Neuroprognostication via Radiomics and Machine Learning Following Neonatal Hypoxic-Ischemic Insult.

Area of Science:

• Neonatal neurology
• Radiomics
• Machine learning in medicine

Background:

• Perinatal hypoxic-ischemic insult (HIE) can lead to significant neurodevelopmental deficits.
• Accurate prediction of outcomes is crucial for guiding clinical management and therapeutic strategies.
• Current prediction methods may not fully capture the spectrum of HIE severity.

Purpose of the Study:

• To develop a machine learning model for predicting neurodevelopmental outcomes in infants with HIE.
• To utilize only MRI-based radiomic measures for outcome prediction.
• To cover the full spectrum of HIE severity.

Main Methods:

• Retrospective cohort study of 167 infants with HIE (gestational age ≥35 weeks).
• Post-rewarming brain MRIs analyzed for radiomic features.
• Elastic-net penalized linear regression model used for prediction with 10-fold cross-validation.
• Developmental outcomes assessed at 18 months using Bayley Scales.

Main Results:

• High accuracy in predicting outcomes across cognitive, language, and motor domains.
• Mean correlation between predicted and observed outcomes was 0.94.
• Mean predictive R-square was 0.87, indicating strong model performance.

Conclusions:

• MRI-based radiomic features reliably predict 18-month developmental outcomes in HIE infants.
• The model's accuracy is high across all severity levels of brain injury.
• Generated brain atlases highlight regions associated with impairments, potentially guiding novel intervention development.