Objective To investigate the predictive value of a model constructed by combining MRI-based radiomics and clinical risk factors for neonatal white matter injury (WMI).

Methods A total of 154 neonates with WMI admitted to the Second Affiliated Hospital of Shaanxi University of Chinese Medicine from January 2022 to September 2024 were retrospectively enrolled. The Neonatal Behavioral Neurological Assessment (NBNA) was performed within 7 days after birth (for preterm infants, corrected gestational age to term was needed). According to NBNA scores, infants were divided into an early favorable neurodevelopment group (score ≥35, n = 112) and a risk group ( score ＜ 35, n = 42). MRI imaging data and clinical data were collected. After delineating regions of interest, radiomic features were extracted. Optimal features were selected using the Pearson correlation coefficient and least absolute shrinkage and selection operator (LASSO) regression. Radiomics, clinical, and radiomics-clinical combined models were constructed using a random forest algorithm, and performance metrics of each model were calculated in the training and test sets. A multivariable logistic regression model was built based on the selected predictors, and a nomogram was generated.

Results A total of 2,286 radiomics features were extracted; 2,081 features were retained after consistency screening, and 35 key features were further selected using the Pearson correlation coefficient and LASSO regression. Clinical factors were analyzed using univariate and multivariable logistic regression, and hypoproteinemia, respiratory distress syndrome, and neonatal anemia were selected as independent clinical risk factors for model construction. The areas under the curve (AUC) of the radiomics, clinical, and radiomics-clinical combined models in the training and test sets were 0.959 and 0.906, 0.923 and 0.896, and 0.939 and 0.932, respectively. In the test set, the sensitivities of the three models were 60.0%, 72.0%, and 88.0%, and the specificities were 92.1%, 87.1%, and 79.4%, respectively. The DeLong test showed that the difference in AUC between the radiomics-clinical combined model and the clinical model was statistically significant (P = 0.044), the other pairwise comparisons showed no statistically significant differences (P = 0.689, 0.200).

Conclusion The MRI radiomics-clinical integrated model demonstrates good predictive value for early neurodevelopmental risk in neonates with WM, and the nomogram may facilitate clinical risk stratification.