Novel features of early burst suppression predict outcome after birth asphyxia.

Burst suppression patterns in the electroencephalogram are a reliable marker of recent severe brain insult. Here we analyze statistical properties of bursts occurring in 20 electroencephalographic recordings acquired from hypothermic asphyxic newborns in the hours immediately following birth. We show that the distributions of burst area and duration in these acute data predict later clinical outcome in both structural neuroimaging and neurodevelopment. Our findings indicate the first early electroencephalographic metrics that offer outcome prediction in asphyxic neonates undergoing hypothermia treatment.

Introduction

Hypoxic ischemic encephalopathy (HIE) following perinatal asphyxia is the leading cause of acquired, severe neurodevelopmental disability.1 The neurodevelopmental outcome of HIE depends on how quickly neuronal function recovers from the asphyxic insult. Continuous electroencephalography (EEG) monitoring has become a widely accepted practice in neonatal intensive care units (NICUs).2–4 A key EEG hallmark of severely injured brain function is a pattern of low amplitude (suppressed) EEG punctuated by irregular high amplitude bursts (Fig.1A) known as burst suppression (BS).5 Several studies have shown that the resolution of BS in EEG recordings within 6–12 h of birth correlates strongly with a favorable neurodevelopmental outcome.4,6 Identification of BS and its recovery time in asphyxic newborns has thus been used as a prognostic indicator of brain function to guide therapeutic decisions. While clinical decisions to commence therapeutic hypothermia after birth asphyxia are often based on observing BS in the early EEG, hypothermia treatment itself was recently found to significantly delay BS recovery, hence compromising the utility of early EEG in outcome prediction7,8 (see also Ref. 9).

Figure 1

Extraction of burst suppression metrics. (A) Example EEG epoch with burst suppression pattern (P3–P4 derivation). (B) Instantaneous power (amplitude-squared) of the same EEG epoch obtained from the amplitude envelope (via Hilbert transform). (C) Instantaneous power over 80 min of a recording to illustrate the variability in the size of bursts over time. (D) Sample burst taken from graph (B) to illustrate the burst measures. The automated threshold was used for defining the burst area (BA) and the burst duration (BD). (E) Upper cumulative distribution function (CDF) of BA in one infant shows that it follows an exponentially truncated power law distribution (fit shown in green) over several orders of magnitude. On the basis of these observations (refer to methods14,16,17), we extracted the scaling exponent (α) of the distribution. Here the dashed line illustrates the corresponding α for the BA as given by the CDF. (F) Relationship between BAs and BDs in one infant, plotted in double logarithmic coordinates. Points are median durations and median areas calculated after dividing the data into 250 logarithmic bins, with the robust linear relationship quantified by the slope of the linear least-squares fit (red line).

The observation of BS in EEG recordings has long been identified as indicating acute neuronal compromise.1,10–12 Neurophysiological interpretation of BS conventionally rests upon a dichotomous presence-versus-absence classification. While qualitative assessment of “burst sparseness” – a low overall frequency of bursts – may be taken as an additional sign of severity,12 no properties of the bursts themselves have been shown to provide useful prognostic indicators. Exploiting recent advances in complex systems analysis,13–16 we sought to characterize statistical properties of BS, in particular features related to burst duration (BD) and area (i.e., burst size). These features of BS carry novel, clinically important information neither provided by current brain monitoring paradigms nor evident by visual inspection of EEG alone. Our findings provide a novel approach for predicting outcome, based on identifying information embedded in BS activity early after brain injury, a task that has proven thus far elusive.

