Amplitude‐integrated electroencephalography compared with conventional video‐electroencephalography for detection of neonatal seizures

Abstract

Background

Conventional video‐electroencephalography (cEEG) is the reference standard for diagnosing and managing neonatal seizures. However, continuous bedside cEEG services are not available in most neonatal units. Hence, an alternative and relatively simple method called amplitude‐integrated EEG (aEEG), which uses a limited number of scalp electrodes, has become popular. aEEG allows continuous bedside monitoring of the electrical activity of the brain in neonates.

Objectives

The primary objective of the review was to assess the accuracy of aEEG against the reference standard cEEG for the detection of 'neonates with seizures' and 'individual seizures'. Detection of 'neonates with seizures' refers to the ability of the test to correctly identify a 'neonate' as 'seizure positive' or 'seizure negative' based on the detection of at least one seizure episode in the entire aEEG recording. Detection of 'individual seizures' refers to the ability of the test to correctly identify 'individual' seizure episodes within the same neonate rather than just diagnosing the neonate as 'seizure positive' or 'seizure negative'.

Search methods

We searched CENTRAL, MEDLINE, Embase, clinical trials registries, and grey literature (Open Grey, Trove, and American Doctoral Dissertations) to 26 July 2022. We did not apply any language or publication status restrictions or any other filters.

Selection criteria

We included prospective and retrospective studies investigating the accuracy of aEEG (index test) against the reference standard cEEG for the detection of neonatal seizures. To be eligible for inclusion, the studies must have compared aEEG with simultaneously recorded cEEG. There was no restriction on the number of leads, use of raw EEG traces, or experience and training of an aEEG interpreter. cEEG should have been recorded using at least nine electrodes and interpreted by a qualified person experienced in the interpretation of neonatal cEEG.

Data collection and analysis

Working independently, two review authors collected data from the included studies in a prespecified form and assessed the quality of the included studies using the QUADAS‐2 tool.

For the outcome of 'neonates with seizures', we used a bivariate mixed‐effects regression model to conduct a meta‐analysis to derive pooled sensitivity, specificity, positive and negative likelihood ratios (LR), and their respective 95% confidence intervals (CI). We generated a summary receiver operating characteristic (SROC) curve to display the results of individual studies. We calculated post‐test probabilities based on Bayes' theorem through Fagan nomograms.

For the outcome of 'individual seizures', pooling of data was not possible because of the 'unit of analysis' issue. Instead, we performed a narrative synthesis.

We assessed the certainty of the evidence using GRADE guidelines.

Main results

We included 16 studies (562 infants) in the systematic review, of which only two studies interpreted the aEEGs prospectively at the bedside. Out of 16 studies, three studies (97 infants) described the accuracy of aEEG only for detecting ‘infants with seizures’, three studies (72 infants) described only ‘individual seizures’, while 10 studies (393 infants) described the accuracy of aEEG for detecting both.

Ten of 16 studies were conducted in term and late preterm infants. Half of the included studies did not use raw EEG traces. Fourteen studies reported outcomes based only on retrospective interpretation. Ten of 16 studies used four electrodes (making this the most common approach amongst the included studies), and 10 studies' aEEG recordings exceeded six hours. Only two included studies used a seizure detection algorithm. In 14 studies, a neonatal or neurology consultant performed aEEG interpretation, and most (in 10 of 14 studies) had prior experience in aEEG interpretation.

Accuracy of aEEG to diagnose 'neonates with seizures'. The only two prospective studies (53 participants) which interpreted aEEGs 'live' at the bedside, reported sensitivities of zero and 0.57 and specificities of 0.82 and 0.92, respectively.

Meta‐analysis of 13 studies (490 neonates) found that aEEG had a pooled sensitivity of 0.71 (95% CI 0.57 to 0.83), specificity of 0.84 (95% CI 0.59 to 0.95), positive LR of 4.50 (95% CI 1.55 to 13.04), and negative LR of 0.34 (95% CI 0.22 to 0.53) for the detection of 'neonates with seizures'. However, when we analysed only studies with low risk of bias (3 studies), sensitivity (0.56, 95% CI 0.02 to 0.99) and specificity (0.78, 95% CI 0.60 to 0.90) were even lower.

There was significant statistical heterogeneity, which could not be explained based on threshold effect and exploratory analyses of forest plots. We graded the certainty of the evidence as low, in view of the high or unclear risk of bias in many studies, imprecision, and significant heterogeneity.

Accuracy of aEEG to detect 'individual seizures' in neonates. The reported sensitivities of aEEG for the detection of 'individual seizures' ranged from 0 to 0.86 (13 studies, 465 neonates). We rated the certainty of the evidence as low.

The common causes of missed seizures on aEEG (i.e. false negative) as reported by studies were short duration of seizures, localisation of seizures away from aEEG leads, low voltage, and inexperienced interpreter. The false‐positive rates were high when interpreted live at the bedside and if the interpreters were inexperienced. Artefacts resulting from muscle movement, patting, hiccups, and insufficient electrode attachment were other common causes of false‐positive results.

Authors' conclusions

Low‐certainty evidence suggests that aEEG has only moderate sensitivity and specificity for detecting 'neonates with seizures', and its ability to detect 'individual seizures' varies widely. These findings suggest that aEEG may not be sufficiently accurate for diagnosing neonatal seizures as it can under‐diagnose or over‐diagnose seizures. Studies with low risk of bias are needed to address the issue definitively.

Plain language summary

How accurate is a simplified and easily available electroencephalography (called amplitude‐integrated EEG) in detecting seizures in newborn babies?

Key messages

• Amplitude‐integrated electroencephalography (aEEG) may not be sufficiently accurate to identify babies with seizures and also individual seizure episodes in a baby.

• Treatment based only on aEEG may lead to inadequate or unnecessary treatment with anti‐seizure medications.

Why is it important to accurately detect seizures?

Seizure is an abnormal and short‐term electrical activity of the brain. They are uncommon, but a serious problem in the newborn period. They occur in response to problems such as reduced oxygen or blood supply to the brain, low blood sugar, and brain infections. Babies can have a single or many seizures. Each seizure can be as short as 10 seconds or as long as several minutes.

Seizures in newborn babies can harm the brain and have long‐lasting effects. Therefore, it is important to identify seizures accurately.

What is the aEEG test?

Electroencephalography (EEG) is the non‐invasive measurement of the brain’s electric signals. Recording the electrical activity of the brain with 10 to 20 leads (small wires attached to the head) is the best way to detect seizures. This is called conventional electroencephalography, or cEEG. However, cEEG is not always easily available in neonatal intensive care units, is difficult to perform, and needs a high level of expertise to interpret.

Amplitude‐integrated electroencephalography (aEEG) is a simplified and easy‐to‐perform form of EEG but provides less information than cEEG. aEEG is prepared from EEG signals and needs only 2 to 4 leads, which neonatal nurses can apply. Bedside doctors caring for babies can interpret aEEG to detect seizures.

There are several different aEEG recording machines: some use 2 electrodes while others use 4 electrodes. Some machines also show the original EEG signal from which aEEG was prepared. aEEG can be prepared from already recorded cEEG for research purposes.

What did we want to find out?

How accurate is aEEG in identifying babies who are having seizures and individual seizure episodes in a baby?

What did we do?

We searched for studies that compared aEEG with cEEG for detecting seizures in newborn babies. We read the details of how the studies were conducted and their findings. We summarised the results and estimated their reliability.

What did we find?

We found 16 relevant studies with a total of 562 newborn babies. Out of 16 studies, 3 described the accuracy of the aEEG only for identifying babies with seizures, 3 described it only for detecting individual seizures, and 10 for detecting both babies with seizures and individual seizures.

The duration of recording, the number of aEEG leads, the use of unprocessed EEG signals, and the training and experience of aEEG interpreters were variable in the studies.

In only 2 studies, bedside doctors read the aEEG, reflecting real‐life clinical practice. In the rest of the studies, newborn baby specialists interpreted the aEEG at a later stage. The method of later interpretation of the aEEG is not very useful for the immediate management of the babies.

On average, aEEG detected 71 out of 100 babies who were having seizures. That means it missed 29 babies with seizures. It labelled 16 babies as having seizures when, in fact, they did not have seizures.

Included studies reported variable accuracy of aEEG in detecting individual seizures, ranging from as low as 0 to as high as 86 correctly detected seizures out of 100.

The review findings suggest that aEEG may not be sufficiently accurate for identifying babies with seizures and individual seizure episodes in a baby.

What are the limitations of the evidence?

The review had several limitations, the most important of which was that the results of studies varied widely without any obvious reason, decreasing the reliability of the evidence. Another limitation of the evidence was that only 2 studies reported the effect of real‐life bedside aEEG interpretation on the accuracy of aEEG.

How current is this review?

The evidence is current to July 2022.

Summary of findings

Summary of findings 1. Summary of findings: should amplitude‐integrated EEG be used to detect 'neonates with seizures'?

Patient or population: newborn infants at risk of seizures
Prior testing: none required
Setting: neonatal intensive care unit
Study design: prospective and retrospective studies investigating the accuracy of aEEG (index test) against the reference standard cEEG for detecting neonatal seizures.
Index test: amplitude‐integrated electroencephalography
Reference standard: conventional video‐electroencephalography
Target condition: neonates with seizures
Importance: identifying “neonates with seizures” will help prioritise the infants who need close clinical monitoring and management
Pooled sensitivity: 0.71 (95% CI 0.57 to 0.83); Pooled specificity: 0.84 (95% CI 0.59 to 0.95)
Limitations: unexplained heterogeneity in the study results and variable duration of recording, < 6 hours (4 studies), > 6 hours (7 studies), and unknown (2 studies)
Test result	Number of results per 1000 infants tested (95% CI)			Number of participants with cEEG seizures/Total number of participants (Number of studies)	Certainty of evidence (GRADE)	Comments
	Prevalence 15%a Typically seen in studies with low true seizure rate	Prevalence 25%a Typically seen in studies with medium true seizure rate	Prevalence 35%a Typically seen in studies with high true seizure rate
True positives	107 (85 to 124)	178 (143 to 208)	248 (199 to 291)	300/490 (13)	⨁⨁◯◯ Lowb,c	These infants are correctly identified as having seizures. Therefore, the clinician can then arrange close clinical monitoring and further testing with cEEG to identify individual seizure episodes. The diagnosis will help to treat seizures with appropriate anti‐seizure medications.
False negatives	43 (26 to 65)	72 (42 to 107)	102 (59 to 151)			These infants are not identified as having seizures despite the fact they were having seizures. Therefore, they may not receive close clinical monitoring and further testing to identify individual seizures leading to undertreatment with anti‐seizure medications.
True negatives	714 (501 to 708)	630 (443 to 712)	546 (384 to 617)	300/490 (13)	⨁⨁◯◯ Lowb,c	These infants are correctly identified as not having seizures. Therefore, unnecessary further investigations to identify individual seizures can be avoided.
False positives	136 (42 to 349)	120 (38 to 307)	104 (33 to 266)			These infants were wrongly labelled as having seizures when, in fact, they were not having seizures. They will be falsely labelled as infants with seizures, may receive unnecessary anti‐seizure medications, and undergo testing to identify seizures.

aEEG: amplitude‐integrated electroencephalography; cEEG: conventional video‐electroencephalography; CI: confidence interval

^aThe prevalence frequencies of 15%, 25%, and 35% are based on studies by Buttle Retro 2019, Rakshasbhuvankar 2017, and Shah 2008, respectively. 
^bDowngraded by one level for some study limitations and imprecision, not serious enough to downgrade by one level each.
^cDowngraded by one level for serious heterogeneity.

Summary of findings 2. Summary of findings: should amplitude‐integrated EEG be used to detect 'individual seizures' in high‐risk infants?

