Beyond visually negative MRI: a clinical–radiomics combined model for identifying hypoxic–ischemic encephalopathy in term neonates with birth asphyxia

Abstract

In term neonates with birth asphyxia but normal visual findings on conventional magnetic resonance imaging (MRI), accurately identifying of hypoxic-ischemic encephalopathy (HIE) remains challenging. Therefore, this study aims to construct and validate a combined clinical-radiomics model for differentiating HIE from non-HIE cases. This retrospective study included 297 term neonates with birth asphyxia treated between January 2020 and December 2025. The cohort was randomly divided into a training set and a test set at a ratio of 7:3. Clinical predictors were identified via univariate and multivariate logistic regression. Regions of interest (ROIs) were manually delineated within the basal ganglia (BG) and thalami (TH) on T1-weighted (T1WI), T2-weighted (T2WI), and diffusion-weighted (DWI) images. Extracted radiomics features were screened to generate six sequence- and ROI-specific radiomics scores (Rad-scores). Five models were developed: a Clinical model, two single-ROI radiomics models (Rad-BG and Rad-TH), a radiomics fusion model (Rad-fusion), and a model integrating clinical predictors with radiomics features (Combined). Model performance was evaluated using receiver operating characteristic (ROC) analysis, area under the curve (AUC), the DeLong test, calibration curves, decision curve analysis, and SHapley Additive exPlanations (SHAP) interpretability analysis. Furthermore, a total of 54 neonates were recruited from the Children’s Hospital of Soochow University and Yulin Children’s Hospital between January and April 2026 to serve as the prospective validation set. The Combined model demonstrated superior performance, yielding AUC values of 0.958 in the training set, 0.915 in the test set and 0.878 in the prospective validation set. The DeLong test confirmed that the Combined model outperformed all of the comparator models. Calibration curves showed acceptable agreement between predicted and observed outcomes, while DCA suggested a potential clinical net benefit. SHAP analysis indicated that both clinical variables and radiomics features contributed to the predictions, with T2WI and DWI features exhibiting the highest importance. The combined model showed superior discriminative performance in identifying HIE in term neonates with birth asphyxia and visually negative MRI findings. As an objective quantitative tool, this model may complement routine clinical assessment and help guide follow-up strategies for this high-risk population.