Patients and Methods

We studied EEG recordings of 20 consecutively admitted newborns (gestational age 39 ± 2 weeks) monitored due to HIE following perinatal asphyxia at the tertiary level NICU of Helsinki University Central Hospital. HIE was assessed according to conventional criteria,17 including lethargy, stupor, or coma, with one or more findings of hypotonia, abnormal reflexes or clinical seizures. The EEG was acquired from biparietal electrodes at 250 Hz using a NicOne or Olympic EEG device (Cardinal Healthcare, Nicolet Biomedical, Madison, WI and Natus Medical, San Carlos, CA). Lengthy epochs of BS that were relatively artifact-free (120 ± 90 min, range 30–600 min) were filtered at 1–20 Hz and further analyzed with custom software in Matlab (Mathworks, Natick, MA). BS was determined visually as an EEG pattern consisting of activity bursts (mixture of sharp, slow waves) periodically interrupted by multiple episodes of suppressed cortical activity (∼<10 μV). Although BS somewhat resembles trace alternant and trace discontinue activity patterns, these are distinct phenomena. First, they are part of the pseudo-periodic vigilance state cycling that is absent early after asphyxia.3,7 Second, they only occur in younger preterm babies (trace discontinue), or their suppression period is insufficiently suppressed for classification as BS (trace alternant). Cooling onset preceded EEG recordings in all babies (see Table S1). Clinical data collated from patient reports included: reports from magnetic resonance imaging (MRI) scans during the first week after birth; birth details; NICU drug treatments; and outcome description from the last visit to routine neonatal outpatient clinic (age 12–39 months). Patients were divided into two categories based on MRI findings (presence/absence of a thalamic lesion), and further into four categories based on clinical outcomes: normal versus mildly, moderately, and severely abnormal (see Table S1 for full details). Four of the 20 available neonatal data were excluded from further analyses due to the presence of excessive seizure activity (n = 1), early administration of sedative drugs that likely affected BS (n = 1), or missing reliable clinical or MRI information (n = 2). The use of patient data for this study was approved by the Ethics Committee of the Hospital for Children and Adolescents, Helsinki University Central Hospital.

To extract statistical properties of the bursts, we first developed an adaptive thresholding method to detect the occurrence of bursts in the EEG (Fig.1, and see Fig. S1 for preprocessing steps). We then extracted three burst metrics for each infant: (1) mean BD and its coefficient of variation (CV), (2) scaling exponent (α) of the cumulative distribution function (CDF)18 of the burst areas (BA) in each recording (Fig.1E), and (3) slope value based on the relationship between BD and BA (Fig.1F). Here, the scaling exponent (α) captures how quickly the number of bursts diminishes as a function of increasing area: A large exponent (α ≫ 1.5) indicates that most bursts are very small in area, whereas the converse (α ∼ 1.5) indicates that larger bursts occur relatively more often. Since the distribution closely follows a truncated power law, there is no single characteristic size but rather a broad “scale-free” regime, limited only by finite brain capacity. Our novel metrics focus on properties of bursts, specifically the heterogeneity of BAs, as well as the relationship between BAs and BDs. In addition, we benchmarked these novel measures against those established in prior research, namely the mean and CV of interburst interval (IBI) durations.10 The slope values from BA versus BD plots and BA CDF plots were extracted as described in Figure1E and F. Notably, the members in our team scripting and running the analysis algorithms did not perform visual EEG interpretation nor clinical data collection.

Group comparisons were conducted using one-way analysis of variance (ANOVA) and verified with nonparametric Mann–Whitney U tests. Correlations of clinical outcomes to EEG metrics were measured using Spearman's rank correlation coefficient.

Results

Visual inspection of bursts occurring during post asphyxic EEG characteristically showed erratic and irregular fluctuations of BS in the amplitude trace (Fig.1A), as well as the corresponding power trace (Fig.1B and C). We hence examined the variability in BA by plotting its CDF (Fig.1E) and the relationship between BA and BD (Fig.1F). We first observed that the CDFs of BA show a robust, linear scaling relationship over several orders of magnitude (Fig.1E). We also observed a robust power-law relationship between BA and BD over four orders of magnitude (Fig.1F). Comparison with subsequent MRI findings (Fig.2A and B) showed that α was significantly higher (F1,16 = 5.39, P = 0.035), and the slope between BAs versus BDs was significantly higher (F1,16 = 7.89, P = 0.013) in babies with thalamic lesions. Moreover, both of these metrics showed a significant correlation with neurodevelopmental outcome (r = 0.81, P < 0.0001, and r = 0.58, P = 0.022, respectively; Fig.2D and E). Visual comparison of different outcome groups shows considerable variability over time and between records, due to both temporal evolution of BS patterns and physical factors (e.g., subcutaneous swelling). Two extreme cases (normal MRI and outcome versus abnormal MRI and outcome) are shown in Figure S4, emphasizing the need for objective, statistically sound methods for their interpretation.

