Early Electroencephalography and Amplitude-Integrated Electroencephalography for the Prediction of Neurodevelopmental Outcomes in Neonates with Hypoxic Ischemic Encephalopathy: A Systematic Review and Diagnostic Test Accuracy Meta-Analysis

Authors: Vijay Kumar Krishnegowda, Viraraghavan Vadakkencherry Ramaswamy, Prathik Bandiya, Tapas Bandyopadhyay, Thangaraj Abiramalatha, Arun Prasath, Daniele Trevisanuto

Abstract

Introduction

Electroencephalography (EEG), including both conventional EEG (cEEG) and amplitude-integrated EEG, is early prognostic tools utilized in neonates with hypoxic ischemic encephalopathy (HIE). However, the reported predictive accuracy of EEG varies widely.

Methods

We evaluate the diagnostic accuracy of EEG in predicting neurodevelopment impairment (NDI) among neonates ≥35 weeks with any stage HIE. MEDLINE, Embase, Cochrane Library, and Scopus were searched from inception until 24th December 2024. Observational studies evaluating EEG performed in the first 72 h of life in neonates with HIE, and reporting NDI outcomes assessed after 12 months were included. Two authors independently extracted data. A Bayesian random-effects bivariate model was used for diagnostic test accuracy meta-analysis. Risk of bias was assessed using QUADAS-2, and certainty of evidence (CoE) with GRADE. NDI defined as cognitive/motor scores <1 SD below the mean or presence of motor disability.

Results

Sixty-two studies (n = 3,929) were included. In neonates who underwent therapeutic hypothermia (TH) (34 studies, n = 2,538), EEG showed a sensitivity of 88.3% (95% credible interval [CrI]: 83.7%, 92.8%) and specificity of 63.9% (53.6%, 72.8%). In no TH group (33 studies, n = 1,391), the sensitivity was 87.2% (77.5%, 93.5%) and specificity was 76.3% (61.5%, 86.8%). Further, in neonates who received TH (12 studies, n = 868), cEEG had an acceptable sensitivity of 84.1% (77.3%, 89.9%) and specificity of 76.7% (66.9%, 84.3%). CoE being predominantly moderate.

Conclusion

EEG has good sensitivity in predicting NDI regardless of TH status and may aid in identifying high-risk neonates for further evaluation.

Introduction

Neonatal hypoxic ischemic encephalopathy (HIE) is one of the important causes of mortality and, short- and long-term morbidity in term neonates [1]. Therapeutic hypothermia (TH) is the only proven neuroprotective therapy in neonates with moderate to severe HIE and is the standard of care in high-income countries (HICs). Despite TH, nearly 30–50% die or survive with neurodevelopmental impairment (NDI) [1]. Thus, it is crucial to identify neonates at risk of NDI so as to initiate prompt diagnosis-specific early intervention to improve their outcomes. Additionally, early prognostication aids in clinical decision making and effective parental communication.

When compared to magnetic resonance (MR) imaging, a simpler prognostic tool electroencephalography (EEG) could be performed bedside even in the early days of postnatal life [2]. The use of EEG in HIE serves multiple other purposes as well, namely, identifying neonates eligible for TH, monitoring evolving brain injury, detecting, and managing seizures, and possibly predicting the risk of death and childhood epilepsy [3]. However, conventional EEG (cEEG) monitoring is resource-intensive and presents challenges in its implementation. Henceforth, amplitude-integrated EEG (aEEG) has been widely adopted in NICUs offering TH, as its interpretation is easier for clinicians while allowing continuous monitoring.

Consensus guidelines have recommended cEEG as the reference standard for monitoring of neonates with HIE [4, 5], suggesting its use along with aEEG. Nevertheless, these guidelines were based on data from neonates receiving TH, with no specific recommendations for those who were not treated with TH. This systematic review and diagnostic test accuracy (DTA) meta-analysis aims to comprehensively review the prognostic utility of EEG performed in the first 72 h of life in neonates with HIE who did or did not receive TH for predicting NDI.

Methods

Protocol Registration

The systematic review protocol was registered with PROSPERO (CRD42024577001) [6]. The reporting adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies (PRISMA check list in the online supplement; for all online suppl. material, see https://doi.org/10.1159/000548737) [7].

