A comparison of automatically extracted quantitative EEG features for seizure risk stratification in neonatal encephalopathy

Abstract

Introduction:

Seizures occur in up to 40% of neonates with neonatal encephalopathy. Earlier identification of seizures leads to more successful seizure treatment but is often delayed due to limited availability of continuous EEG monitoring. Clinical variables poorly stratify seizure risk, and EEG use to stratify seizure risk has previously been limited by need for manual review and artifact exclusion. The goal of this study is to compare the utility of automatically extracted quantitative EEG (qEEG) features for seizure risk stratification.

Methods:

We conducted a retrospective analysis of neonates with moderate to severe neonatal encephalopathy who underwent therapeutic hypothermia at a single center. The first 24 hours of EEG underwent automated artifact removal and qEEG analysis, comparing qEEG features for seizure risk stratification.

Results:

The study included 150 neonates and compared the 36 (23%) with seizures to those without. Absolute spectral power best stratified seizure risk with area under the curve ranging from 63–71%, followed by range EEG lower and upper margin, median and standard deviation of the range EEG lower margin. No features were significantly more predictive in the hour prior to seizure onset. Clinical exam was not associated with seizure risk.

Conclusions and Significance:

Automatically extracted qEEG features were more predictive than clinical exam in stratifying neonatal seizure risk during therapeutic hypothermia. qEEG represents a potential practical bedside tool to individualize intensity and duration of EEG monitoring and decrease time to seizure recognition. Future work is needed to refine and combine qEEG features to improve risk stratification.

Introduction

Neonatal seizures occur in 1–3.5/1000 term births^1–3 and up to 40% of children with neonatal encephalopathy (NE). Higher seizure burden is independently associated increased morbidity or neurodevelopmental disability^4,5. Recent evidence suggests rapid treatment of neonatal seizures increases the odds of successful treatment and decreases total seizure burden^6,7.

Continuous, neonatal montage EEG is the standard of care during therapeutic hypothermia (TH)⁸. However, optimal EEG monitoring is limited by availability of technologists and epileptologists experienced in neonatal EEG^9–11, leading to issues of resource allocation, triaging, and potentially delayed interpretations¹². Although augmentation of human monitoring with automatic neonatal seizure detection has been tried, even seizure detection algorithms with excellent performance¹³ have failed to meaningfully augment human EEG reading¹⁴ and are plagued by frequent false positives¹².

A high-fidelity biomarker of seizure risk would improve neonatal seizure care by allowing care teams to individualize intensity and duration of EEG monitoring^15,16. Accurate risk stratification would improve time to seizure recognition and treatment, particularly in settings where 24-hour-a-day continuous monitoring is not available. Quantitative evaluation of neonatal EEGs presents an ideal modality for seizure risk stratification. Previous work by Jain et al¹⁷ compared early qualitative amplitude-integrated EEG (aEEG) background and quantitatively calculated EEG power from conventional continuous EEG (cEEG), finding that the best predictors of subsequent seizures were a flat tracing pattern on aEEG and total EEG power <10μV². However, practical clinical application of quantitative EEG (qEEG) has been hampered by the need for manual selection of artifact-free EEG epochs for analysis¹⁸.

The goal of this project is to compare the utility of automatically extracted qEEG features for seizure risk stratification in neonates undergoing therapeutic hypothermia for moderate to severe NE.

Materials and Methods

2.1. Patient population

We conducted a retrospective analysis of all neonates undergoing TH for NE at a single institution over a three-year period (2018–2021, N=178). Neonates were excluded if EEG was initiated after 12 hours of life (N=12), no encephalopathy examination was recorded (N=2), or seizures were evident within 10 minutes of EEG initiation (N=3) (Figure 1). Neonates underwent TH if they had evidence of perinatal compromise (cord pH < 7.10, > 10 minutes resuscitation, 10-minute Apgar score ≤5), met Sarnat criteria for moderate to severe neonatal encephalopathy ^19,20, were ≥35 weeks gestation, and were within 6 hours of birth. The standardized protocol for NE included neonatal montage cEEG monitoring starting as soon as possible after the decision to cool and continuing for at least 24 hours. Clinical features were extracted from the medical record. The total Sarnat score²¹ was calculated based upon the worst documented Sarnat exam. MRIs were clinically categorized as no, minor (punctate), moderate (intermediate) or profound injury (>50% of white matter and deep grey affected) based on Trivedi et al²², although no numerical score was assigned. Initial categorization was assigned based on review of the clinical MRI report read by pediatric neuroradiologists provided with GA and clinical context of NE but no details of the clinical course. A neonatal neurologist (JK) reviewed the MR images and made a final determination of categorization when the clinical report was unclear. MRIs were obtained between days 4–6 or after day of life 10 if clinically unstable for transport following rewarming. This study was approved by the Washington University Institutional Review Board.

