Background: Hypoxic-ischemic encephalopathy (HIE) is a severe brain injury that occurs in neonates due to perinatal oxygen deprivation, often leading to adverse neurodevelopmental outcomes or death. The therapeutic window for HIE is within the first 6 hours of life but implementing therapeutic hypothermia up to 12 hours after birth has been shown to be effective in reducing the severity of brain injury. To ensure the timely and effective recognition of HIE during this critical therapeutic window, it is essential to use objective methods for diagnosis. Near-infrared spectroscopy (NIRS) provides a non-invasive continuous regional measurement of cerebral oxygenation.
Objective: We have sought to assess the potential clinical utility of NIRS as an additional tool in the diagnosis of HIE.
Methods: We analysed 53 infants with all grades of HIE (>36 weeks GA) enrolled in the Multimodal Assessment of Newborns at Risk of Neonatal Hypoxic Ischaemic Encephalopathy (Monitor) trial. All infants had continuous cerebral oxygenation monitoring for at least 2 hours in their first 12 hours after birth. HIE was graded (mild, moderate, severe) based on assessment using the modified Sarnat score at 1 hour of life. The NIRS signals recorded in the first 12 hours of life were pre-processed, and quantitative features were extracted. Furthermore, prolonged relative desaturations (PRDs; data-driven desaturations lasting 2-15 minutes) were identified and removed from NIRS signals, termed filtered NIRS. The quantitative features were combined in a machine-learning model using a leave-one-out cross-validation approach to determine the likelihood of requiring hypothermia treatment, distinguishing between mild vs moderate and severe HIE. We used logistic regression models to control for the potentially confounding effects of clinical features on the NIRS machine-learning model. We controlled for Apgar score (5min) and mode of delivery for the NIRS and filtered NIRS models for detecting mild HIE. In all models, the significance level was set at p < 0.05.
Results: Logistic regression analysis revealed that features extracted from NIRS were significant predictors of requiring hypothermia in this population (β = 0.61, p = 0.01). Furthermore, features extracted from filtered NIRS were found to be significant predictors of mild HIE (β = 0.72, p < 0.001). The predictability of the Apgar score when assessed independently was significant (β = -0.11, p < 0.001), while the mode of delivery did not demonstrate a significant impact. The regression model, which included filtered NIRS, Apgar score, and delivery mode, accounted for 50.4% of the variance in the outcome variable (R-squared = 0.504), with the root mean squared error (RMSE) of 0.36. This model performed better than both the logistic model based on filtered NIRS and Apgar score (R-squared = 0.48, RMSE = 0.37) and the model based on Apgar score and delivery mode (R-squared = 0.41, RMSE = 0.40).
Conclusion: Utilizing machine-learning methods to analyse NIRS in the first 12 hours of life, allows for early objective identification of infants at risk of adverse short-term outcomes and may aid in the stratification of infants for intervention in the effective therapeutic window.