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

To produce objective predictions of neurodevelopmental outcomes after perinatal hypoxic-ischemic insult for the full spectrum of severity of hypoxic-ischemic encephalopathy (HIE) using a machine learning model incorporating only magnetic resonance imaging (MRI)-based radiomic measures.

This was a retrospective cohort study of infants born between January 2018 and January 2022 who experienced HIE. Neonates with a gestational age of ≥35 weeks and a diagnosis of neonatal encephalopathy were treated with therapeutic hypothermia, after which post-rewarming brain MRIs were acquired. At 18 months of age, developmental outcomes were assessed with the Bayley Scales of Infant and Toddler Development. The MRI data were preprocessed, and radiomic measures were extracted from labels covering the entire brain. The radiomic measures together with the outcome measures were provided to an elastic-net penalized linear regression model to predict the outcomes within a 10-fold cross-validation framework.

A total of 167 neonates were included. Across cognitive, language, and motor domains, the mean correlation between the predicted outcomes and the observed outcomes was 0.94, and the mean predictive R-square was 0.87.

A machine learning model using only MRI-based radiomic measures from infants with HIE can reliably predict their 18-month developmental outcomes with high accuracy across motor, cognitive, and language domains, regardless of the severity of their brain injury. In addition, our approach allowed us to produce atlases of the brain regions responsible for the developmental impairments, which may prove useful in the search for novel interventions.