Figure 1: Patient inclusion and analysis flowchart.

178 patients who underwent therapeutic hypothermia were screened with 150 undergoing full analysis of multiple quantitative EEG (qEEG) measures including range EEG (rEEG) measures using the first hour of EEG and the hour prior to seizure onset. The presence of macroperiodic oscillations (MOs) was evaluated over the first 24 hours. SD = Standard deviation, GA = gestational age, HOL = hours of life

2.1.1. Data Availability Statement:

Limited, de-identified data are available on reasonable request by qualified researchers. These data are not publicly available due to restrictions concerning the privacy of research participants and release of medical records.

2.2. Neurophysiologic Analysis

All cEEGs were evaluated for seizures during clinical review by a board-certified pediatric epileptologist or neurophysiologist who documented the presence or absence of seizure and timing of first seizure. Quantitative EEG analysis was conducted for the first 24 hours of EEG monitoring using neonatal montage EEG processed using the NEURAL toolbox^23,24 for MATLAB. The NEURAL toolbox was specifically designed for neonatal EEG analysis and is freely available. Briefly, all channels were filtered with a 0.5–50 Hz bandpass filter. Channels with evidence of improper electrode placement or electrode coupling were automatically removed from further analysis²³. For remaining channels, continuous rows of zeros, high-amplitude artifacts and sudden jumps in amplitude were automatically removed along with a 5 second collar of time surrounding each artifact²³ (Figure 1).

After artifact removal, features were extracted for each channel across standard frequency bands (delta 0.5–3 Hz, theta 4–7 Hz, alpha 8–12 Hz, beta 13–30 Hz) using 30 second epochs over a 1-hour evaluation period. The features evaluated include the absolute and relative spectral power for each frequency band, connectivity, range EEG (rEEG) measures including median, upper and lower margin, width, asymmetry, standard deviation (SD) and SD of the rEEG lower margin. rEEG is closely related to aEEG, using a peak-to-peak measure of voltage over a short window which is transformed to a linear-log amplitude. However, rEEG is preferred for quantitative analysis due to inconsistent definitions of aEEG and the ability to analyze multiple frequency bands²⁵.

The median value of all 30 second epochs over the first hour of EEG excluding outliers was used for each patient. Quantitative evaluation was performed on limited channel montage for comparability to previous studies^17,26 in alignment with the typical aEEG electrode placement (C3-P3 and C4-P4). In N=11 neonates both channels were removed by automatic artifact removal with evidence of diffuse artifact on manual review and were excluded from further analysis. In N=41 patients either C3-P3 or C4-P4 were removed, and the remaining channel was used. In N=109 patients the value from both channels were averaged and used as the feature value for that individual (Figure 1).

To analyze change in qEEG features with approaching seizure onset a second timepoint was evaluated consisting of one hour prior to seizure onset in N=36 children with seizures. To control for EEG evolution²⁷; patients with seizures were matched by sex and gestational age at a 1:2 ratio to neonates without seizures who had their EEG evaluated at the same age (in hours of life).

Macroperiodic oscillations (MOs) are a recently described EEG pattern characterized by ultralow frequency oscillations which have been associated with both seizures²⁸ and with poor outcomes after ICU stay²⁹. MOs were evaluated as previously described in Lo et al²⁹. Briefly, a time-varying power envelope signal was extracted for each channel across each frequency band. The strength of MOs was quantified as the extent to which the envelope signal is sinusoidal over 40-minute epochs using linear regression to estimate the signal as a weighted sum of sinusoids across our frequency range of interest. An index value, q, was calculated using the degree of fit. The strength of MOs, qmax, was defined as the average of the 9 epochs over the 24 hours of the study with the highest q values in each frequency band. For each patient, Qmax was defined as the highest qmax in any frequency band and values >0.65 were defined as demonstrating MOs.

All analysis was completed in MATLAB (MATLAB 2021a, The MathWorks, Inc., Natick, Massachusetts, United States).

2.3. Statistical Analysis

Categorical variables are reported as the number of patients and percentage of the total. Continuous variables were reported as a median along with the first and third quartile. Group comparisons were conducted using Kruskal-Wallis and Fisher’s exact tests. The receiver operating curve (ROC) was plotted and the area under the curve (AUC) was calculated using the Youden index. AUC 95% confidence intervals were calculated using 2000 stratified bootstrap replicates³⁰. Evaluation of EEG features at multiple timepoints was conducted using a repeated measures ANOVA. Significance was defined as P value <0.05. Statistical analysis was completed in R Studio (R Core Team (v 1.4.1103, 2021), Vienna Austria).

