth, under the prespecified threshold of 0.50, the model demonstrated limited sensitivity and PPV, and therefore is not currently suitable as a standalone clinical screening or decision-making tool. Finally, this study primarily used MRI-defined WMI as the outcome and did not incorporate long-term neurodevelopmental follow-up outcomes at 1–3 years of age to evaluate its predictive value for later functional outcomes.

Conclusions

In conclusion, this study developed and internally validated an exploratory Logistic regression model for predicting MRI-defined WMI in MLPT infants based on 201 complete cases and compared its performance with that of a CART model. After bootstrap optimism correction, the Logistic regression model demonstrated limited discriminative ability, and the calibration slope suggested potential overfitting. At the prespecified threshold of 0.50, the model showed low sensitivity and is therefore not currently suitable as an independent clinical screening or decision-making tool. Further optimization and validation in larger cohorts, multicenter studies, and external validation populations are required before potential clinical application.

Supplementary

The article’s supplementary files as

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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Jinjiang Municipal Hospital (Shanghai Sixth People’s Hospital, Fujian Hospital) (No. JJSYYYXLL-2023022). Informed consent was obtained from all patients’ parents or legal guardians.

Data Sharing Statement

Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0503/dss

tp-15-9-370-dss.pdf ^{(68.2KB, pdf)}

DOI: 10.21037/tp-2026-0503