22 - Lab Values in Neonates With Hypoxic Ischemic Encephalopathy Over Time During and After Therapeutic Hypothermia

Publication Number: 1019.22

Michael Elias, University of Florida College of Medicine, MIami, FL, United States; Nikolay Bliznyuk, University of Florida, Gainesville, FL, United States; Daphna Barbeau, Nova Southeastern University Dr. Kiran C. Patel College of Allopathic Medicine, Pembroke Pines, FL, United States; Sarah Sukumar, UF Health Shands Children's Hospital, Gainesville, FL, United States; Juan C. Roig, University of Central Florida College of Medicine, Gainesville, FL, United States; Dhanashree Rajderkar, University of Florida College of Medicine, Gainesville, FL, United States; James L. Wynn, University of Florida College of Medicine, Gainesville, FL, United States; Michael D.. Weiss, University of Florida, Gainesville, FL, United States

Background: Hypoxic-ischemic encephalopathy (HIE) is a leading cause of neonatal neurological injury. However, the relationship between systemic laboratory values (LV) measured during therapeutic hypothermia and the extent of brain injury remains poorly defined.

Objective: To characterize LV trends during and after TH in neonates with HIE, compare groups by the presence or absence of sentinel birth events, and determine their associations with MRI-defined brain injury severity using ROC analysis and machine-learning (ML) models.

Design/Methods: Single-center retrospective cohort of 152 neonates meeting TH criteria (2012-2024 at UF Health). LV were extracted from the electronic records at time points (T1: 0-6 h; T2: 6-12 h; T3: 12-24 h; T4: 24-48 h; T5: 48-72 h; T6: 72-96 h; T7: >96 h), with T1 as baseline for comparison across subsequent time points. Temporal trends were analyzed using linear mixed-effects models, and associations with MRI-defined injury severity (Weeke score) were evaluated using logistic regression. Analyses were stratified by sentinel event status (definite, probable, none). ML models using LV from T1-T3 were developed with Bayesian Information Criterion-based variable selection and five-fold cross-validation, and model performance evaluated using ROC curves and AUC.

Results: ALT increased from baseline (median 28 U/L at T1) and peaked at T2 (64 U/L; p < 0.001), while AST rose from 75 U/L at T1 to a peak of 135 U/L at T2 (p < 0.001) before declining. Lactate peaked at T1 (median 6.8 mmol/L) and decreased through T3 (2.1 mmol/L; p < 0.0001). PT-INR was highest at baseline (T1) and decreased through T3-T4 (p < 0.0001), and D-dimer rose at T3 (p < 0.05) before normalization after rewarming. Lower pH (T1-T2) and greater base deficit (BD) (T1-T3) were associated with higher MRI injury scores across white matter, grey matter, and cerebellar regions (p < 0.05). Subgroup analysis revealed significant differences in pH, pCO₂, WBC, PTT, and D-dimer trends, with infants without a sentinel event. Lactate (T1, AUC 0.80), PT-INR (T1, AUC 0.73), and base deficit (T1, AUC 0.74), were early predictors of MRI-defined brain injury within the T1-T3 timepoints (Figure 1). ML models combining pH (T1), pCO₂ (T1), and lactate (T3) improved prediction accuracy compared with single LV (adjusted R² = 0.47).

Conclusion(s): LV demonstrate temporal patterns during and after TH in neonates with HIE, with early abnormalities correlating with MRI-defined brain injury. Integration of ML analysis may enhance early prognostication and guide personalized neuroprotective care.

ROC Curves for Predictive Biomarkers
Figure 1. Receiver operating characteristic (ROC) curves for early biomarkers at baseline (T1). Lactate (AUC = 0.80), PT-INR (AUC = 0.73), and base deficit (AUC = 0.74). The diagonal dashed line represents the line of no discrimination (AUC = 0.5).