Results

3.1. Patient Demographics

Patient demographics are summarized in Table 1. 150 neonates with neonatal encephalopathy were included in the analysis, of which 36 developed seizures. 67% were male (N=100), the median gestational age was 38 weeks, 88% had moderate encephalopathy, and 12% had clinically severe encephalopathy at enrollment. Profound injury on magnetic resonance imaging (MRI) was relatively rare in this cohort (N=7, 5%), and 66% had no injury on MRI. 91% of patients were discharged without oxygen and on full oral feeding, with 4% dying prior to discharge, and 5% requiring enteral feeding.

Table 1 –

Characteristics of study patients

Cohort Characteristics	Seizure on EEG (N=36)	No seizure on EEG (N = 114)	Total (N = 150)	P-value
Male gender (%)	22 (61%)	78 (68%)	100 (67%)	0.72
Gestational Age in weeks - median (1st/3rd quartile)	39 (38–39)	38 (37–39)	38 (37–39)	0.06
Degree of encephalopathy at enrollment (%)				0.92
- Moderate encephalopathy	31 (86%)	101 (89%)	132 (88%)
- Severe encephalopathy	5 (14%)	13 (11%)	18 (12%)
Age in hrs at cEEG start - median (1st/3rd quartile)	6.0 (4.9–7.5)	6.0 (4.9–8.0)	6.0 (4.5–8.0)	0.97
Age in hrs at seizure start - median (1st/3rd quartile)	12.0 (8.0–17)	N/A	N/A	N/A
MRI degree of injury (%)				<0.01
- No Injury	16 (44%)	83 (73%)	99 (66%)
- Minor injury	11 (31%)	26 (23%)	37 (25%)
- Moderate Injury	5 (14%)	2 (2%)	7 (5%)
- Profound/Global injury	4 (11%)	3 (3%)	7 (5%)
Length of stay in survivors in days - medican (1st/3rd quartile)	18 (13–32)	15 (9.0–20)	15 (11–22)	0.02
Status at discharge				0.06
- RA and oral feeding	29 (81%)	107 (94%)	136 (91%)
- NG/G-tube fed	4 (11%)	4 (3%)	8 (5%)
- Died	3 (8%)	3 (3%)	6 (4%)

3.2. Absolute spectral power is associated with seizure risk

Decreased absolute power was associated with an increased seizure risk in the subsequent 24 hours in all frequency bands, most significantly associated in the alpha band (p = <0.0001). There were also associations in the delta (p = 0.002), theta (p = 0.016), and beta (p = 0.0008) bands (Figure 2). When examining the ROC for the delta band, 93 μV² best differentiates neonates who did vs did not have seizures with a specificity of 80% and sensitivity of 55% (Supplementary Figure 1), and sensitivity drops to 33% to achieve a 95% specificity. In the theta band, 6 μV² best differentiates (specificity, 91% sensitivity 44%) In the alpha band 3 μV² is optimal (specificity 92%, sensitivity 50%), and in the beta band 4 μV² (64% specificity, 64% sensitivity) (Supplementary Table 1).

Figure 2 A-D: Comparison of absolute spectral power during the first hour of EEG in neonates with and without seizure in the subsequent 24 hours.

In each frequency band (delta, theta, alpha and beta) the absolute power in μV² was lower in patients with subsequent seizures. P<0.05 = *, P<0.01 = **, P<0.001 = ***, P=<0.0001 = ****.

3.3. Comparison of qEEG features for evaluating seizure risk

qEEG features were evaluated during the first hour of EEG to compare their efficacy for differentiating neonates who would develop seizures in the subsequent 24 hours (Figure 3). In addition, degree of encephalopathy as defined by a composite Sarnat score and MOs as evaluated over the first 24-hour period were compared to qEEG features.

Figure 3: Univariate comparison of features for evaluation of seizure risk during initial TH EEG monitoring.

Clinical exam with Sarnat scoring, MOs strength over the first 24 hours and quantitative EEG features over the first hour of EEG including absolute and relative spectral power, connectivity and multiple rEEG features were compared. Absolute spectral power in all frequency bands was most strongly associated with seizure risk with AUCs ranging from 63–71%, followed by rEEG lower and upper margin, median and SD of the rEEG lower margin.

