40 - Brain Magnetic Resonance Imaging–Based Radiomics Risk Prediction for Neurodevelopmental Outcomes in Neonates

Publication Number: 1033.40

Vivek V. Shukla, University of Alabama at Birmingham, Birmingham, AL, United States; Avinash Singh, Children's of Alabama, Birmingham, AL, United States; Sarah Engel, Prisma Health, Greenville, SC, United States; Hannah Vickland, University of Alabama at Birmingham, Irondale, AL, United States; Tao Chen, UAB, Spanish Fort, AL, United States; Stephen A. Walker, Children's of Alabama, Birmingham, AL, United States; Fabricio G. Goncalves, Children's of Alabama, Bryn Mawr, PA, United States; Salman Rashid, University of Alabama School of Medicine, Birmingham, AL, United States; Arie Nakhmani, University of Alabama at Birmingham, Birmingham, AL, United States; Waldemar Carlo, University of Alabama, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States

Background: Prior studies have combined clinical and qualitative brain magnetic resonance imaging (MRI) scores; however, use of quantitative radiomics alongside clinical data for outcome predictions in neonatal hypoxic-ischemic encephalopathy (HIE) remains limited. Developing models integrating multimodal data may enable early, accurate prediction of severe neurodevelopmental outcomes in neonates with moderate-severe HIE, thereby guiding timely, targeted interventions.

Objective: Test the hypothesis that radiomics-based analysis of brain MRI performed within the first two weeks after birth can predict mortality or moderate to severe neurodevelopmental disability by 18-24 months in neonates with HIE.

Design/Methods: Neonates (≥35 weeks gestation) with moderate-to-severe HIE who underwent brain MRI within 2 weeks were included. Neurodevelopmental outcomes at 18-24 months were evaluated using the Bayley-III, following Neonatal Research Network (NRN) therapeutic hypothermia definitions. Axial T1, T2, DWI, and ADC images underwent threshold-based brain masking and 128×128 standardization. From 5,368 analyzable scans, 147 radiomics features were extracted in SimpleITK and averaged per participant. Clinical variables included perinatal data. LASSO feature selection, and classification models including logistic regression, random forest, XGBoost, histogram gradient boosting, decision tree, and SVM models were trained using nested 5-fold cross-validation. AUC was primary; accuracy, precision, recall, precision-recall (PR) AUC, F1, Brier score, and SHAP were secondary metrics.

Results: Among 70 neonates, 15 (21.4%) had death or disability. The radiomics-only XGBoost model achieved an AUC of 0.726, a PR AUC of 0.405, and a Brier score of 0.219 (Table 1, Figure 1), with top predictive features including canny_edge_count, sphericity, mean_intensity, entropy, and contrast. The clinical-only Random Forest model achieved an AUC of 0.867, a PR AUC of 0.725, and a Brier score of 0.129, with top predictive features including discharge neurologic examination, early-intervention history, bicarbonate infusion, encephalopathy score, and birthweight. The multimodal logistic regression model achieved an AUC of 0.875, a PR AUC of 0.715, a Brier score of 0.107, with top predictive features including image contrast, mean absolute deviation, mean intensity, discharge neurologic examination, and encephalopathy score (Figure 2).

Conclusion(s): Brain MRI-based radiomics and multimodal models can meaningfully predict neurodevelopmental outcomes in neonates with HIE, supporting early identification of high-risk infants for targeted intervention.

Table 1 Performance Comparison of Top-Performing Models


Figure 1: AUC Curves of the Top Performing Models


Figure 2: SHAP Analysis of the Top Predictive Multimodal Model (Logistic Regression)
Left panel ranks features by mean absolute SHAP value. Right panel displays SHAP summary plots in which each dot represents a case (red = high feature value; blue = low feature value) and horizontal position indicates whether the feature value increases (right) or decreases (left) predicted risk of a severe outcome.