Eligibility Criteria

We included observational studies with various study designs. Only published primary literature was considered. Case reports, case series, and conference abstracts were excluded. No time or language restrictions were applied.

Participant Selection Criteria

We included studies that had enrolled neonates of gestational age ≥35 weeks diagnosed with HIE of any grade. Inclusion criteria for labelling HIE encompassed the need for resuscitation, low Apgar scores, umbilical cord blood gas acidemia, and requirement of prolonged ventilation at birth. Studies that had evaluated neonates who received TH treatment and those who did not were included.

Index Test(s)

Studies reporting on EEG (cEEG or aEEG) performed within the first 72 h after birth were included. Only background parameters as described by the study authors were included, provided they could be dichotomized. For aEEG studies reporting both voltage and pattern, background pattern was prioritized over voltage. Studies that provided only quantitative EEG data were excluded.

Reference Standard(s)

We included all validated neurodevelopmental assessment tools and clinical neurological examinations, conducted at least after 12 months of age. While no specific test was designated as the gold standard, assessments evaluating both cognitive and motor domains between 18 and 30 months of age were considered the ideal reference standards.

Target Condition

NDI was defined as a developmental quotient below 1 standard deviation or presence of any motor disability on neurological examination.

Literature Search

Electronic databases namely MEDLINE, Embase, the Cochrane Library, and Scopus were searched from the inception until 24th December 2024. The literature search strategy is provided in online supplementary Table 1. A forward citation search was also done to identify potentially eligible studies. Rayyan software (Rayyan QCRI, Doha) was used for the literature screening. Titles and abstracts were screened independently by two authors, with conflicts resolved by a third author. Full-text articles of potentially eligible studies were independently assessed for inclusion, and disagreements were resolved through consensus.

Study Selection and Data Collection

Two authors independently extracted the data in duplicate using a prespecified proforma. The extracted data were validated by a third author. For the reported EEG parameters, true positives, false positives, false negatives, and true negatives were recorded at all available time points. However, amongst the available time points, the earliest DTA measures were included in the data synthesis.

Quality Assessment

The risk of bias and applicability of the included studies were evaluated using the QUADAS-2 tool [8]. To ensure consistency, we predefined standardized criteria for the signalling questions (online suppl. Table 2). Two authors independently conducted the quality assessment with discrepancies resolved through consensus.

Statistical Analysis

The included studies were categorized into two groups based on TH those who received TH and those who did not receive TH. We estimated individual study sensitivity and specificity with corresponding 95% confidence intervals (CIs). These estimates were visually represented in forest plots.

Bayesian meta-analysis was conducted using MetaBayesDTA and the R software (R Foundation for Statistical Computing, Vienna, Austria) [9, 10]. For Bayesian inference, we applied weakly informative prior distributions combined with the observed data likelihood to derive posterior estimates for the unknown parameters. Sampling was performed using the Stan Hamiltonian Monte Carlo sampler, employing a maximum tree depth of 10 and an adapt_delta of 0.8. Four Markov chains were run, each with 500 warm-up iterations and 1,500 total iterations. A truncated normal prior was assigned for between-study standard deviations, while the Lewandowski-Kurowicka-Joe (LKJ) prior was used for between-study correlations [10].

Model fit was assessed through correlation residual plots and frequency table probability residual plots. Convergence diagnostics included examination of trace plots, density plots, and R-hat values. For each meta-analysis, we reported the posterior median and 95% credible interval (CrI) for pooled sensitivity and specificity. Summary receiver operating characteristic (SROC) curves were generated, displaying the posterior median estimates along with the 95% prediction intervals.

Subgroup Analyses, Meta-Regression, and Sensitivity Analyses

Prespecified subgroup analyses were conducted based the index test (cEEG vs. aEEG) and assessment time points (postnatal 1–6, 1–24, 1–48, and 1–72). A prespecified sensitivity analysis was also performed by excluding studies with applicability concerns.

To explore heterogeneity, we conducted meta-regression analysis to examine potential associations between study-level EEG modality (cEEG vs. aEEG), HIE severity (I, II, and III vs. II and III) and timing of neurodevelopmental assessment (18–30 months vs. before 18 months and after 30 months).