Clinical exam was not significantly associated with seizure risk (p = 0.71). MOs strength over the first 24 hours as measured by Qmax was associated with seizure (p = 0.008). Absolute spectral power in all frequencies was significantly associated with seizure risk with an AUC ranging from 63% to 71%. Relative spectral power was not significantly associated with seizure risk at any frequency. Connectivity was also not associated with seizure risk (p = 0.71), nor was rEEG standard deviation (p = 0.07), rEEG asymmetry (p = 0.19) or rEEG width (p = 0.05). rEEG lower margin, median and upper margin were all significantly associated (p = 0.001, p=0.003 and p = 0.03 respectively) with AUCs ranging from 62%−68%. The standard deviation of the rEEG lower margin was associated with seizure risk with p = 0.005 and an AUC of 65%.

3.4. Comparison of qEEG features for evaluating seizure risk the hour prior to seizure onset

EEG features evolve over time for infants with neonatal encephalopathy and can be predictive of outcome³¹. A second time point one hour prior to seizure onset was examined and compared to a similar timepoint in matched neonates without seizures in order to evaluate for evolution of features with respect to prediction of seizure risk (Figure 4). A very similar profile of feature performance was seen prior to seizure onset as observed during the initial hour of EEG. Absolute spectral power continued to be significantly associated with seizure risk in all frequencies although relative spectral power was not. rEEG lower margin, median and upper margin as well as standard deviation of the rEEG lower margin continued to be significantly associated with seizure risk.

Figure 4: Univariate comparison of features an hour prior to seizure onset for evaluation of seizure risk.

Clinical exam with Sarnat scoring, MOs strength over the first 24 hours and quantitative EEG features in the hour prior to seizure onset including absolute and relative spectral power, connectivity and multiple rEEG features were compared. Absolute spectral power in all frequency bands remained most strongly associated with seizure risk with AUCs ranging from 62–70%, followed by rEEG lower and upper margin, median and SD of the rEEG lower margin.

3.5. Evolution of features over time did not improve seizure risk stratification

Repeated measures ANOVA for each feature over all frequency bands demonstrated that many EEG features evolved temporally, as expected. Features which differentiated neonates who would and would not go on to have seizures remained stable from the first hour to the hour preceding first seizure. There were no features which changed significantly more between the first hour and the hour prior to seizure in neonates with seizures than those without seizures (Supplementary Table 2).

Discussion

NE is a common, high morbidity neonatal diagnosis with up to a 40% risk for seizures. Current tools for predicting seizure risk, including clinical exam and qualitative EEG interpretation, are either poorly predictive or require complex, subjective scoring with significant inter-rater reliability issues³². This study compares qEEG features extracted with automated artifact removal to evaluate a potentially practical approach to stratification of seizure risk. We found seizures were associated with lower absolute spectral power and MOs strength, as well as rEEG features, including lower and upper margin, median and SD of the rEEG lower margin.

4.1. Clinical exam and data use in seizure risk stratification

Extensive evaluation of clinical variables has not revealed a strong and consistent predictive marker of seizure risk during TH, including degree of encephalopathy, clinically suspected seizures prior to monitoring, initial pH, base excess, or 10-min Apgar score³³. Due to poor correlation between clinical variables and seizure risk, the American Clinical Neurophysiology Society recommends use of cEEG for at least 24 hours in NE⁸. Our results also demonstrated no significant association between clinical exam and seizure risk. This underlines the need to identify alternative ways to stratify seizure risk to optimize and individualize duration and intensity of EEG monitoring.

4.2. Absolute vs relative spectral power use in seizure risk stratification

Previous work comparing relative and absolute spectral power in the neonatal population has found that absolute spectral power differentiates severity of encephalopathy better than relative spectral power, although it is more sensitive to artifacts³⁴. Jain et al¹⁷ found that a predictor of subsequent seizures included total EEG power less than 10μV² (sensitivity 52%, specificity 98%). Our results support that lower absolute power in the first hour of EEG was useful in predicting seizure risk in the following 24 hours, albeit with low sensitivity. This finding is also in line with qualitative EEG evaluation identifying extremely low voltage EEG background as a risk factor for subsequent seizures³³.

In NE, absolute total and delta band power are equally associated with the severity of encephalopathy²⁶; it is therefore unsurprising that both delta power and absolute power in other spectral bands correlate with seizure risk. In rat models of neonatal hypoxic injury, initial suppression across all frequency bands is seen, but alpha and beta bands remained suppressed in animals with significant hypoxic injury³⁵. In adult rat rodent models, theta band changes have more prominent changes prior to seizure onset in post-injury epilepsy models³⁶, suggesting that more work is needed to explore differences in spectral band dynamics.