Patient or population: neonates at risk of seizures
Setting: neonatal intensive care unit
New test: amplitude‐integrated electroencephalography (aEEG)
Reference standard: conventional video‐electroencephalography (cEEG)
Target condition: individual seizure episode
Importance: identifying 'individual seizures' will help to detect seizures that may not be identified by clinical observation alone and will also help to rule out clinical phenomena that mimic seizures.
Study designs: prospective and retrospective studies
Limitations: unexplained heterogeneity in the study results, inability to calculate specificitya and pool the data for sensitivity due to unit of analysis issuea
Test resultsa	Interpretation	Number of participants (studies)	Certainty of the evidence (GRADE)
The studies calculated the sensitivity of aEEG for detecting 'individual seizures' by dividing the total number of seizures correctly detected using aEEG by total number of cEEG seizures in the "cohort", thus ignoring the 'unit of analysis' issue.a The reported sensitivities ranged from 0% to 86%. The proportion of false‐positive seizures amongst all seizures detected on aEEG varied across studies and ranged from 0% to 100%.	On average, aEEG detects half of the cEEG seizures and may also falsely detect seizures when they are absent. The accuracy of aEEG is likely to vary between infants and may also depend on technical issues such as number of leads used, availability of the raw trace and seizure‐detection software, experience and training of aEEG interpreter, and other factors. Therefore, aEEG may not be sufficiently accurate for diagnosing 'individual seizures' as it may under‐ or over‐diagnose seizures.	4210 seizures in 465 infants (13 studies)	⨁⨁◯◯ Lowb,c

^aDiagnostic or experimental studies are designed to make inferences about 'independent' experimental units, i.e. infants in our case. Multiple observations taken from the same experimental unit tend to be more similar than observations from different experimental units; i.e. multiple seizures in the same infant tend to be more similar than seizures in different infants. Therefore, calculating sensitivity by adding all seizures in a cohort is inappropriate and will lead to a 'unit of analysis' issue (Parsons 2018). 
In addition, we did not report specificity for the outcome of 'individual seizures' due to an inability to meaningfully define 'true‐negative' episodes. A 'negative seizure' has no independent existence; it is the segment of EEG in between two seizure episodes. There is no consensus on how to define a 'true‐negative' seizure for diagnostic accuracy studies. 
^bDowngraded by one level in view of serious imprecision. Not downgraded for risk of bias as most risks of bias occurred in the 'participant selection' domain, which we judged to be less likely to affect the outcome of detection of 'individual seizures'.
^cDowngraded by one level for unexplained heterogeneity.

Background

Electroencephalography is the non‐invasive measurement of the brain’s electric signals (Biasiucci 2019). Conventional video‐electroencephalography (cEEG) is recorded in neonates with at least nine surface electrodes placed on the scalp using a 10 to 20 electrode placement system and accompanied by a video to ascertain the clinical correlates of seizures. cEEG is the reference standard for diagnosing neonatal seizures. However, continuous bedside cEEG monitoring and interpretation services are not available in most neonatal units. Hence, an alternative and relatively simple method called amplitude‐integrated EEG (aEEG), which uses fewer scalp electrodes, has become popular. aEEG allows continuous bedside monitoring of the electrical activity of the brain in neonates (Hellstrom‐Westas 2018). However, there is debate regarding the accuracy of the aEEG to diagnose neonatal seizures.

A systematic review without meta‐analysis included 10 studies (433 infants) and reported a median sensitivity of 76% (range: 71 to 85) and specificity of 85% (range: 39 to 96) for the detection of ‘individual seizures’ when aEEG with raw trace was used, and sensitivity of 39% (range: 25 to 80) and specificity of 95% (range: 50 to 100) when aEEG was used without raw trace (Rakshasbhuvankar 2015). The authors, appropriately, did not pool the sensitivity data for the outcome of ‘individual seizures’. The authors included a study by Toet and colleagues that used an unconventional definition of seizures (epileptic discharges of 6 to 9 seconds were also considered as seizures) (Toet 2002).

Another systematic review and meta‐analysis included 14 studies with a total sample size of 550 infants and reported an overall sensitivity of aEEG between 31.25% and 90% (Falsaperla 2022). The review, however, had multiple limitations. It included two studies, Lee 2021 and Toso 2014, that were not relevant. Lee 2021 reported on the EEG and aEEG profiles of sick neonates admitted into the neonatal intensive care unit (NICU) and their correlation with neurological outcomes. The Toso 2014 study, in which no simultaneous cEEG was recorded, reported the accuracy of aEEG for predicting abnormal neurological outcomes. In addition, the Falsaperla 2022 review included the Toet 2002 study, which used an unconventional definition of seizure. The review also made no distinction between 'individual seizures' and 'patients with seizures', and included data from one study in a meta‐analysis multiple times (Shah 2008).

Considering the limitations of the previous systematic reviews, it was essential to synthesise the current evidence of aEEG accuracy for detecting neonatal seizures using a robust methodology.

Target condition being diagnosed

A seizure is a paroxysmal transient occurrence of abnormal excessive or synchronous neuronal activity in the brain of at least 10 seconds' duration (Volpe 2008). Seizures associated with physical signs are called 'clinical' or 'electro‐clinical' seizures, while those without physical signs are called 'subclinical', 'electrographic‐only', 'silent', or 'occult' seizures. Seizures are one of the most common manifestations of central nervous system dysfunction in newborn infants. Detection of individual seizure episodes is important to assess the need for seizure medications and help with prognostication. Hypoxic‐ischaemic encephalopathy (HIE) is the most common cause of neonatal seizures. Other causes include cerebral infarction, intracranial haemorrhage, meningitis and septicaemia, acute metabolic disturbances like hypoglycaemia and hypocalcaemia, maternal drug withdrawal, inborn errors of metabolism, epilepsy syndromes, and idiopathic aetiology (Vasudevan 2013). The risk of seizures in the neonatal period is higher than in any other period, with a reported incidence between 1 and 5 per 1000 live births (Silverstein 2008; Vasudevan 2013).

The detection and monitoring of seizures is a vital component of neurocritical care for the following reasons: (a) seizures could adversely affect vital functions, including respiration and circulation; (b) seizures may disrupt the laying down of neural circuitry and hence adversely affect neurodevelopmental outcomes (Miller 2002; Nagarajan 2010; Nardou 2013; Sankar 2007; Shah 2014; Srinivasakumar 2015; Van Rooij 2010; Wirrell 2001); and (c) seizures may have a kindling effect on the neonatal brain, making the infant prone to develop seizures later in life (Levene 2002).

Clinical observation alone can lead to under‐diagnosis since 65% to 80% of neonatal seizures are occult (Clancy 2005; Murray 2008; Nagarajan 2011). It can also result in over‐diagnosis since many clinically suspected episodes will not have corresponding electrographic correlation (Murray 2008). Evidence from animal studies has shown that anti‐seizure medications can cause apoptosis of developing neurones (Bittigau 2002; Kim 2007). Hence, over‐diagnosis of seizures could result in inappropriate and excessive use of anti‐seizure medications, thereby adversely affecting the infant's developing brain. As reliance on clinical observation alone is inadequate, it is essential to monitor brain electrical activity to detect neonatal seizures. Electrical activity of the brain can be monitored using either cEEG or aEEG to detect neonatal seizures (Boylan 2010).

cEEG is the reference standard for detecting neonatal seizures (Wusthoff 2025). cEEG uses the full complement of scalp electrodes applied according to the International 10‐20 System of Electrode Placement (modified for neonates) by skilled technologists and interpreted by experienced neurologists or clinical neurophysiologists (Wusthoff 2025). The American Clinical Neurophysiology Society has recommended that neonates at high risk of seizures should be monitored with cEEG for 24 hours to screen for seizures (Wusthoff 2025). However, due to resource limitations, 24‐hour cEEG monitoring is not feasible in many centres (Boylan 2010; Buttle 2019; Dilena 2019). In addition, the cEEG results are often not available in real‐time to assist with patient management (Jobe 2009).

If aEEG or cEEG recording does not show any seizures, it is called a seizure‐negative recording. If aEEG or cEEG recording shows at least one seizure, it is called a seizure‐positive recording. In a seizure‐positive recording, there may be a single seizure or multiple seizures. 'Negative' seizure (in other words: 'no seizure') is the segment of EEG or aEEG in between two 'positive' seizures (in other words: 'seizures'). Therefore, as the number of seizures (i.e. 'positive seizures') increases, the number of 'negative seizures' also increases (Negative seizures = Positive seizure ± 1). The issue of dependence of 'negative seizures' on positive seizures is complicated by difficulties in converting a continuous variable into a categorical one. Therefore, there is no consensus on how to define a 'true‐negative' seizure episode for calculating accuracy indices for 'individual seizure' episodes.

Index test(s)

aEEG is a simplified method of monitoring the electrical activity of the brain using commercially available portable bedside machines. However, for research purposes, it can be derived from an already recorded cEEG. The effect of such a “derived” aEEG on the accuracy of interpretation is not known.

For generating aEEG, brain electrical activity, called ‘raw trace’, is recorded by an aEEG machine, and then filtered, rectified, and compressed in time (Shah 2008b). The aEEG display is in the form of a band with the upper and lower margins corresponding to the highest and lowest voltage of the raw trace. The pattern and voltage of the upper and lower margins of the band are used to classify the background pattern. Seizures appear as transient elevations or peaks in the lower margin of the band, often accompanied by the upper margin. During an episode of a seizure, the raw (unprocessed) EEG signal shows repetitive evolving spikes and waveforms lasting for at least 10 seconds with a definite beginning, middle, and end. Commonly used aEEG machines utilise two or four electrodes, giving rise to one or two aEEG bands, respectively.

The main advantages of aEEG over cEEG are the following: fewer scalp electrodes that can be applied by neonatal staff and maintained for extended periods for continuous monitoring; the output can be interpreted by neonatal staff at the bedside; and less time is required to review the trace because of aEEG's time‐compressed nature. Because of these advantages, aEEG is gaining popularity worldwide in NICUs (Shah 2008b). However, the use of fewer leads and the time‐compressed nature of aEEGs are also responsible for the following drawbacks: seizures with short duration, especially those less than 30 seconds, may be difficult to identify; localised seizure activity in brain areas (such as frontal and temporal lobes) away from the aEEG leads is likely to be missed; artefacts, such as those arising from leads, movements, and the like, may be misinterpreted as seizures; and there is a tendency to underestimate the duration of seizures (Hellstrom‐Westas 2018; Rakshasbhuvankar 2017).

The newer aEEG monitoring systems display one or more channels (bands) of aEEG along with the unprocessed raw EEG trace to assist in the diagnosis of seizures. In addition, algorithms to help seizure detection have been developed (El‐Dib 2009; Lawrence 2009; Lommen 2007; Navakatikyan 2006). C3, P3, C4, and P4 lead positions of the International 10‐20 system are commonly used since they are close to the watershed area of the brain, which is commonly affected in infants with HIE, and these positions are less susceptible to artefacts.

The number and position of aEEG leads, the availability of raw traces, the use of seizure detection algorithms, conditions of interpretation such as bedside real‐time versus retrospective, and experience and training of interpreters may influence the accuracy of aEEG for identifying seizures.

Clinical pathway

High‐risk neonates require careful monitoring for the detection of seizures. The risk factors for seizures include HIE, meningitis, intracranial bleeding, cerebral infarction, metabolic encephalopathy, cardiac surgery, and critical illness. Infants at risk of seizures should ideally be monitored using cEEG (Abend 2013), since this approach will provide the most accurate information regarding the diagnosis and seizure burden, enabling management decisions. However, this approach requires considerable infrastructure and cost, which may not be available in many neonatal units (Boylan 2010).