Publication Bias

We assessed publication bias using Deeks’ funnel plot.

Certainty of the Evidence

Certainty of evidence (CoE) was evaluated for sensitivity and specificity using the GRADE framework for DTA meta-analysis [11].

Results

The initial search yielded 4,315 records. After excluding duplicates, 62 studies were included in the meta-analysis (Fig. 1). Excluded studies with valid reasons are provided in online supplementary Table 3 [12–79]. Of the 62 included studies [42, 71, 80–139], 34 studies [83, 99, 103, 108, 109, 111, 114–126, 128–131, 133–136, 138, 139] evaluated EEG in neonates treated with TH, and 33 studies [42, 80–82, 84–98, 100–102, 104, 107, 110, 112, 132, 137] focused on neonates not receiving TH.

Amongst the studies that had enrolled neonates who received TH, cEEG was performed in 12 studies [83, 99, 105, 114–116, 119, 121, 124, 127, 135, 136] and aEEG in 22 studies [71, 103, 106, 108, 109, 111, 113, 117, 118, 120, 122, 123, 125, 126, 128–131, 133, 134, 138, 139]. The timing of assessment with EEG in the aforementioned studies was within 6 h in 12 studies [71, 103, 108, 109, 113, 115, 120, 123, 125, 128, 133, 138], 24 h in 14 studies [105, 106, 111, 116–118, 126, 127, 129, 131, 134–136, 139], 48 h in 2 studies [83, 99],and 72 h in 6 studies [114, 119, 121, 122, 124, 130]. Neurodevelopmental assessments included the Bayley Scales of Infant and Toddler Development (BSID) II [71, 106, 113, 125–127, 133], BSID III [109, 114, 117, 120, 122–124, 126–129, 136, 138], Griffiths scales [108, 111, 118, 123, 130], Brunet-Lézine scale [83], Kyoto scale of neurodevelopment assessment [131], the Denver Developmental Screening Test (DDST) [116], and neurological examinations performed by clinicians [99, 103, 105, 115, 119, 121, 124, 134, 135, 139].(Table 1).

Amongst the studies that had included neonates who did not receive TH, cEEG was performed in 21 [42, 81, 84, 86, 88–92, 96–98, 100, 101, 104, 105, 107, 110, 112, 127, 137] and aEEG in 12 studies [71, 80, 82, 85, 87, 93–95, 102, 106, 113, 132]. The index test was conducted within 6 h in 12 studies [71, 82, 86, 87, 89, 94, 95, 100, 102, 106, 113, 132], within 24 h in 6 studies [85, 98, 101, 105, 107, 127], within 48 h in 5 studies [42, 84, 88, 90, 92], and within 72 h in 10 studies [80, 81, 91, 93, 96, 97, 104, 110, 112, 137]. The reference standards included BSID II [71, 95, 96, 101, 106, 107, 113, 127, 132], BSID III [127, 137], Griffiths scales [81, 85–87, 89, 91, 94, 100, 102], DDST [42, 80, 97, 110, 112], Brunet Lézine scale [92] and clinician performed neurological examinations [82, 84, 88, 90, 93, 98, 104, 105]. The background parameters utilized in each study are summarized in Table 1 and in cEEG in online supplementary Figure 1.

Risk of Bias and Applicability

QUADAS-2 assessment of all domains is detailed in online supplementary Table 4 and Figure 2.

Meta-Analysis

The DTA meta-analysis of studies evaluating EEG in neonates treated with TH for predicting NDI [71, 83, 99, 103, 105, 106, 108, 109, 111, 113–131, 133–136, 138, 139] (34 studies, n = 2,538) showed a median pooled sensitivity of 88.3% (95% CrI: 83.7%, 92.8%; CoE: moderate) and a specificity of 63.9% (53.6%, 72.8%; CoE: moderate) (Fig. 3; online suppl. Table 5). Analysis of studies that had included neonates who did not receive TH indicated that EEG [42, 71, 80–82, 84–98, 100–102, 104–107, 110, 112, 113, 127, 132, 137] (33 studies, n = 1,391) had a median pooled sensitivity of 87.2% (77.5%, 93.5%; CoE: moderate) and specificity of 76.3% (61.5%, 86.8%; CoE: low) for the prediction of NDI (Fig. 4; online suppl. Table 5).

