MACHINE LEARNING models incorporating MRI-based measures improve the accuracy of predicting neurodevelopmental outcomes in neonates with hypoxic-ischemic encephalopathy (HIE), surpassing the predictive power of demographic and laboratory data alone.
Hypoxic-ischemic encephalopathy (HIE) affects approximately 1.5 infants per 1,000 live births and is a leading cause of both death and neurodevelopmental impairment, such as cerebral palsy, cognitive delays, and speech and language difficulties. Despite advancements in therapeutic hypothermia, which has improved survival rates, many infants still experience significant long-term impairments. Brain MRI is commonly used in clinical settings to support diagnosis and predict long-term outcomes in neonates with HIE, although its interpretation can be subjective, time-consuming, and prone to inter-rater variability. Recent advancements in machine learning and radiomics offer a potential solution to enhance the accuracy of neuroprognostication by incorporating complex, objective data from neuroimaging.
In this study, an MRI template was created using data from 286 neonates who received therapeutic hypothermia. The deep gray matter structures of the brain were labelled, and quantitative data, including shape-related measures and radiomic features, were extracted. These data were then used to train an elastic net machine learning model to predict neurodevelopmental outcomes at 18 months, as assessed by the Bayley Scales of Infant and Toddler Development. Results showed that MRI-based measures alone were able to predict Bayley scores for cognitive, language, and motor outcomes with greater correlation than demographic and laboratory data alone. Combined models using both MRI measures and clinical data provided similar or slightly improved predictive accuracy, reducing the prediction error and explaining a larger proportion of the variance in the outcomes.
The findings indicate that machine learning models using MRI-based measures significantly improve neurodevelopmental outcome predictions for neonates with HIE, with the potential to aid clinical decision-making and enhance prognostication accuracy. This approach can provide a more objective and reliable means of predicting long-term outcomes, which could inform early interventions and treatment planning. Future studies should focus on validating these models across multiple institutions and exploring their integration into routine clinical practice to support healthcare professionals in managing HIE and optimising outcomes for affected infants.
Katrina Thornber, EMJ
Reference
Lewis JD ET AL. Automated neuroprognostication via machine learning in neonates with hypoxic‐ischemic encephalopathy. Annals of Neurology. 2024;DOI:10.1002/ana.27154.