(1) If aEEG has high sensitivity and specificity for detecting neonatal seizures AND

1. If 24X7 (i.e. 24 hours a day, seven days a week) cEEG monitoring is available, aEEG may be able to replace cEEG for the detection and management of neonatal seizures with anti‐seizure medications. aEEG monitoring is more efficient, economical, and requires less resources compared with cEEG monitoring.

2. If 24X7 cEEG monitoring is NOT available, aEEG will provide an excellent diagnostic tool to monitor and treat seizures.

(2) If aEEG has high sensitivity but low specificity AND

1. If 24X7 cEEG monitoring is available, aEEG could be used as a screening tool to prioritise babies for cEEG monitoring, thus enabling better use of cEEG resources.

2. If 24X7 cEEG monitoring is NOT available, aEEG may be helpful in identifying infants who require close clinical monitoring and also in identifying seizures.

(3) If aEEG has low sensitivity but high specificity AND

1. If 24X7 cEEG monitoring is available, aEEG will not add any benefit to cEEG monitoring for the purpose of detecting neonatal seizures.

2. If 24X7 cEEG monitoring is NOT available, aEEG may be helpful as an adjunct to clinical observation for detecting and managing seizures.

(4) If aEEG has low sensitivity and specificity AND

1. If 24X7 cEEG monitoring is available, aEEG will not add any benefit to cEEG monitoring for the purpose of detecting neonatal seizures.

2. If 24X7 cEEG monitoring is NOT available, aEEG will not add any benefit to clinical observation for the purpose of detecting neonatal seizures.

Prior test(s)

No prior tests are expected. The high‐risk infants are identified based on history (e.g. birth asphyxia), investigations (e.g. cerebrospinal fluid examination suggestive of meningitis), or physical examination (e.g. signs of encephalopathy).

Role of index test(s)

An index test (aEEG) is used to identify electrographic seizures to recognise 'neonates with seizures' and 'individual seizures'.

aEEG’s ability to detect ‘neonates with seizures’ will be helpful in identifying infants who require close monitoring, further investigations such as cEEG and neuroimaging, and therapy such as therapeutic hypothermia.

aEEG’s ability to detect individual seizures will be helpful in identifying aetiology based on specific seizure patterns (Pijpers 2023), treatment with anti‐seizure medications, and prognostication based on total seizure burden (Chalak 2021; Pisani 2020).

Alternative test(s)

There are no alternative tests to diagnose neonatal seizures apart from cEEG and aEEG.

Rationale

Bedside monitoring of aEEG background pattern helps predict neurodevelopmental outcomes in newborn infants with HIE (Van Laerhoven 2013). However, the usefulness of aEEG for detecting neonatal seizures is not well established. The need for a rigorous evaluation of aEEG for the diagnosis of neonatal seizures is well recognised (Freeman 2007; Rakshasbhuvankar 2015; Sanchez Fernandez 2015; Silverstein 2008). The published systematic reviews have limitations, as described earlier (Falsaperla 2022; Rakshasbhuvankar 2015). Therefore, we conducted this systematic review to synthesise the current evidence regarding the accuracy of aEEG versus the reference standard cEEG for the diagnosis of 'neonates with seizures' as well as 'individual seizures' in neonates.

Objectives

The primary objective was to assess the accuracy of aEEG against the reference standard cEEG for the detection of 'neonates with seizures' and 'individual seizures'.

1. Detection of 'neonates with seizures' refers to the ability of the index test (aEEG) to correctly identify a 'neonate' as 'seizure positive' or 'seizure negative' based on the detection of at least one seizure episode in the entire aEEG recording.

2. Detection of 'individual seizures' refers to the ability of the index test (aEEG) to correctly identify all 'individual' seizure episodes within the same neonate rather than just diagnosing the neonate as 'seizure positive' or 'seizure negative'. Identifying 'individual seizure' episodes is critical because a higher seizure burden is associated with worse outcomes in neonates.

Secondary objectives

Our secondary objective was to investigate variation in the accuracy of aEEG to address the potential sources of between‐study heterogeneity listed below.

1. To compare aEEG in term versus preterm infants

2. To compare risk factors for seizure (HIE versus other causes)

3. To compare 'aEEG without raw trace' versus 'aEEG with raw trace'

4. To compare 'aEEG with raw trace' versus 'aEEG with raw trace and seizure detection algorithm'

5. To compare single versus two versus more than two channels

6. To compare different aEEG lead positions

7. To compare surface electrodes versus needle electrodes

8. To investigate the effect of training (yes/no) in the interpretation of aEEG

9. To investigate the impact of clinicians' experience (yes/no) in the interpretation of aEEG

10. To compare prospective ('real‐time' or 'bedside') versus retrospective interpretation

11. To study the effect of the duration of aEEG recording (≤ six hours versus > six hours) on the detection of 'neonates with seizure.' This is important as six hours is the cut‐off point to decide whether infants with HIE require therapeutic hypothermia (Shankaran 2005; Shankaran 2011).

12. To compare studies with high versus low risk of bias

Methods

Criteria for considering studies for this review

Types of studies

We included prospective and retrospective studies evaluating the accuracy of aEEG against the reference standard cEEG for the detection of neonatal seizures. To be eligible for inclusion, the studies must have compared aEEG with a simultaneously recorded cEEG.

We excluded studies that focused only on the background EEG pattern without addressing the detection of neonatal seizures and studies in which there was no simultaneous recording of aEEG and cEEG.

The duration of aEEG monitoring can influence the outcome of detection of 'neonates with seizure'. Ideally, aEEG should be able to detect seizures as soon as possible, but no one knows the critical time window within which seizures should be identified and treated. Currently, there are no standard guidelines for the duration of aEEG monitoring in 'at‐risk' neonates. In clinical practice, the duration of monitoring is variable and is guided by risk factors for seizures (e.g. HIE), frequency of continuing seizures, seizure‐free period, aEEG background activity, and local practices (Shah 2015). Hence, we included studies with any duration of simultaneous monitoring with aEEG and cEEG.

Participants

The study participants were newborn infants (postmenstrual age under 44 weeks) admitted to NICU with suspected seizures or at risk of seizures. The risk factors included HIE, infectious or metabolic encephalopathy, intraventricular haemorrhage, intracranial malformations, sick infants receiving muscle paralysis, and cardiac surgery (Clancy 2005; Shah 2012). In some studies, initially, neonates were confirmed to have had seizures based on the 'reference standard' cEEG and subsequently were monitored with simultaneous aEEG and cEEG to identify ongoing seizure activity.

Index tests

The index test was aEEG. We included studies in which a separate machine was used to record aEEG, as well as studies in which aEEG was derived from signals from the cEEG recordings. We included studies in which aEEG was interpreted 'real‐time' (bedside) and studies in which aEEG was analysed retrospectively. We did not apply any restrictions regarding the aEEG interpreter qualification, experience, or training. To be clinically applicable, aEEG should be interpreted by neonatal staff; ideally, prospectively by bedside neonatal staff.

aEEG could be either single or multichannel EEG systems.

aEEG could have been recorded with or without raw EEG traces.

aEEG could have been measured with or without automated seizure detection algorithms.

Target conditions

The target condition for the review was electrographic seizures of at least 10 seconds' duration. A diagnostic test should identify patients with the target condition (detection of 'neonates with seizures') and every time a target condition occurs in a patient (detection of 'individual seizures'). The definitions of 'neonates with seizures' and 'individual seizures' are set out in the Objectives.

Reference standards

The reference standard was cEEG recorded using at least the following nine electrodes (Fp1, Fp2, C3, C4, Cz, T3 (or T7), T4 (or T8), O1, and O2) and interpreted by neurologists or clinical neurophysiologists (or equivalent personnel) experienced in the interpretation of neonatal cEEG (Tekgul 2005).

Search methods for identification of studies

Electronic searches

We searched the following databases:

1. Cochrane Central Register of Controlled Trials (CENTRAL) (CRS Web, 26 July 2022) using the search strategy shown in Appendix 1.

2. MEDLINE (Ovid, 1946 to 25 July 2022) using the search strategy shown in Appendix 2.

3. ClinicalTrials.gov (26 July 2022) using the search strategy shown in Appendix 3.

4. World Health Organization: International Clinical Trials Registry Platform(ICTRP, 26 July 2022) using the search strategy shown in Appendix 4.

5. Embase (Ovid, 1980 to 26 July 2022) using the search strategy shown in Appendix 5.

6. Grey literature: Open Grey (www.opengrey.eu; grey literature database for Europe); www.trove.nla.gov.au (Trove: Australian online library database aggregator); American Doctoral Dissertations (available from EBSCOhost).

The CRS Web includes randomised or quasi‐randomised controlled trials from PubMed, Embase, ClinicalTrials.gov, the World Health Organization International Clinical Trials Registry Platform (ICTRP), the Cochrane Central Register of Controlled Trials (CENTRAL), and the Specialised Registers of Cochrane Review Groups, including Epilepsy. In MEDLINE (Ovid), the coverage end date always lags a few days behind the search date.

We did not apply any language or publication status restrictions or any other filter.

Searching other resources

We searched the reference lists of primary articles for relevant publications not captured by the literature search.

Data collection and analysis

Selection of studies

Two review authors (AR and SR) independently reviewed the titles and abstracts of all articles obtained in the broad initial screening to identify potentially eligible studies. The same two review authors independently retrieved and read the full texts of all articles deemed to be potentially eligible for detailed screening. AR and SR resolved any disagreement regarding study selection by discussing it with the review co‐authors (ZZ and LN).

Data extraction and management

Two review authors (AR and SR) independently collected data from the included studies in a prespecified form (Appendix 6), which had the following information: title of the article, author names, year of publication, sample size, study participant characteristics, details of cEEG (leads, montages, availability of video) and aEEG (leads, channels, availability of raw trace and seizure detection algorithm), characteristics of clinicians who interpreted the aEEG and cEEG, duration of simultaneous recording of aEEG and cEEG, data related to seizure detection including 2x2 tables, limitations, and original authors' conclusions. A third review author (ZZ) extracted data from included study Rakshasbhuvankar 2017 to avoid a conflict of interest.

For the outcome of the detection of 'neonates with seizure,' we constructed 2x2 tables that detailed the proportion of true‐positive, false‐positive, true‐negative, and false‐negative results. If the studies had not reported adequate information to populate a 2x2 table, we contacted the study authors requesting them to provide the missing data. If the information was unavailable even after contacting the authors, we included the study in the systematic review but excluded it from the meta‐analysis.

For the outcome of the detection of 'individual seizures', we extracted the following data: percentage/number of 'individual seizures' detected correctly using aEEG, and percentage/number of false‐positive seizures on aEEG.

Assessment of methodological quality

Two review authors (AR and ZZ) assessed all the included studies using the QUADAS‐2 (Quality Assessment of Diagnostic Accuracy Studies) tool (Whiting 2011; Appendix 7). For participant selection, if a study recruited consecutive or random infants at risk of seizures, we judged the risk of bias to be low. On the other hand, if the study recruited infants based on cEEG (case‐control design) or aEEG findings, we considered the risk of bias for the domain to be high. For the index test domain, if aEEG was interpreted without the knowledge of cEEG findings, we considered the risk of bias to be low. Signalling questions regarding threshold effect were not relevant to the detection of seizures. aEEG derived from cEEG and retrospective interpretation were considered to be of high concern for applicability, whereas prospective interpretation and aEEG recoded with a separate specialised machine (usual clinical practice) were considered of low concern for applicability. If both the ‘index test’ and ‘reference standard’ were interpreted by the same person, we judged blinding as inadequate.

We resolved any disagreements through discussion with the review co‐authors (SR and LN).