Subgroup Analyses

Index Test: cEEG vs. aEEG

Synthesis of data from studies evaluating neonates receiving TH showed that cEEG [83, 99, 105, 114–116, 119, 121, 124, 127, 135, 136] (12 studies, n = 868) possibly had a pooled sensitivity of 84.1% (77.3%, 89.9%; CoE: moderate) and specificity of 76.7% (66.9%, 84.3%; CoE: moderate) for the outcome of NDI. For aEEG [71, 103, 106, 108, 109, 111, 113, 117, 118, 120, 122, 123, 125, 126, 128–131, 133, 134, 138, 139] (22 studies, n = 1,670), the pooled sensitivity was 87.4% (82.9%, 91.5%; CoE: moderate), and the pooled specificity was 56.6% (48.1%, 64.8%; CoE: low) (online suppl. Fig. 2A; online suppl. Table 5).

Results from studies that had evaluated neonates who did not receive TH indicated that cEEG [42, 81, 84, 86, 88–92, 96–98, 100, 101, 104, 105, 107, 110, 112, 127, 137] (21 studies, n = 673) had a pooled sensitivity of 87.2% (82.8%, 91.0%; CoE: moderate) and a pooled specificity of 77.4% (69.4%, 84.2%; CoE: low) for predicting NDI. For aEEG [71, 80, 82, 85, 87, 93–95, 102, 106, 113, 132] (12 studies, n = 709), the pooled sensitivity was 88.4% (83.3%, 92.9%; CoE: moderate), and specificity being 77.0% (64.4%, 86.6%; CoE: low) (online suppl. Fig. 2B; online suppl. Table 5).

Timing of Index Test

EEG performed at 1–6 h, 1–24 h, 1–48 h, and 1–72 h showed variations in the sensitivity and specificity. For neonates who received TH, the pooled sensitivity was higher when performed in first 6 h of life. In contrast, specificity was higher when EEG was performed at later time points (online suppl. Fig. 3A; online suppl. Table 6).

For neonates who did not receive TH, the sensitivity remained the same. Though specificity did not show any distinctive pattern, it was highest when performed between 1 and 72 h (online suppl. Fig. 3B; online suppl. Table 6).

Reference Standard

A subgroup analysis based on the neurodevelopment assessment tool was performed, which showed similar sensitivity and specificity across all the reference standards in neonates who received TH and those who did not (online suppl. Table 7).

Heterogeneity Exploration by Meta-Regression

We assessed heterogeneity by visualization of forest plots and distribution of study estimates in SROC curves [148]. Heterogeneity was observed for the specificity parameter in both neonates receiving TH and those who did not. To explore the plausible causes of heterogeneity, we performed meta-regression based on various prespecified covariates. In neonates receiving TH, the index test (aEEG vs. cEEG) showed significant differences in specificity −20.5% (−35.1%, −4.3%), with improved specificity with cEEG. Whereas, in neonates who did not receive TH, the severity of HIE (I, II, and III vs. I and II) showed a significant effect on the specificity (30.3% [6.6%, 57.2%]). However, the timing of the reference test assessment had minimal influence on the observed heterogeneity (online suppl. Table 8).

Sensitivity Analysis

We performed a predefined sensitivity analysis by excluding studies with applicability concerns. Detailed results are provided in online supplementary Table 9.

Publication Bias

Publication bias was not detected for studies that had included either neonates who received TH (p = 0.92) or those who did not receive TH (p = 0.63) (online suppl. Fig. 4, 5).

CoE Assessment

The CoE was assessed for EEG, as well as for cEEG and aEEG separately, in two sub-groups of neonates, namely, those who underwent TH and those who did not, in predicting the outcome of NDI (online suppl. Table 10).