Figure 2

Prediction of outcome with early burst suppression metrics. (A and B) Comparison of burst suppression metrics during first hours after birth to subsequent MRI changes observed several days later. Infants who developed thalamic lesion (Th lesion) apparent in MRI had significantly higher exponents α (A) in the BAs, and a significantly higher slope values (B) in the relationship between BA and BD. The graphs show mean ± SEM (circles and lines) and the individual data points (crosses). (C–E) Comparison of burst suppression metrics during first hours after birth to clinical outcome categories later in infancy (1 = normal; 2 = mildly abnormal; 3 = moderately abnormal; 4 = severely abnormal; see also Table S1). There was a statistically significant correlation between clinical outcome and (C) α (D) slope and a positive correlation in (E) BD CV. Significant group differences were also found when dividing outcomes into two categories, good (1 and 2) and poor (3 and 4).

We then benchmarked these findings against conventional measures of bursts (BDs, BD CVs), and their intervals (IBIs). Statistical analyses revealed that BD CV was significantly higher (F1, 16 = 4.92, P = 0.042) in babies with thalamic lesions compared to those with normal MRI findings. However, no difference was found in the mean duration of bursts (F1,16 = 0.17, P = 0.68) or the mean IBIs (F1,16 = 0.05, P = 0.81). In addition, BD CV showed a positive correlation with the developmental outcome (r = 0.37, P = 0.17; Fig.2C). No correlation was seen between mean BD and any measure of IBIs.

Discussion

Our findings show that novel burst statistics derived objectively from EEG soon after perinatal asphyxia correlate significantly with MRI findings obtained only several days later, as well as with neurodevelopmental outcomes obtained later in infancy. The measures of burst areas and durations, and their interrelationships, were derived from routinely collected EEG records in cooled asphyxic newborns where traditional measures of IBIs or resolution of BS have insufficient prognostic reliability.6,7 The present observations have both theoretical and practical implications.

Theoretically, the presence of robust scale-free statistics is compatible with the notion that a severely compromised brain operates in a critical balance between excitation and inhibition. While such properties of cortical dynamics cannot be assessed using traditional EEG measures, they are rapidly extracted with novel and robust mathematical tools recently developed in computational neuroscience and physics.14,18,19 Pragmatically, extracting clinically informative statistics from a very large number of bursts engenders such metrics insensitive to common noise-related confounders. Moreover, our observations significantly challenge the clinical practice of perceiving BS as a uniform entity.12 Rather, the present findings are compatible with the alternative notion that functional brain recovery is already ongoing during BS where it can be assessed from the statistical characteristics of cortical bursting rather than from the traditionally measured lengths of inactive (IBI) periods.

Practically, our measures present several clinical advantages. First, because they focus on the bursts only, they avoid a need for uninterrupted recordings. Hence, these measures are robust to intermittent artifacts present in the electrically noisy NICU environment and challenge traditional analyses of EEG monitoring. Second, our EEG-derived measures provide outcome predictions 4–7 days earlier than MRI findings, possibly enabling early guidance of individually targeted therapeutic decisions. Third, our paradigm is data-driven, reducing interindividual variability in the overall level of EEG amplitudes caused by technical variations (e.g., interelectrode distance, constant artifact, or subcutaneous swelling). Amplitude variance caused by artifacts is a serious confounder prevalent in many existing approaches, particularly amplitude-integrated EEG (aEEG, a.k.a. Cerebral Function Monitor [CFM]20). The data-driven algorithm employs burst detection via an adaptive threshold, where scaling properties reported were reliable over a range of thresholds (see Supplementary Material). Notably, the EEG measures introduced here are designed for rapid automated prognostication, and thus could be readily implemented as software features in currently used brain monitoring systems.

While our findings are encouraging, the dataset presently studied has limitations, such as a limited sample size and outcome classifications that were retrospectively extracted from clinical archives instead of prospectively collected standardized neurodevelopmental tests. Future studies on larger patient series with standardized outcome measures are needed to validate the present observations. It will also be important to examine the clinical potential of our measures in differential diagnosis of BS etiologies (e.g., asphyxia, trauma, sedative drugs), a common diagnostic challenge in acute patient care. Finally, it is likely that these same statistical properties of BS are also found in adult neuro-intensive care, expanding the clinical implications of our findings.