Statistical analysis and data synthesis

One review author (AR) entered data into Review Manager (RevMan) software (RevMan 2025), and a second review author (ZZ) checked the data entry.

Detection of 'neonates with seizures'

We used the Metandi command in Stata SE v16 to conduct a meta‐analysis using a bivariate mixed‐effects regression model to derive individual and pooled sensitivity, specificity, positive and negative likelihood ratios (LR), and their respective 95% confidence intervals (CI) (Stata 2017). We inputted the summary statistics in RevMan to generate a summary receiver operating characteristic (SROC) curve to display the results of individual studies (Bossuyt 2013; Cronin 2018). Since the Metandi and Meqrlogit commands do not have the ability to derive Fagan nomograms and Deeks' test for funnel plots, we used the Midas command to derive them (Dwamena 2007). Midas has been used in Cochrane reviews (Gokulakrishnan 2022).

A rough guide to the interpretation of LR is as follows: LR (+) of > 10 or LR (‐) < 0.1 generate large and often conclusive changes in the post‐test probability, LR (+) from > 5 to 10 or LR (‐) from 0.1 to 0.2 generate moderate shifts in post‐test probability, LR (+) from > 2 to 5 or LR (‐) from > 0.2 to 0.5 generate small (but sometimes important) changes in probability, and LR (+) from > 1 to 2 or LR (‐) from > 0.5 to 1 alter post‐test probability to a very small degree (Bhandari 2003; Rubinstein 2018).

We generated a summary receiver operating characteristic (SROC) curve to display the results of individual studies in the receiver operating characteristic (ROC) space (Bossuyt 2013; Cronin 2018).

We assessed statistical heterogeneity by visually inspecting forest plots and plotting the prediction region in the ROC space. We considered heterogeneity to be high if the 95% prediction region in the ROC space was much larger than the 95% confidence region. The prediction region illustrates the extent of statistical heterogeneity by depicting an area within which, assuming the model is correct, we have 95% confidence that a future study's true sensitivity and specificity should lie.

The Fagan nomogram for diagnostic accuracy was derived using likelihood ratios to calculate post‐test probability based on Bayes' theorem (Fagan 1975).

Detection of 'individual seizures'

An experimental unit is an entity which can be randomly and independently assigned to two diagnostic or intervention arms under study. A study can only make inferences about the experimental unit. In the diagnostic studies included in this review, the experimental unit was a neonatal infant. A neonatal infant may not have any seizures, and if it does have, there may be a few correctly detected seizures (true positive), a few not‐detected seizures (false negative) and a few false‐positive seizures, thus leading to multiple observations regarding 'individual seizures'. Multiple observations or measurements taken from the same experimental unit tend to be more similar than observations taken from different experimental units, leading to the 'unit of analysis' issue (Parsons 2018). Since these data were non‐independent, we did not pool the data through meta‐analysis for the detection of 'individual seizures'. Instead, a narrative synthesis was performed.

Calculation of specificity for 'individual seizures' poses additional challenges. The first problem in calculating the specificity of 'individual seizures' was that specificity requires the number of 'True‐negative' seizures which have no 'independent existence'. 'Negative' seizure (in other words: 'no seizure') is the segment of EEG or aEEG in between two 'positive' seizures (in other words: 'seizure'). Therefore, as the number of seizures (i.e. 'Positive seizures') increases, the number of 'negative seizures' also increases (Negative seizures = Positive seizure ± 1). The second problem was that converting a continuous variable into a categorical one is not straightforward and studies may differ on how they label a segment as ‘True negative’. No study has reported how they defined and labelled an aEEG segment as a 'True negative'. Therefore, it was not possible to calculate specificity for 'individual seizures'.

We created summary of findings tables for the outcomes of the detection of 'neonates with seizures' and 'individual seizures'.

Investigations of heterogeneity

As noted in Statistical analysis and data synthesis, we assessed heterogeneity by visually inspecting forest plots and plotting the prediction region in the ROC space. We considered heterogeneity to be high if the 95% prediction region in the ROC space was much larger than the 95% confidence region. This prediction region illustrates the extent of heterogeneity by depicting an area within which, assuming the model is correct, we have 95% confidence that a future study's true sensitivity and specificity should lie. Due to the small sample size and the lack of relevant data, we were unable to do meta‐regression to explore heterogeneity. We explored heterogeneity by performing sensitivity analyses and visual analysis of forest plots with reference to covariates.

We explored the following sources of heterogeneity.

1. Gestational age (term versus preterm) of infants

2. Risk factors for seizure (HIE versus other causes)

3. Use of 'raw trace' along with aEEG

4. Use of seizure detection algorithm

5. Use of multiple channels

6. Use of different lead positions

7. Use of surface versus needle electrodes

8. Interpretation training (yes/no) in the interpretation of aEEG

9. Impact of experience of aEEG interpretation (yes/no) on the interpretation of aEEG

10. Bedside interpretation compared with retrospective interpretation

11. Duration of monitoring (≤ six hours versus > six hours)

12. Risk of bias

Sensitivity analyses

We planned to perform the following sensitivity analyses for the outcome of the detection of 'neonates with seizures'.

1. Sensitivity analysis by including only those studies with a low risk of bias. If a study had a high risk of bias in any of the four QUADAS‐2 domains of risk of bias (participant selection, index test, reference standard, and flow and timing), we judged the overall risk of bias as high. If a study was not deemed to be high risk but had an unclear risk of bias in any of the four QUADAS‐2 domains of risk of bias, we judged the overall risk of bias as unclear. If a study had a low risk of bias in all four QUADAS‐2 domains, we judged the overall risk of bias as low.

2. Sensitivity analysis by excluding studies in which participants were enroled based on knowledge of cEEG findings. While such studies are not the ideal design, they are still valuable for evaluating the efficacy of aEEG in diagnosing seizures that can subsequently occur in those neonates. Hence, we included such studies in the primary analysis and subsequently conducted a sensitivity analysis by excluding them.

3. Sensitivity analysis by excluding studies in which either aEEG was derived from cEEG or aEEG was interpreted retrospectively. This was done since such studies have applicability concerns and are at high risk of incorporation bias.

We planned to perform additional sensitivity analyses if we observed variations that could alter the accuracy of the index test in study participants, characteristics of the index and reference tests, and methodology.

Assessment of reporting bias

In accordance with Cochrane recommendations, we used Deeks' method for formal testing of publication bias in diagnostic studies with a significant (at 10% level) slope coefficient indicating the presence of publication bias (Deeks 2005). However, we acknowledge that current knowledge of the formal tests for assessing publication bias in diagnostic accuracy reviews is limited (Macaskill 2010). To mitigate the risks due to missed studies, we conducted comprehensive searches, including searches of the grey literature, and contacted experts in the field for unpublished data. If an adequate number of studies with different publication statuses were available (e.g. studies published as full text versus studies published as abstracts), we planned to include publication status as a covariate in the meta‐regression.

Assessment of the certainty of evidence and preparation of summary of findings tables

We assessed the certainty of the evidence using GRADE guidelines, and prepared summary of findings tables using the GRADEpro GDT online application (https://gdt.gradepro.org/app). We assessed the evidence according to the following domains: risk of bias, indirectness, inconsistency, imprecision, and publication bias. We decreased the certainty of the evidence for the domains based on the level of concern. If the concerns were serious, we downgraded the evidence by one level and when the concerns were very serious, we downgraded the evidence by two levels. The GRADE guidelines also recommend adopting a conservative approach when downrating evidence and considering the risk of bias in the context of other limitations (Guyatt 2011). We based our judgements about the level of seriousness on the magnitude of risk posed to the outcome estimate. We documented the reasons for downgrading in the footnotes of the table.

We prepared summaries of findings tables for both primary outcomes. For the outcome of ‘neonates with seizures’, we planned to report the number of infants with true‐ and false‐positive and true‐ and false‐negative diagnoses for low (15%), medium (25%), and high (35%) prevalence of neonates with seizures amongst infants at risk of seizures. The prevalence frequencies of 15%, 25%, and 35% were based on studies by Buttle Retro 2019, Rakshasbhuvankar 2017, and Shah 2008, respectively. For the outcome of ‘individual seizures’, we planned to provide a narrative summary of the findings.

Results

Results of the search

We identified 849 articles following the literature search in March 2020 and 179 additional articles after an updated search on 26 July 2022. After removing duplicates, two authors read the titles and abstracts of the remaining 764 articles to identify 50 potentially eligible articles. The authors reviewed the full texts of the potentially eligible articles and finally included 16 studies (562 participants). We excluded 33 studies (Al Naqeeb 1999; Alfonso 2001; Bogaarts 2014; Bourgoin 2020; Bridgers 1986; Cagdas 2011; Chalak 2021; Clancy 2011; Eken 1995; Falsaperla 2022; Fenichel 1987; Griesmaier 2011; Hellstrom‐Westas 1992; Hirasawa 2009; Jalisatgi 2015; Kadivar 2019; Kidokoro 2013; Klebermass 2001; Lee 2021; Lloyd 2017; Meledin 2017; Osredkar 2006; Quigg 2009; Rakshasbhuvankar 2015; Ray 2011; Shellhaas 2007b; Sisman 2005; Toet 2002; Toso 2014; Van Rooij 2005; Wusthoff 2009; Wusthoff 2017; Zimbric 2011). The reasons for exclusion were as follows: did not investigate the accuracy of aEEG for the detection of neonatal seizures (12 studies); not a primary study – review article (seven studies); no simultaneous recording of cEEG (six studies); aEEG was not investigated – only raw traces were studied (three studies); unable to retrieve relevant data (one study); inadequate reference standard (one study); not related to neonatal seizures (one study); study did not aim to investigate the diagnostic accuracy of aEEG for seizure detection (one study); ineligible definition of seizure (one study). See Characteristics of excluded studies for details. One ongoing study was available only in protocol form and, therefore, could not be included in the current review (NCT05079971) (see Characteristics of ongoing studies). The study selection process is shown in Figure.

1.

Study flow diagram

We contacted the authors of three studies (Mendelsohn 2019; Ponten 2010; Yamamoto 2015), and received the following additional information: number of neonates with seizures and individual seizures, experience and training of aEEG interpreters, number of aEEG leads and use of raw trace (Mendelsohn 2019); number of neonates with seizures, aEEG interpreter qualification, experience and training in aEEG interpretation, gestational age of infants (Yamamoto 2015). We could not obtain any additional information for the Ponten 2010 study.

Methodological details of included studies

The included studies were published between 1994 and 2019. Five studies each were performed in North America and Europe, three in Asia, two in Australia, and one in South America. Characteristics of included studies are summarised in Table.

