Automated Neuroprognostication via Machine Learning in Neonates with Hypoxic-Ischemic Encephalopathy

Abstract

AbstractObjectivesNeonatal hypoxic-ischemic encephalopathy is a serious neurologic condition associated with death or neurodevelopmental impairments. Magnetic resonance imaging (MRI) is routinely used for neuroprognostication, but there is substantial subjectivity and uncertainty about neurodevelopmental outcome prediction. We sought to develop an objective and automated approach for the analysis of newborn brain MRI to improve the accuracy of prognostication.MethodsWe created an anatomical MRI template from a sample of 286 infants treated with therapeutic hypothermia, and labeled the deep gray-matter structures. We extracted quantitative information, including shape-related information, and information represented by complex patterns (radiomic measures), from each of these structures in all infants. We then trained an elastic net model to use either only these measures, only the infants’ clinical data, or both, to predict neurodevelopmental outcomes, as measured by the Bayley Scales of Infant and Toddler Development at 18 months of age.ResultsAmongst those infants who survived and for whom Bayley scores were available for cognitive, language, and motor outcomes, we found sets of MRI-based measures that could predict their Bayley scores with correlations that were more than twice the correlations based on only the clinical data, and explained more than four times the variance in the observed scores; predictions based on the combination of the clinical and MRI-based measures were similar or marginally better.InterpretationOur findings show that machine learning models using MRI-based measures can predict neurodevelopmental outcomes in neonates with hypoxic-ischemic encephalopathy across all neurodevelopmental domains and across the full spectrum of outcomes.