Discussion

Our systematic review and DTA meta-analysis with moderate certainty showed that EEG has good sensitivity but poor specificity. In neonates who did not receive TH, the pooled sensitivity was also good, with evidence certainty being moderate, while specificity was relatively better, with evidence certainty being low. The observed differences in specificity may be due to the influence of TH on EEG pattern recovery, potentiated by the fact that majority of these studies (26 out of 34 studies) assessed EEG in the first 24 h in the TH group.

cEEG has a high sensitivity in the identification of seizures [149]. However, the role of either cEEG or aEEG in HIE neonates for predicting NDO is unclear. In neonates treated with TH, cEEG performed within the first 72 h showed an improved specificity of 76% while maintaining a good sensitivity of 84%. This contrasts with studies utilizing aEEG, indicating that cEEG retained its ability to discriminate between neonates with and without NDI. To translate these findings, cEEG improved the correct identification of neonates with NDI by 15% (prevalence of 25% NDI) without increasing the proportion (7%) of false negatives. Conversely, in neonates who did not receive TH, both cEEG and aEEG exhibited comparable sensitivity and specificity. More importantly, the observed improvement in the prognostic performance of EEG beyond 24 h further supports the need for prolonged monitoring, preferably with cEEG.

It is a standard practice to use cEEG or aEEG along with serial neurological examinations during the initial days and MR neuroimaging between 5 and 7 days of life in neonates with HIE for prognostication [150]. However, there is currently no established clinical pathway to determine which neonates may require MR imaging for prognostic purposes. Our review highlights the role of EEG, which, due to its reasonably high sensitivity, can be valuable tool for stratifying high-risk neonates for additional high accuracy prognostic tests. A high sensitivity of EEG would translate to fewer eligible high-risk neonates being missed, while an acceptable specificity could reduce the number of neonates undergoing unnecessary invasive tests later for prognostication. Nevertheless, MR neuroimaging provides additional value, such as identifying alternative etiologies, as well as the location and extent of injury which may serve as a valuable tool for prognostication.

In our review, a key challenge in synthesizing EEG data was the heterogeneity of background parameter definitions, as we could not explore them statistically. For cEEG studies, the voltage cut-off for low voltage and flat/inactive trace varied. In contrast, aEEG cut-offs for voltage and pattern recognition remained largely consistent across studies. These variations might have influenced the DTA estimates. To address this, we followed an approach adopted by previous studies [61, 100], where we defined abnormal background parameters as reported by the study authors. This could improve the generalizability of the findings of this meta-analysis as different cut-offs when ascertained under a unified classification (such as a flat trace could correspond with other parameters like low voltage) would aid in interpreting cEEG or aEEG findings.

Amongst the neonates treated with TH, meta-regression showed that the index tests, namely, cEEG and aEEG, influenced the observed heterogeneity. This difference may be attributed to the timing of the index tests, with only one small sample sized study having performed cEEG in the first 1–6 h [115], majority of the others having studied them between 1 and 48 h [83, 99] and 1–72 h [114, 119, 121, 124]. Whereas in neonates who did not receive TH, the severity of HIE based on clinical criteria influenced the observed heterogeneity in the specificity with the inclusion of full spectrum (grade I, II, and III) improving the specificity. Though different reference standards were used for assessment of NDO, subgroup analysis indicated that these did not influence the results. The same was true for the timing of the NDO assessment.

The prognostic utility of cEEG or aEEG has been previously studied [2, 151–153]. Chandrasekaran et al. [2] included nine studies evaluating aEEG within the first 72 h in neonates receiving TH. Their findings demonstrated a high pooled sensitivity of 96% at 6 h, which declined over time, while specificity was 96% at 72 h. Similarly, Del Río et al. [152] included 17 studies that performed aEEG in neonates who received TH and those who did not. This review also demonstrated a similar trend in sensitivity and specificity. Awal et al. [151] and Han et al. [153] evaluated individual EEG parameters (burst suppression, low voltage, and flat or inactive trace). Awal et al. [151] reported a sensitivity of 78%–92% and a specificity of 82%–99%, while Han et al. [153] found sensitivity in the range of 78%–87%, with a specificity being 56%–85%. While Awal et al. [151] analyzed 31 studies including neonates who were treated with TH and those who were not, Han et al. [153] included 18 studies of TH neonates. Differences in the time frame of index testing and inclusion criteria likely contributed to variations in the reported DTA measures in these meta-analyses. There were some major differences between our systematic review and the others. First, these reviews included substantially fewer studies compared to that of ours. Further, we assessed both cEEG and aEEG in combination and separately. Also, we studied the effect of TH vs. no TH on the index tests. Finally, we explored heterogeneity for multiple clinically relevant covariates.