1. Characteristics of the included studies.

Study, sample size	GAa	Independent aEEG machine	Use of raw trace	Monitoring Duration	Leads	aEEG interpreterb	Trainingc	Experiencec	Neonates with seizure ‐ Sensitivity	Neonates with seizure ‐ Specificity	Individual seizures ‐ Sensitivity
Bourez‐Swart 2009d, n = 12	>35	No	No	< 6 hr	C3,C4	Neonatal consultant	NR	NR	0.91	NR	0.30
Buttle Pros 2019e, n = 13	Any	Yes	Yes	> 6 hr	C3,C4,P3,P4	Fellow	NR	NR	0	0.82	0
Buttle Retro 2019, n = 13	Any	Yes	Yes	> 6 hr	C3,C4,P3,P4	Neonatal consultant	NR	Yes	0.50	0.64	0.13
Eaton 1994d, n = 13	Any	Yes	No	NR	C3,C4	Neonatal consultant	NR	NR	1.00	NR	NR
Evans 2010, n = 44	Any	No	No	> 6 hr	O1,O2,C3,C4	Neurology consultant	NR	NR	0.80	0.50	NR
Frenkel 2011, n = 38	Any	Yes	Yes	< 6 hr	P3,P4	Neonatal consultant	NR	NR	0.90	0.93	0.76
Lawrence 2009f, n = 34	>35	Yes	Yes	> 6 hr	C3,C4,P3,P4	Neurology consultant	Yes	Yes	NR	NR	0.76
Mastrangelo 2013d, n = 28	>35	No	No	> 6 hr	C3,C4,T3,T4	Neurology consultant	No	No	NR	NR	0.38
Mendelsohn 2019e,g, n = 40	>35	Yes	Yes	> 6 hr	C3,C4,T3,T4	Fellow	No	Yes	0.57	0.92	0.46
Ponten 2010d,h, n = 10	>35	NR	No	> 6 hr	NR	Neonatal consultant	NR	Yes	NR	NR	0.62
Rakshasbhuvankar 2017f, n = 35	>35	Yes	Yes	> 6 hr	C3,C4,P3,P4	Neonatal consultant	Yes	Yes	1.00	0.75	0.34
Rennie 2004, n = 40	Any	No	No	NR	P3,P4	Neonatal consultant	Yes	No	0.37	0.90	NR
Schmidt 2015d, n = 25	>35	NR	NR	> 6 hr	NR	Neonatal consultant	NR	Yes	0.40	NR	0.29
Shah 2008, n = 21	>35	Yes	Yes	> 6 hr	C3,C4,P3,P4	Neonatal consultant	No	Yes	0.86	0.86	0.76
Shellhaas 2007, n = 121	>35	No	No	< 6 hr	C3,C4	Neonatal consultant	NR	Yes	0.40	1.00	0.25
Yamamoto 2015d,g, n = 26	Any	No	No	< 6 hr	C3,C4,Fp1,Fp2	Neurology consultant	No	Yes	0.81	NR	0.60
Zhang 2011, n = 62	>35	No	Yes	> 6 hr	C3,C4,P3,P4	Electroencephalographer	NR	Yes	0.71	NR	0.86

R: reported; NR: not reported or could not be calculated

^aGestational age in weeks
^bMost senior interpreter qualification where aEEG was interpreted by multiple readers
^cRefers to training or experience in aEEG interpretation
^dStudy did not report funding source
^eProspective interpretation of aEEG by bedside staff; in other studies, aEEG was interpreted at a later date i.e. retrospectively
^fStudy used aEEG seizure detection algorithm
^gAuthors were contacted, and published and unpublished data reported
^hAuthors were contacted, but only published data reported as no unpublished data could be retrieved

Participant characteristics

All studies included newborn infants at risk of seizures, such as infants with HIE and other forms of encephalopathy, infants with clinical seizures, and infants receiving paralytic agents. However, a few studies had some additional inclusion criteria: one study included at‐risk infants only if they had abnormal or equivocal aEEG (Frenkel 2011), four studies included at‐risk infants only if they had cEEG‐confirmed seizures (Bourez‐Swart 2009; Shellhaas 2007a; Yamamoto 2015; Zhang 2011), one study included at‐risk as well as non‐risk infants (Eaton 1994), and one study had a case‐control design and included EEGs with and without seizures (Rennie 2004). Ten studies included term or late preterm infants. Four studies included moderate or very preterm, and two included extremely preterm infants and more mature newborns. The sample size of the infants included in the studies ranged between 10 and 121.

Index test (aEEG) characteristics

Eight studies used separate aEEG recording devices, while six studies derived aEEG from the raw trace of cEEG utilising the software. Two studies did not report the method of obtaining aEEG (Ponten 2010; Schmidt 2015).

Studies differed in the number of leads and lead positions used for the aEEG. Three studies used aEEG from two electrodes (P3‐P4 by two studies, and C3‐C4 by one study) (Frenkel 2011; Rennie 2004; Shellhaas 2007a), nine studies used four electrodes (C3‐C4, P3‐P4 by five studies, C3‐C4, T3‐T4 by two studies, and C3‐C4, O1‐O2 and C3‐C4, Fp1‐Fp2 by one study each), while one study used aEEG from nine electrodes but also reported analysis with two (C3‐C4) electrodes (Bourez‐Swart 2009). Two studies did not report the number of leads (Ponten 2010; Schmidt 2015).

Two studies reported using needle electrodes (Buttle Pros 2019; Mendelsohn 2019). Seven studies used raw traces to help aEEG interpretation. Two studies reported using an automated seizure detection algorithm to assist aEEG analysis (Lawrence 2009; Rakshasbhuvankar 2017).

aEEG interpretation

All but two studies reported only retrospective aEEG interpretation (Buttle Pros 2019; Mendelsohn 2019). The study by Buttle and colleagues reported both prospective interpretation by bedside clinical staff and retrospective interpretation by senior clinicians. However, we included only the results from the retrospective interpretation in the meta‐analyses, as most estimates from other studies in the meta‐analyses were also based on retrospective interpretations. Mendelsohn 2019 reported only prospective interpretation of aEEG by bedside clinical staff (see primary and secondary references).

The aEEG interpretation was performed retrospectively by neonatal consultants (10 studies), paediatric neurologists (four studies: Evans 2010; Lawrence 2009; Mastrangelo 2013; Yamamoto 2015), or an electroencephalographer (one study; Zhang 2011). In Buttle Pros 2019 and Mendelsohn 2019, prospective interpretation was performed by bedside residents or fellows in consultation with an attending neonatologist. Frenkel and colleagues reported interpretation by a medical student and neonatal fellow in addition to a neonatal consultant (Frenkel 2011).

Studies were variable with regard to the experience and training in interpreting aEEGs (Table).

Methodological quality of included studies

Our assessments of the quality of the included studies are shown in Figure and Figure.

2.

Risk of bias and applicability concerns summary: review authors' judgements about each domain for each included study

3.

Risk of bias and applicability concerns graph: review authors' judgements about each domain presented as percentages across included studies

Participant selection

Six studies had a high risk of bias in the domain of participant selection because of the inappropriate exclusions of eligible participants: four studies selected infants with cEEG‐confirmed electrographic seizures (Bourez‐Swart 2009; Shellhaas 2007a; Yamamoto 2015; Zhang 2011); one study selected infants with pathological or equivocal aEEG findings (Frenkel 2011); and the final study used a case‐control design (Rennie 2004). The bias in participant selection raises questions regarding applicability in the clinical setting for the participant selection domain. We judged the risk of bias for Mastrangelo 2013 to be unclear as the study excluded infants who had simultaneous recording of aEEG and cEEG for less than 12 hours without providing reasons for exclusion.

Index test

Two studies carried a high risk of bias as aEEG was interpreted with knowledge of cEEG findings (Eaton 1994; Lawrence 2009). We rated two studies as having an unclear risk in this domain (Buttle Retro 2019; Evans 2010). In Buttle Retro 2019, the aEEG interpreter was blinded to bedside aEEG interpretation, but the study did not report if the interpreter was blinded to cEEG findings. In Evans 2010, aEEG and cEEG were analysed independently, but by the same investigator. Hence, it was not clear whether blinding was successful.

All but two studies interpreted aEEG retrospectively (Buttle Pros 2019; Mendelsohn 2019), which is an applicability concern because, in real life, aEEG is analysed prospectively at the bedside. Thus, the effect of bedside analysis of aEEG by 'clinicians on the floor' (the real‐life situation) was not apparent in most studies. Buttle Pros 2019 reported prospective bedside interpretation by 'clinicians on the floor' and retrospective interpretation by an experienced neonatologist. Mendelsohn 2019 reported only prospective bedside interpretation by clinical staff.

Reference standard

Two studies had a high risk of bias as clinicians who interpreted cEEG were aware of the aEEG results (Eaton 1994; Mastrangelo 2013). We judged four studies to have an unclear risk of bias in this domain. In Evans 2010, aEEG and cEEG were analysed independently, but by the same investigator. Therefore, it was unclear if blinding was successful. Shah 2008 and Yamamoto 2015 did not mention if cEEG interpreters were unaware of aEEG findings. In Mendelsohn 2019, the bedside clinical team could opt to contact an on‐call neurologist after suspected seizure activity on aEEG to confirm or rule out seizures. It was unclear whether the knowledge of suspected seizure activity on aEEG could have influenced cEEG interpretation. No study had applicability concerns regarding the reference standard.

Flow and timing

In all studies, both the index test and reference standard were recorded simultaneously.

Findings

Ten studies (393 participants) reported on both 'neonate with seizures' and 'individual seizures' (Bourez‐Swart 2009; Buttle Pros 2019; Frenkel 2011; Mendelsohn 2019; Rakshasbhuvankar 2017; Schmidt 2015; Shah 2008; Shellhaas 2007a; Yamamoto 2015; Zhang 2011). Three studies (97 participants) reported only on the outcome of 'neonates with seizures' (Eaton 1994; Evans 2010; Rennie 2004), and three studies (72 participants) reported only on the outcome of 'individual seizures' (Lawrence 2009; Mastrangelo 2013; Ponten 2010).

aEEG for diagnosing 'neonates with seizures'

Thirteen studies (490 infants) reported data on the accuracy of aEEG for detecting 'neonates with seizures'. The meta‐analysis estimated that the aEEG had a sensitivity of 0.71 (95% CI 0.57 to 0.83), specificity of 0.84 (95% CI 0.59 to 0.95), positive LR of 4.50 (95% CI 1.55 to 13.04), and negative LR of 0.34 (95% CI 0.22 to 0.53) (Figure) for the detection of 'neonates with seizures'. The LR scatter plot (Figure) suggested that aEEG did not help confirm or exclude 'neonates with seizures'. The LR estimates suggest that aEEG is likely to alter post‐test probability only to a small extent. To assess the clinical implications of aEEG's estimated accuracy, we calculated post‐test probability values for a group of pre‐test probabilities using the Fagan nomogram (Table and Figure). Figure shows the summary receiver operating characteristics (SROC) plot with a 95% confidence interval region and 95% prediction region. The confidence interval region represents the area where the 'true' sensitivity and specificity will fall 95% of the time based on available data. The confidence area was confined to the left upper triangular part of the square, indicating the usefulness of the test based on the current data. If the confidence interval region occupies both the left upper and right lower triangular areas of the square, the test would not be clinically useful. The prediction region represents the area where the results of 95% of future studies will fall. The shape of the prediction region depends on the assumption of a bivariate normal distribution for the random effect and should not be over‐interpreted (Harbord 2009). The region is intended to visually represent the extent of between‐study heterogeneity, which is often considerable in diagnostic accuracy reviews (Harbord 2009).

4.

Forest plot for positive and negative likelihood ratios for the outcome of neonates with seizures

5.

Likelihood ratios scatter plot for the outcome of neonates with seizures

2. The post‐test probability values for a group of pre‐test probabilities using the Fegan nomogram.

Pre‐test probability of “true” seizure	Post‐test probability of “true” seizure if aEEG shows seizure	Post‐test probability of “true” seizure if aEEG does not show seizure
10 %	32	4
20 %	51	8
30 %	64	13
40 %	74	20

6.

Fegan nomogram for the outcome of neonates with seizures

7.

SROC plot (1 Detection of 'neonates with seizures')

Assessment of heterogeneity

There was significant heterogeneity, as indicated by the following:

1. widely differing sensitivity and specificity across the studies;

2. the 95% prediction region in the ROC space was much larger than the 95% confidence region.

Heterogeneity is often seen and expected in diagnostic accuracy studies as they often differ in the threshold used for positive tests, design, conduct, population, and reference standard (Lee 2015). Other sources of heterogeneity include bias, chance variation, and unexplained variation (Reitsma 2012).