Neonatal specific algorithms for seizure detection have previously found that relative spectral power outperforms absolute spectral power for neonatal seizure identification^37,38. It is logical that seizures, a high-frequency, high-amplitude phenomenon, would fundamentally alter relative spectral power. However, our work suggests that there are not consistent and predictive changes in the pre-ictal relative frequency band contributions, either in the first hour of EEG monitoring or shortly preceding ictal onset. This may be because in the acute period of injury there is baseline suppression of all frequencies, as seen in neonatal hypoxic rat models³⁵, without time for differential recovery of higher vs lower frequencies in the median 6 hours prior to seizure development.

4.3. aEEG and rEEG use in seizure risk stratification

Although cEEG is the gold standard for monitoring in NE, aEEG is widely utilized due to its ease of bedside interpretation and in settings where cEEG is unavailable³⁹. Jain et al¹⁷ found that a qualitatively flat aEEG tracing pattern during the first hour on aEEG had a high risk of subsequent seizure. Several studies have identified qualitative cEEG features associated with low risk for seizures which could potentially be extrapolated to aEEG. Benedetti et al found that infants with state cycling, synchronous and symmetric activity, typical amplitudes and quiet sleep inter-burst-intervals < 7 seconds on cEEG by 24-hours of life were not found to have subsequent seizures¹⁶, with other authors similarly identifying normal or essentially normal EEG as very low risk for seizure^40,41. A recent study from Lacan et al explored recreating the cEEG qualitative grading system with quantitative variables and found that the minimal amplitude and the burst suppression ratio both significantly differentiated normal/minimally abnormal EEGs from moderately abnormal EEGs⁴². In our study, lower margin of rEEG was associated with higher risk of seizures, which is consistent with cEEG patterns of burst suppression and extremely low voltage³³ that have similar associations. Similarly, lack of normal state cycling can be assessed quantitatively by measuring the standard deviation of the lower margin on rEEG over the hour-long evaluation period, as we observed here. In future work we hope to develop more sensitive quantitative measures of state cycling on rEEG.

4.4. Macroperiodic oscillations are associated with seizures

MOs are a measure of ultralow frequency oscillations^29,43 which have been associated with seizures in the ICU²⁸, and impairment in neurodevelopmental outcome after neonatal encephalopathy. In this cohort the presence of MOs (present in 82 of 150 neonates) was significantly associated with the presence of seizures. Ongoing work into the underlying pathophysiology of MOs is needed to better understand whether these ultraslow oscillations represent a paraepileptic phenomenon, evidence of acute brain injury, or both.

4.5. Future directions

Automatically filtered neonatal qEEG demonstrates promise as a bedside tool for seizure risk stratification, but further work is needed before clinical utilization. This study focused on analyzing limited channel EEG in order to build on previous work, but the use of full neonatal montage analysis and evaluation of spatial distribution of qEEG features may further improve seizure risk stratification, particularly as neonatal seizures are typically very focal and may be missed on limited montage⁴⁴. The identification of neonates at a very low risk of seizure will particularly benefit from development of more sophisticated measures of state cycling and sleep architecture as these have been shown to qualitatively portend a very low risk for seizure¹⁶. Finally, while many of the individual qEEG features identified here were significantly associated with seizure risk, ongoing work is focused on improving our seizure prediction through development of models incorporating multiple clinical and qEEG features, including utilization of machine learning to optimize the use of combined features for predict seizure risk, an approach which has been successful in neonatal neurodevelopmental risk stratification⁴⁵. We hope to also develop models examining prediction of post-injury epilepsy and further refining neurodevelopmental risk stratification.

4.6. Limitations

Our study does have limitations. Eleven of the original 161 identified neonates (7%) had EEGs with diffuse artifact that could not be corrected by automatic artifact removal. Although patients with clinically mild NE represent an increasing percent of patients undergoing TH and have a known seizure risk, our sample was limited to neonates with moderate to severe NE as our center does not cool nor routinely perform EEG on neonates with mild NE. This study evaluated for the risk of seizure, but we did not evaluate for the subsequent seizure burden nor responsiveness of anti-seizure medications. Finally, although NEURAL is a publicly available Matlab toolbox, the features identified here cannot currently be evaluated in real time at bedside. We hope that this work represents a first step in identifying a clinically useful qEEG based model of seizure risk which could be translated to and implemented in commercially available bedside qEEG software.

Conclusion

In this cohort of neonates with moderate to severe NE, we found that multiple automatically extracted qEEG features, particularly absolute spectral power and rEEG lower margin and variability, can be used for seizure risk stratification in NE. Further work is needed to optimize, refine and combine these measures to provide a practical clinical tool for neonatal seizure risk stratification.