Novel methods for predicting clinical outcome in neonates from clinical electroencephalography recordings

Abstract

In this doctoral thesis I investigate a key issue in neonatal intensive care monitoring - the prediction of clinical outcome for infants soon after preterm or compromised full-term birth. With an ever increasing need for robust and reliable bedside monitoring, this dissertation develops a set of novel methods to quantify the electrical activity of the neonatal brain. In particular, I draw upon concepts established in physics and neuroscience known as criticality and crackling noise. These concepts, observed in many other dynamic natural systems, are used here to characterize bursts of cortical activity that are present in the neonatal electroencephalogram (EEG). Henceforth, in this thesis I test two related hypotheses: First, that cortical bursts in neonatal EEG yield important insights into critical brain states and underlying neurophysiology; Second, I propose that the robust characterization of these cortical bursts yield predictive markers of clinical outcome soon after birth, in those neonates at risk of poor outcome. Previous studies of complex natural systems, such as earthquakes, have drawn upon analyses of stochastic, bursty activity known formally as “crackling noise”. These analyses characterize the statistical distribution of the bursts (their area and duration) as well as the average shape of the bursts across a hierarchy of time scales. Furthermore, the study of crackling noise processes allow one to determine whether the bursts have a characteristic scale, or whether they are scale-free and consistent with criticality. In this thesis, I apply both these techniques – analyses of statistical distribution and of average burst shape to quantify the temporal progression of cortical bursts in EEG, from birth to recovery periods in the neonatal intensive care unit. The data driven methods developed in this thesis are henceforth specifically aimed at analyzing the durations, areas and average shapes of EEG bursts in two neonatal populations; 1) Full-term neonates following birth asphyxia, and 2) Extremely preterm infants. In both cohorts, I characterize empirical data features present in neonatal EEG, i.e. durations and areas of cortical bursts by analyzing their probability distributions. Statistical quantifiers of average burst shapes - such as burst symmetry and sharpness - are used to distinguish between healthy and abnormal brain states. The analysis of statistical distributions of bursts and the quantification of average burst shapes leads to an automated metric for outcome prediction. The key, original contributions of this thesis are that EEG bursts following full-term hypoxia and preterm birth are inherently scale-free, that is, their statistical distributions have no characteristic scales and robustly conform to the theoretical exponentially truncated power-law. In full-term hypoxic infants, burst shapes are nearly scale-invariant although show changes in symmetry and sharpness for longer burst durations. These scale-free distributions and average shapes reflect fundamental, stochastic properties of critical brain states that provide insight into important neurobiological processes. For example, following recovery from hypoxia, average burst shapes in full-term neonates become symmetrical and scale invariant across all time scales. I study simulations of stochastic models to investigate possible mechanisms underlying the resumption of healthy cortical activity states such as recovery from metabolic depletion. In the preterm neonate, average shapes reveal a temporal progression in cortical bursts with respect to gestational age, and significantly pre-empt the occurrence of intra-ventricular brain hemorrhage. In these cohort studies, I present the utility of novel methods developed to predict acute brain injury and long-term neurodevelopment. This work establishes statistically significant differences between burst shape indices in healthy versus poor clinical outcomes in both full-term hypoxic and preterm populations. These findings indicate a practical use for prediction and classification of at-risk neonates not readily available in current clinical settings. Moreover, the fundamental understanding of the neonatal brain is enriched by presenting unique features and insights into cortical burst generation, allowing a better understanding of neonatal neurophysiology.