We investigated heterogeneity in the following ways.

1. Visual inspection of the paired forest plots for sensitivity and specificity. This inspection did not show a pattern of negative correlation resulting from the trade‐off between sensitivity and specificity when studies vary in their threshold for the positive test. The threshold was not explicit, as in a numerical value to define a test as positive, but implicit variations in the interpretation and judgement. The forest plot was not suggestive of a 'threshold effect'.

2. Exploratory analyses of forest plots with studies sorted based on the presence or absence of covariates were unable to explain the heterogeneity. However, the number of studies in each group was small.

Thus, heterogeneity observed across the studies in the aEEG accuracy parameters remained unexplained. However, at the individual study level, studies reported factors influencing aEEG accuracy. Buttle Pros 2019 compared prospective bedside interpretation to retrospective interpretation by an expert. Bedside interpreters could not detect any (0/2) infants with cEEG seizures, whereas retrospective expert interpretation detected one out of two infants with cEEG seizures. Frenkel 2011 compared the two‐channel aEEG versus the two‐channel aEEG with the raw trace, and reported that sensitivity was significantly higher when the raw trace was used (70.9% versus 17.7%; P < 0.05). They also compared interpretations by a medical student, fellow, and neonatal consultant for detecting neonates with seizures using a single channel aEEG with raw trace. The sensitivities were similar for the student, fellow, and consultant (0.8, 0.9, 0.9, respectively), but specificity was lower when interpreted by students as compared to fellows and consultants (0.43, 0.86, 0.93, respectively).

Sensitivity analysis

We performed a sensitivity analysis by including studies with a low risk of bias (Buttle Pros 2019; Rakshasbhuvankar 2017; Schmidt 2015), and excluding: (a) studies that did not use raw EEG traces; (b) studies where the duration of monitoring was less than six hours; (c) studies in which participants were enroled based on knowledge of cEEG findings; (d) studies in which aEEG was derived from cEEG; and (e) studies in which aEEG was interpreted retrospectively (see Table).

3. Sensitivity analysis.

		No. of studies	Sensitivity (95% CIs)	Specificity (95% CIs)	Positive LR (95% CIs)	Negative LR (95% CIs)
All studies		13	0.71 (0.57, 0.83)	0.84 (0.59, 0.95)	4.50 (1.55, 13.04)	0.34 (0.22, 0.53)
Duration of recording	Studies where the duration of recording > 6 hr (Buttle Retro 2019; Evans 2010; Mendelsohn 2019; Rakshasbhuvankar 2017; Schmidt 2015; Shah 2008; Zhang 2011)	7	0.71 (0.60, 0.80)	0.78 (0.61, 0.89)	3.20 (1.7, 6.1)	0.37 (0.26, 0.54)
	Studies where the duration of recording was < 6 hr (Bourez‐Swart 2009; Frenkel 2011; Shellhaas 2007a; Yamamoto 2015)	4	0.77 (0.48, 0.92)	0.69 (0.00, 1.00)	2.5 (0.00, 1825)	0.34 (0.02, 4.8)
Risk of bias	Studies with low risk of bias (Buttle Pros 2019; Rakshasbhuvankar 2017; Schmidt 2015)	3	0.56 (0.02, 0.99)	0.78 (0.60, 0.90)	2.6 (0.40, 17.50)	0.56 (0.05, 6.08)
	Studies with high or uncertain risk of bias (Bourez‐Swart 2009; Buttle Retro 2019; Eaton 1994; Evans 2010; Frenkel 2011; Mendelsohn 2019; Rennie 2004; Shah 2008; Shellhaas 2007a; Yamamoto 2015; Zhang 2011	11	0.71 (0.56, 0.82)	0.83 (0.51, 0.96)	4.2 (1.20, 14.60)	0.35 (0.23, 0.55)
Use of raw trace	Studies which used raw trace (Buttle Retro 2019; Frenkel 2011; Mendelsohn 2019; Rakshasbhuvankar 2017; Shah 2008; Zhang 2011)	6	0.75 (0.64, 0.84)	0.85 (0.74, 0.92)	5.0 (2.70, 9.30)	0.29 (0.19, 0.45)
	Studies which didnot use raw trace (Bourez‐Swart 2009; Eaton 1994; Evans 2010; Rennie 2004; Shellhaas 2007a; Yamamoto 2015)	6	0.68 (0.45, 0.84)	0.73 (0.20, 0.99)	2.5 (0.20, 34.00)	0.44 (0.18, 1.04)
Availability of cEEG findings at enrolment	Studies in which cEEG findings were not available at enrolment (Buttle Retro 2019; Eaton 1994; Evans 2010; Frenkel 2011; Mendelsohn 2019; Rakshasbhuvankar 2017; Schmidt 2015; Shah 2008)	8	0.79 (0.59, 0.90)	0.83 (0.66, 0.92)	4.6 (2.10, 10.00)	0.26 (0.12, 0.55)
aEEG recording method	Studies in which aEEG was recorded using an independent aEEG machine (Buttle Retro 2019; Eaton 1994; Frenkel 2011; Mendelsohn 2019; Rakshasbhuvankar 2017; Shah 2008)	6	0.84 (0.66, 0.94)	0.86 (0.76, 0.93)	6.2 (3.50, 10.90)	0.18 (0.08, 0.43)
	Studies in which aEEG was derived from cEEG (Bourez‐Swart 2009; Evans 2010; Rennie 2004; Shellhaas 2007a; Yamamoto 2015; Zhang 2011)	6	0.69 (0.50, 0.83)	0.51 (0.06, 0.94)	1.4 (0.40, 4.90)	0.61 (0.21, 1.76)
Studies in which aEEG was interpreted prospectively at bedside	Mendelsohn 2019	1	0.57 (0.18, 0.90)	0.91 (0.78, 0.98)	7.05 (2.00, 24.90)	0.46 (0.19, 1.10)
	Buttle Pros 2019	1	0 (0.00, 0.84)	0.81 (0.48, 0.97)	0	1.2 (0.92, 1.61)

CI: confidence interval; LR: likelihood ratio

Assessment of publication bias

A P value of 0.99 in the Deeks' funnel plot asymmetry test indicated that there was no significant publication bias (Figure).

8.

Deeks' funnel plot for the outcome of neonates with seizures

aEEG for detecting 'individual seizures'

Thirteen studies reported the outcome of the detection of 'individual seizures'. The reported sensitivity of aEEG for detecting 'individual seizures' ranged from 0% to 86%. However, all 13 studies calculated sensitivity by dividing the total number of true seizures detected using aEEG (true positive) in all recruited infants by the total number of true seizures (true positive+false negative), without taking into account the clustering of seizures in infants – that is, the relatedness of the multiple seizures (non‐independent nature) occurring in an infant, leading to unit of analysis issues (Parsons 2018).

The average number of seizures reported per recruited infant had a wide range: 0.6 to 44.3 on cEEG, and 0.6 to 17.7 on aEEG. The seizure detection rate in an individual infant ranged from 0 to 100%; that is, aEEG could not detect any seizure in some infants whereas it detected all seizures in some infants (Bourez‐Swart 2009; Buttle Pros 2019; Rakshasbhuvankar 2017; Shah 2008). The reported reasons for the inability of aEEG to detect seizures were the short duration of seizures (seven studies), localisation of seizures away from aEEG leads (five studies), low voltage (five studies), and inexperience of interpreters in interpreting aEEG (Frenkel 2011; Shellhaas 2007a).

It was not possible to calculate the specificity of aEEG for detecting 'individual seizures' because of the difficulty in estimating true negatives. Therefore, we reported false‐positive seizures. Two studies did not report false‐positive seizures (Bourez‐Swart 2009; Zhang 2011). The proportion of false‐positive seizures amongst all seizures detected on aEEG varied across studies and ranged from 0% (Shellhaas 2007a) to 100% (Buttle Pros 2019). Shah 2008 reported one false‐positive seizure per 39 hours of recording, whereas Mastrangelo 2013 reported one false‐positive seizure per 15 hours of aEEG recording. Frenkel 2011 reported a low false‐positivity rate when aEEG was interpreted by a neonatologist (11%), and a high false‐positivity rate (87%) when interpreted by students. The reported causes of false positivity were artefacts resulting from muscle movement, patting, hiccough, and electrode movement (Lawrence 2009; Mastrangelo 2013; Rakshasbhuvankar 2017; Shah 2008).

Bourez‐Swart 2009 reported sensitivity for different lead positions. Sensitivity to detect individual seizures was highest for C3‐C4 (30%), followed by T3‐T4 (20%), O2‐O1 (17%), Cz (13%), and Fp2‐Fp1 (10%). Frenkel 2011 reported that the interpretation of aEEG by a student with basic training but no experience of aEEG interpretation resulted in the overdiagnosis of seizures. Two studies compared seizure detection using only aEEG and aEEG plus raw EEG trace (Shah 2008; Zhang 2011). Sensitivity to detect individual seizures was better when aEEG was used along with the raw EEG trace. Lawrence 2009 reported the sensitivity of the seizure detection algorithm. The algorithm detected 171/426 (40%) cEEG seizures. Only two studies reported prospective aEEG interpretation (Buttle Pros 2019; Mendelsohn 2019). Of these, Buttle Pros 2019 reported that none of the eight cEEG‐confirmed seizures were detected on aEEG. On the other hand, Mendelsohn 2019 reported that six of 13 cEEG‐confirmed seizures were detected using aEEG with prospective interpretation.

Summary of evidence

The evidence generated by the systematic review is summarised in Table and Table (Rakshasbhuvankar 2021b). Overall, the certainty of the evidence was low due to some risk of bias in included studies, high heterogeneity, and wide confidence intervals around effect size estimates. We did not downgrade the evidence certainty for risk of bias for the outcome of the detection of 'individual seizures' as most risk of bias was in the 'participant selection' domain, which we judged to be less likely to affect the outcome of the detection of 'individual seizures'.

Discussion

This systematic review summarises current evidence for the accuracy of aEEG for detecting neonatal seizures and highlights the limitations of the existing evidence. Knowing the limitations and advantages of aEEG for detecting seizures is essential if we want to improve its use in clinical practice and determine the future direction of research.

Summary of main results

We included 16 studies (562 neonates) in our meta‐analysis comparing aEEG with cEEG for the detection of neonatal seizures. Of these, 10 studies were conducted in term and late preterm infants. Half of the included studies did not use raw EEG traces. Fourteen studies reported outcomes based only on retrospective interpretation. Ten of 16 studies used four electrodes (making this the most common approach amongst the included studies), and 10 studies' aEEG recordings exceeded six hours. Only two included studies used a seizure detection algorithm. In 14 studies, neonatal or neurology consultants performed aEEG interpretation, and most (in 10 of the 14 studies) had prior experience in aEEG interpretation.

We found that aEEG had a sensitivity of 0.71 (95% CI 0.57 to 0.83), specificity of 0.84 (95% CI 0.59 to 0.95), positive LR of 4.50 (95% CI 1.55 to 13.04), and negative LR of 0.34 (95% CI 0.22 to 0.53) for the detection of 'neonates with seizures', but the certainty of evidence was low due to high and unexplained heterogeneity (Table). The only two prospective studies (53 participants) which interpreted aEEGs 'live' at the bedside reported sensitivities of zero and 0.57 and specificities of 0.82 and 0.92.

aEEG can detect a variable number of 'individual seizures' in an infant, ranging from zero to 100% (Table). The common causes of missed seizures on aEEG were short duration of seizures, localisation of seizures away from aEEG leads, low voltage, and inexperience in interpreting them. The rate of false‐positive seizures was variable but could be high, especially for bedside and inexperienced interpreters. Artefacts resulting from muscle movement, patting, hiccups, and electrodes were the reported causes of false‐positive results.

