22 - Predictive Modeling for Death or Severe Neurodevelopmental Impairment (NDI) in Neonates with Hypoxic-Ischemic Encephalopathy (HIE)

Poster Number: 22
Publication Number: 22.302

Background: Outcome after HIE is variable, with ~50% of infants treated with hypothermia having normal neurodevelopment and the remainder sustaining NDI or death at 2 years. Predicting outcome in affected neonates is crucial for guiding clinical management and parent communication.

Objective: To predict death or severe NDI (sNDI) in neonates who receive hypothermia for HIE.

Design/Methods: High-dose Erythropoietin for Asphyxia and Encephalopathy (HEAL) trial participants without ECMO and with EEG were included. The outcome was death/sNDI [Bayley-III cognitive < 70, quadriparesis with Gross Motor Function Classification System (GMFCS)≥1, or hemi/diparesis with GMFCS≥3] at 2 years. Recursive partitioning decision tree prediction models were developed based on data available in the first day (24h Model) or after rewarming (Day 5 Model). Clinical, EEG, and MRI variables were curated for low missingness ( < 10%), high prevalence (>10%), and clinical interpretability. Data were split 70:30 into training and validation datasets with equal proportions of the outcome. Missing MRIs (n=8) were imputed with the worst scores, assuming clinical instability. Missing clinical data were imputed with the median value. For training, a 10-fold cross-validation was used to determine the simplest decision tree resulting in at least one group with >85% death/sNDI. Infants not predicted to have death/sNDI in the 24h Model were included in the Day 5 Model. Model performance [positive, negative predictive value (PPV, NPV)] was assessed in the training and validation datasets.

Results: Among 424 neonates, 59(14%) died and 50(12%) had sNDI (Table 1). In the 24h Model, presence of all three of severely abnormal EEG (burst suppression, flat tracing, or status epilepticus), periodic breathing or intubated/ventilated, and maximum glucose >179 mg/dL (with minimum glucose >54mg/dL) were 99%[98-100%] specific for death/sNDI, with PPV 98%[85-100%] in the training dataset, and 98%[93-100%] specific with PPV 87%[61-97%] in the validation dataset (Figure, Table 2). In the Day 5 Model, diffusion-weighted MRI (DWI) abnormality in the thalamus, caudate, putamen/globus pallidus, and brainstem was 99%[97-100%] specific for death/sNDI with PPV 90%[70-98%] in the training dataset, and 98%[93-100%] specific with PPV 80%[48-95%] in the validation dataset.

Conclusion(s): We show >97% specificity for predicting death/sNDI among newborns with HIE using simple clinical, EEG, and MRI results in a large, multi-center dataset. These models can be used to counsel families about anticipated outcome and goals of care within the first 1-5 days of the hospital admission.