The overall risk of bias was high across the included studies. Future studies should use standardized EEG parameters based on recommended cut-offs [5] while minimizing selection bias and ensuring assessor blinding. Second, evaluating the role of prolonged cEEG in high-income countries (HICs) and aEEG in low- and middle-income countries (LMICs) through interventional diagnostic studies would help assess the impact of such monitoring on long-term NDO, thereby possibly improving the accuracy of prognostic estimates. The inherent difficulty of cEEG interpretation, along with the limited availability of trained EEG interpreters around the clock, poses a significant challenge in its implementation. Therefore, future studies should focus on validating automated cEEG interpretation software and its utility in prognostication.

Our review is the first to assess the CoE for each diagnostic accuracy variable in their prognostic role in neonates with HIE. Further, we also report GRADE assessments for EEG based on whether they received TH or not. The findings of this review may assist in formulating clinical practice guidelines through the evidence to decision parameters of GRADE. With the inclusion of 34 studies enrolling neonates with HIE who received TH and 33 studies evaluating those who did not receive TH, this systematic review with DTA meta-analysis is the most comprehensive one. This DTA meta-analysis was conducted in accordance with the Cochrane Handbook for DTA meta-analysis, thus ensuring methodological rigor. To address variations in the index test threshold, we combined abnormal index test parameters, explored heterogeneity through prespecified subgroup analyses as well as meta-regression, and applied a random-effects bivariate model for data synthesis. Additionally, we incorporated Bayesian methods to estimate the posterior pooled median values, enhancing the robustness of our findings. However, there are many limitations to this systematic review as well. We could not include gray literature, and we did not contact any of the study authors of the included studies for additional data. Furthermore, the influence of seizure burden, anti-seizure medications, and sedative therapies, which may suppress EEG background activity and likely to increase false-positive rates, were not accounted for. Addressing these factors would require access to patient-level data. In addition, we did not assess the EEG background abnormality at multiple time points from individual studies in a standardized manner, given the influence of timing on predictive performance of EEG. Furthermore, pooling studies that used different EEG thresholds may have introduced misclassification bias, leading to inconsistencies in the classification of normal and abnormal EEG patterns.

Conclusion

In this systematic review and DTA meta-analysis, EEG performed within the first 72 h showed a good sensitivity but low specificity for predicting NDI in neonates undergoing TH. In contrast, in neonates with HIE who did not receive TH, EEG demonstrated good sensitivity and acceptable specificity. Subgroup analysis suggested cEEG had improved accuracy compared with aEEG in neonates receiving TH, whereas in those not undergoing TH, both cEEG and aEEG had similar accuracy. To conclude, in the clinical pathway of prognostication of neonates with HIE, the high sensitivity of cEEG and aEEG makes it a useful early prognostic tool for stratification of high-risk neonates.

Statement of Ethics

The paper is exempt from Ethical Committee approval as it is a review of previously published literature.

Conflict of Interest Statement

Daniele Trevisanuto was a member of the journal’s Editorial Board at the time of submission. The other authors have no conflict of interest to declare.

Funding Sources

The authors did not receive any funding.

Author Contributions

Vijay Kumar Krishnegowda and Daniele Trevisanuto conceptualized the systematic review. Vijay Kumar Krishnegowda, Prathik Bandiya, Tapas Bandyopadhyay, and Arun Prasath were responsible for data collection. Vijay Kumar Krishnegowda and Viraraghavan Vadakkencherry Ramaswamy did the analysis and interpretation of data. Vijay Kumar Krishnegowda provided the first draft of the manuscript and had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Daniele Trevisanuto revised the initial draft. Viraraghavan Vadakkencherry Ramaswamy, Tapas Bandyopadhyay, and Thangaraj Abiramalatha critically reviewed the initial manuscript draft by providing further intellectual inputs. All authors approved the final version for submission.