Strengths and weaknesses of the review

Strengths of our review include an extensive literature search, and meta‐analysis of the outcome of detection of 'neonates with seizures'.

Considerable unexplained heterogeneity was a weakness of our review. Heterogeneity is expected in diagnostic accuracy studies as they often differ in the threshold used for positive tests, design, conduct, population, and reference standard (Lee 2015). Other sources of heterogeneity include bias, chance variation, and unexplained variation (Reitsma 2012). The threshold effect, use of the raw EEG trace (Yes/No), duration of recording, and experience and training of the interpreter could not explain the observed heterogeneity; however, the number of studies and sample sizes were small. Therefore, further research is required to assess the effect of the possible sources of heterogeneity on the accuracy of the aEEG for detecting 'infants with seizures'.

Another possible reason for heterogeneity is patient‐related factors such as gestational age (e.g. preterm versus term) and the cause of seizures (e.g. local pathologies such as cerebral infarcts versus generalised pathologies such as HIE and metabolic insults). The fact that the rate of seizure detection by aEEG varied between different infants within a single study suggests the role played by patient‐related factors. We could not perform subgroup analysis based on the diagnosis (generalised versus focal pathology) and gestational age because of the unavailability of relevant data. Considering that the number of variables may affect the accuracy of aEEG, individual participant data meta‐analysis may be more beneficial to tease out the complex interaction between multiple covariates and their effect on aEEG accuracy for diagnosing seizures (Rakshasbhuvankar 2021a). Individual participant data systematic review and meta‐analysis require considerable resources and collaboration between the study investigators, which is often difficult to achieve.

Another weakness was that we could not perform a meta‐analysis for the outcome of 'individual seizures' due to the non‐independent nature of seizures within an infant. The studies that reported accuracy parameters for the diagnosis of 'individual seizures' calculated the parameters without considering the effect due to clustering. Multiple observations made within an infant are positively correlated and, hence, can result in biased estimates of sensitivity and specificity, if the correlated nature of the data is ignored (Genders 2012). Seizures within an infant are likely to have a similar topographic distribution (e.g. focal versus generalised, temporal versus occipital), electrophysiological characteristics (e.g. low amplitude seizures, duration, frequency), occur on a common background pattern, and are often interpreted by the same interpreters, all of which can influence the diagnostic ability of aEEG for the detection of 'individual seizures'. Therefore, calculating accuracy parameters requires adjustment, such as logistic random‐effects regression, to account for the possible correlation between observations (Genders 2012).

Applicability of findings to the review question

The review findings need to be interpreted, taking into consideration the low certainty of evidence and applicability concerns.

Whilst cEEG is the reference standard for monitoring and diagnosing neonatal seizures, it is often not available round‐the‐clock in many NICUs (Boylan 2010; Dilena 2019). In the absence of cEEG, clinical observation is routinely used to diagnose seizures in day‐to‐day practice. However, clinical observation is unreliable for the diagnosis of neonatal seizures (Clancy 1988; Mizrahi 1989). Up to 90% of neonatal seizures may not have clinical manifestations (electrical seizures) (Lawrence 2009). On the other hand, clinical observation can misdiagnose benign neonatal events as seizures. Buttle and colleagues reported that only one episode out of 14 seizure‐like events detected by clinical observation was an actual seizure based on cEEG (Buttle Pros 2019; Buttle Retro 2019). Our review showed that aEEG has a sensitivity of 71% to detect 'neonates with seizures'. Although not very high, the sensitivity may be better than a seizure diagnosis by clinical observation alone. This 'add‐on' sensitivity of aEEG may be helpful for clinical management.

aEEG's sensitivity to detect 'individual seizures' varied widely between studies (0% to 86%) and between infants (0% to 100%). This finding indicates that aEEG may be helpful in detecting seizures in some infants, but not in others. In addition, the proportion of false‐positive seizures could be as high as 100%. Therefore, if aEEG is used in treatment decisions, there is a risk of undertreatment and overtreatment with anti‐seizure medications. Whether aEEG is better or worse than clinical monitoring for the detection of seizures needs further investigation. Based on the available evidence, aEEG cannot be recommended as a mainstay investigation for managing neonatal seizures. A recent RCT that evaluated the effect of treatment of clinical seizures versus electrographic seizures (based on aEEG) found no difference in the risk of death or severe disability between the two groups (Hunt 2021). The investigators acknowledged that the use of aEEG instead of cEEG was an important limitation of their study and that cEEG remains the investigation of choice for the detection of neonatal seizures.

Our review highlights one important applicability concern of current data: namely, retrospective aEEG analysis. In real life, aEEG is interpreted prospectively by bedside clinicians and not retrospectively, which is often by aEEG experts. Variations in the degree of experience and training of bedside clinicians, availability of complete clinical data, including examination findings, and work environment, including the opinion of peers, may affect aEEG interpretation by bedside clinicians. Only two studies reported prospective bedside interpretation by 'clinicians on the floor' (Buttle Pros 2019; Mendelsohn 2019). However, the study sample sizes were small (13 and 40, respectively).

In seven studies, aEEG was derived from cEEG. Clinical studies have not investigated whether aEEG derived from cEEG has a differential effect on aEEG interpretation compared with aEEG recorded with a specialised bedside machine. The practice of deriving the index test from reference standards could make a study prone to incorporation bias. Hence, we performed a sensitivity analysis by excluding the studies where aEEG was derived from cEEG. In addition, as this practice has applicability concerns, we labelled such studies as of high concern for the index test domain of the QUADAS‐2 assessment tool.

Despite its limitations, aEEG is used for making treatment decisions with anti‐seizure medications (Mendelsohn 2019 (see secondary reference); Rakshasbhuvankar 2022; Shah 2012), because of the unavailability of round‐the‐clock cEEG monitoring (Boylan 2010). Our systematic review cautions clinicians about using aEEG for diagnosing individual seizures and the risk of under‐ or overtreatment with anti‐seizure medications. cEEG, although ideal, is labour‐intensive. The inability to initiate cEEG and interpret findings are significant obstacles in using cEEG for clinical purposes. Technological advances in recent years, such as the development of lead placement templates to enable trained neonatal nursing staff to place cEEG leads (e.g. Baby BraiNet® neonatal template, Jordan NeuroScience Inc., USA), display of cEEG with multiple channel aEEG, automated neonatal seizure detection algorithm with direct alerts to neurologists, ability to review cEEG remotely, and commercial cEEG monitoring services have enabled the establishment of cEEG services in many NICUs (Goswami 2018; Kolls 2012; Sharpe 2019). In addition, awareness amongst neonatologists regarding the need for cEEG monitoring, provision of resources, staff education, and ongoing support is essential in establishing and running cEEG monitoring services for clinical purposes. Once established, such services will also be helpful for further clinical research on neonatal seizures and anti‐seizure medications.

In addition to seizure detection, aEEG is used to assess electroencephalographic background in infants at risk of brain dysfunction. aEEG background is helpful in the diagnosis, management, and prognostication of infants with HIE. Abnormal aEEG background pattern was included as one of the criteria to determine the eligibility of therapeutic hypothermia in clinical trials (Azzopardi 2009; Gluckman 2005). A meta‐analysis of nine studies showed that a persistently abnormal aEEG background at 48 hours or more is associated with adverse neurodevelopmental outcomes in infants who were treated with therapeutic hypothermia for HIE (Chandrasekaran 2017). aEEG use for the identification of infants for therapeutic hypothermia has been recommended.

Whether to treat electrographic seizures without clinical manifestations is an unsolved conundrum. Although the aetiology of neonatal seizures plays a vital role in the long‐term outcomes, seizures themselves may contribute to adverse neurodevelopmental outcomes. Whether treatment of electrographic seizures will improve neurodevelopmental outcomes is yet to be proven. Such trials cannot be conducted without an accurate diagnosis of neonatal seizures, which is currently only possible using the cEEG.

Authors' conclusions

Implications for practice.

Low‐certainty evidence suggests that amplitude‐integrated electroencephalography (aEEG) has moderate sensitivity and specificity for detecting 'neonates with seizures'; its ability to detect 'individual seizures' varies widely. The proportion of seizures that are detected and which remain undetected is likely to vary between infants, and may also depend on aEEG methodology (e.g. number of leads used, availability of the raw trace, seizure‐detection software) and aEEG interpreter (e.g. experience and training of aEEG interpretation, and bedside versus retrospective interpretation).

Our review suggests that aEEG can under‐diagnose or over‐diagnose individual seizures, leading to under‐ or over‐treatment with anti‐seizure medications. Therefore, when conventional video‐electroencephalography (cEEG) monitoring is available, it remains the monitoring of choice for detecting neonatal seizures in at‐risk infants. When cEEG monitoring is available, but resources are limited, aEEG may be helpful in detecting neonates who would benefit most from cEEG monitoring. When cEEG monitoring is not available, aEEG may be helpful in detecting neonates who need close clinical observation for the detection of seizures and may also help detect individual seizure episodes. Studies are required to investigate the utility of aEEG for seizure detection when cEEG monitoring is unavailable and neonates are monitored with clinical observation alone.

Implications for research.

Future studies should be designed to meet the QUADAS‐2 (Quality Assessment of Diagnostic Accuracy Studies) criteria so that they will be of high methodological quality with minimal bias. Studies should clearly state the method of enrolment (prospective or retrospective), characteristics of the population assessed, blinding of reference standard and index tests, and explanation of withdrawals. Studies should explicitly state the details of the clinical setting and infant characteristics so that clinicians can determine the generalisability of the diagnostic test for their infant population.

More studies are required to investigate aEEG accuracy when interpreted prospectively by bedside clinicians.

Future studies should report the adjusted sensitivity of aEEG for detecting individual seizures to help interpret and pool the data.

As seizures away from aEEG leads are likely to be missed, studies evaluating the use of more than four leads to cover occipital, temporal, and frontal areas are needed. The incorporation of simultaneous video recording to aEEG in such studies would be helpful in differentiating movement artefacts from true seizures. When cEEG monitoring is not available, neonates at risk of seizures are monitored with clinical observation for seizures. Studies are required to evaluate if the combination of aEEG and clinical observation is better than clinical observation alone.

History

Protocol first published: Issue 3, 2020

Appendices

Appendix 1. Cochrane Central Register of Controlled Trials (CENTRAL) via CRS Web search strategy

1. amplitud* ADJ1 integrat* ADJ3 (EEG* OR electroencephalo* OR electro‐encephalo*) AND CENTRAL:TARGET

2. aEEG* OR "a‐EEG*" AND CENTRAL:TARGET

3. (brain OR cereb*) ADJ1 "function NEXT monitor*" AND CENTRAL:TARGET

4. channel ADJ3 (EEG* OR electroencephalo* OR "electro‐encephalo*") AND CENTRAL:TARGET

5. #1 OR #2 OR #3 OR #4 AND CENTRAL:TARGET

6. MESH DESCRIPTOR Infant, Newborn EXPLODE ALL AND CENTRAL:TARGET

7. MESH DESCRIPTOR Infant, Newborn, Diseases EXPLODE ALL AND CENTRAL:TARGET

8. MESH DESCRIPTOR Intensive Care Units, Neonatal EXPLODE ALL AND CENTRAL:TARGET

9. newborn* OR "new‐born*" OR "new NEXT born*" OR "newly NEXT born*" OR "newly‐born*" OR neonat* OR "neo‐nat*" OR perinat* OR "peri‐nat*" OR birth* OR baby OR babies OR NICU* OR neuroNICU* AND CENTRAL:TARGET

10. #6 OR #7 OR #8 OR #9

11. MESH DESCRIPTOR Epilepsy EXPLODE ALL AND CENTRAL:TARGET

12. MESH DESCRIPTOR Seizures EXPLODE ALL AND CENTRAL:TARGET

13. (epilep* OR seizure* OR convuls*):AB,KW,KY,MC,MH,TI AND CENTRAL:TARGET

14. #11 OR #12 OR #13 AND CENTRAL:TARGET

15. #5 AND #10 AND #14

16. amplitud* ADJ1 integrat* ADJ3 (EEG* OR electroencephalo* OR electro‐encephalo*) AND INSEGMENT

17. aEEG* OR "a‐EEG*" AND INSEGMENT

18. (brain OR cereb*) ADJ1 "function NEXT monitor*" AND INSEGMENT

19. channel ADJ3 (EEG* OR electroencephalo* OR "electro‐encephalo*") AND INSEGMENT

20. #16 OR #17 OR #18 OR #19 AND INSEGMENT

21. MESH DESCRIPTOR Infant, Newborn EXPLODE ALL AND INSEGMENT

22. MESH DESCRIPTOR Infant, Newborn, Diseases EXPLODE ALL AND INSEGMENT

23. MESH DESCRIPTOR Intensive Care Units, Neonatal EXPLODE ALL AND INSEGMENT

24. newborn* OR "new‐born*" OR "new NEXT born*" OR "newly NEXT born*" OR "newly‐born*" OR neonat* OR "neo‐nat*" OR perinat* OR "peri‐nat*" OR birth* OR baby OR babies OR NICU* OR neuroNICU* AND INSEGMENT

25. #21 OR #22 OR #23 OR #24 AND INSEGMENT

26. #20 AND #25

27. #15 OR #26

Appendix 2. MEDLINE search strategy

1. (amplitud* adj1 integrat* adj3 (EEG* or electroencephalo* or electro‐encephalo*)).tw,kf.

2. (aEEG? or a‐EEG?).tw,kf.

3. ((brain or cereb*) adj1 function monitor*).tw,kf.

4. (channel adj3 (EEG* or electroencephalo* or electro‐encephalo*)).tw,kf.

5. or/1‐4

6. exp Infant, Newborn/ or exp Infant, Newborn, Diseases/ or exp Intensive Care Units, Neonatal/

7. (newborn* or new*born* or new*‐born* or neonat* or neo‐nat* or perinat* or peri‐nat* or birth* or baby or babies or NICU* or neuroNICU*).tw,kf.

8. 6 or 7

9. exp Epilepsy/

10. exp Seizures/

11. (epilep$ or seizure$ or convuls$).tw.

12. 9 or 10 or 11

13. 5 and 8 and 12

14. exp animals/ not humans.sh.

15. 13 not 14

16. remove duplicates from 15

Appendix 3. ClinicalTrials.gov search strategy

aEEG OR "a‐EEG" OR (Amplitude integrated (EEG OR electroencephalography OR encephalograph)) | Seizure Newborn

Appendix 4. ICTRP search strategy

(neonatal OR newborn) AND seizure AND (Amplitude integrated electroencephalography OR aEEG OR "a‐EEG")

Appendix 5. Embase search strategy

1. exp amplitude‐integrated EEG/

2. (amplitud* adj1 integrat* adj3 (EEG* or electroencephalo* or electro‐encephalo*)).mp.

3. (aEEG? or a‐EEG?).mp.

4. ((brain or cereb*) adj1 function monitor*).mp.

5. (channel adj3 (EEG* or electroencephalo* or electro‐encephalo*)).mp.

6. 1 or 2 or 3 or 4 or 5

7. exp newborn/ or exp baby/ or exp neonatal care/ or exp neonatal intensive care/ or exp neonatal intensive care unit/

8. (newborn* or new*born* or new*‐born* or neonat* or neo‐nat* or perinat* or peri‐nat* or birth* or baby or babies or NICU* or neuroNICU*).mp.

9. 7 or 8

10. exp seizure/ or exp epilepsy/

11. (epilep* or seizure* or convuls*).mp.

12. 10 or 11

13. 6 and 9 and 12

14. remove duplicates from 13

Appendix 6. Data extraction form

A. Study ID

Authors:

Journal:

Year of publication:

Country/countries where study was done:

Funding:

Ethical approval:

B. Enrolment

Study design:

Study setting:

Simultaneous aEEG and cEEG recordings analyzed: Yes/No (only studies with Yes will be eligible)

Sample size calculation:

Inclusion criteria:

Exclusion criteria:

Recruitment method: consecutive/random/other:

Recruitment period

C. Index test details

aEEG device:

Number of electrodes:

Lead positions:

Use of raw trace to assist interpretation: Yes/No

Use of automated seizure detection algorithm to assist interpretation: Yes/No

If Yes, what algorithm was used:

Interpreter’s qualification:

Interpreter’s aEEG training: Yes/No. If yes, training details:

Interpreter’s experience in aEEG

Were the interpreters “blinded” to the results of cEEG: Yes/No/Not clear

Duration of the aEEG interpreted per patient/sample:

D. Reference standard details

Number of leads:

Interpreter’s qualification:

Were the interpreters “blinded” to the results of aEEG: Yes/No/Not clear

Did all participants receive a reference standard? Yes/No

Did all participants receive the same reference standard: Yes/No

E. Results

Eligible participants, n =

Exclusions, n = …, reasons

Participants receiving index test and reference standard, n =

Participants included in the 2x2 tables, n =

Participants excluded from 2x2 tables, n = …, reasons

Participants’ characteristics:

· Age (mean, SD, range)

· Gestational age at birth (median , range)

· Gestational age at monitoring (median, range)

· Sex (% female)

· Percentage of study infants with a diagnosis of HIE:

· Percentage of infants treated with therapeutic hypothermia:

· Percentage of infants treated with anticonvulsants prior to the study:

· Any other characteristics that might influence test performance or applicability of results (e.g. Cause of seizure/diagnosis, treatment with therapeutic hypothermia, underlying condition, comorbidities etc.)

Were all patients /samples included in analysis?: Yes / No

a) Detection of patients with seizure

True positives (seizures on cEEG and aEEG):

True negatives (no seizures on cEEG and aEEG):

False positives (no seizure on cEEG but seizure on aEEG):

False negatives (seizure on cEEG but NO seizure on aEEG):

Sensitivity:

Specificity:

Interrater agreement (when multiple interpreters):

Characteristics of 'missed participants' (false negative):

b) Detection of individual seizure episode

True positives (seizure on cEEG and aEEG):

False positive (No seizure on cEEG but seizure on aEEG):

False negatives (Seizure on cEEG but NO seizure on aEEG):

Sensitivity:

Interrater agreement (when multiple interpreters):

cEEG characteristics of 'missed seizures' (false negative):

c) Detection of seizure burden

Duration of seizures on cEEG (total positive):

Duration of seizures detected correctly with aEEG (true positive):

Duration of seizures “missed” by aEEG (false negative):

Sensitivity for detection of seizure burden:

Appendix 7. QUADAS 2 checklist

Domain	Signalling question	Criterion
1. Participant selection
1.1. Risk of bias	Was a consecutive or random sample of patients enroled?	Yes if consecutive or random sampling is stated in the paper No if convenient sampling is used instead Unclear if the sampling procedure is not clearly described
	Did the study avoid inappropriate exclusions?	Yes if all infants that meet our definition of target population have been included No if the following groups have been excluded: exclusion of the infants based on cEEG or aEEG findings Unclear if the inclusion and exclusion criteria and methods of selection are not sufficiently clearly described to make a judgement
1.2. Concerns regarding applicability	Is there concern that the included participants do not match the review question?	Target population: infants with clinically suspected seizures or infants at risk of seizures High concern if sample is different from target population Low concern if sample meets the target population criteria Unclear if not sufficient detail reported to make a decision
2. Index test
2.1. Risk of bias	Were the index test results interpreted without knowledge of the results of the reference standard?	Yes if clear from the paper No if clear that there was no blinding Unclear if information is not sufficient to make a decision
	If threshold was used, was it prespecified?	Not relevant, as diagnosis of seizure is not a continuous outcome
2.2. Concerns regarding applicability	Is there concern that the index test, its conduct or interpretation differ from the review question?	Index test: aEEG with or without raw trace/automated seizure detection algorithm, interpreted by neonatal clinicians in real‐time, used for detection of electrical seizures in neonatal infants High concern if index test and/or its performance different from review question e.g. aEEG derived from cEEG; retrospective interpretation of aEEG Low concern if index test and/or its performance relevant to review question Unclear if not sufficient detail reported to make a decision
3. Reference standard
3.1. Risk of bias	Is the reference standard likely to correctly classify the target condition?	Optimal reference standard: cEEG recorded using at least following 9 electrodes (Fp1, Fp2, C3, C4, Cz, T3 (or T7), T4 (or T8), O1, and O2) and interpreted by a specialist (e.g. neurologist, clinical neurophysiologist) trained or experienced in neonatal cEEG interpretation Yes if the reference standard meet the above criteria No if it does not Unclear if information insufficient to make a decision
	Were the reference standard results interpreted without knowledge of the results of the index test?	Yes if clear from the paper No if clear that there was no blinding Unclear if information is not sufficient to make a decision
3.2. Concerns regarding applicability	Is there concern that the target condition as defined by the reference standard does not match the review question?	Target condition: electrographic seizures of at least 10 seconds' duration High concern if target condition different from review question Low concern if target condition relevant to review question Unclear if not sufficient detail reported to make a decision
4. Flow and timing
4.1. Risk of bias	Was there an appropriate interval between index test(s) and reference standard? Did all participants receive a reference standard? Did participants receive the same reference standard? Were all participants included in the analysis?	Appropriate time interval between index test and reference standard: cEEG and aEEG must be simultaneously recorded Yes if all participants who were enroled in the study were also included in the analysis

Data

Presented below are all the data for all of the tests entered into the review.

Tests. Data tables by test.

Test	No. of studies	No. of participants
1 aEEG for the detection of 'neonates with seizures'	13	502
2 aEEG for the detection of 'individual seizures' in neonates	14	4458

1. Test.

aEEG for the detection of 'neonates with seizures'

2. Test.

aEEG for the detection of 'individual seizures' in neonates

Characteristics of studies

Characteristics of included studies [ordered by study ID]

Bourez‐Swart 2009.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG interpretation
Patient characteristics and setting	Term infants with suspected seizures associated with HIE were monitored with continuous cEEG and aEEG. Infants were included in the study only if they had cEEG seizures.
Index tests	aEEG was derived from cEEG signals. It was assessed by two experienced neonatologists.
Target condition and reference standard(s)	cEEG interpreted by two experienced observers (board‐certified clinical neurophysiologist and board‐certified technician)
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by two board‐certified clinical neurophysiologists and one board‐certified technician
Notes	The funding source was not reported.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	No
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Buttle Pros 2019.

Study characteristics
Patient Sampling	Prospective enrolment, prospective aEEG interpretation
Patient characteristics and setting	Neonates admitted to NICU who required EEG monitoring for suspected seizure or therapeutic hypothermia.
Index tests	Four lead aEEG with raw traces interpreted prospectively by bed‐side clinical staff (resident or fellow in consultation with an attending neonatologist).
Target condition and reference standard(s)	Conventional nine lead EEG interpreted by staff neurologist with expertise in reading neonatal EEGs.
Flow and timing	Simultaneous recoding of aEEG and cEEG.
Comparative	cEEG interpreted by staff neurologist with expertise in reading neonatal EEGs
Notes	SGB (first author) received a Faculty of Medicine Student Research Award and Undergraduate Research Opportunity Program funding from the University of Ottawa to assist in completing the project.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			Low concern
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Unclear
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Buttle Retro 2019.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG interpretation
Patient characteristics and setting	Neonates admitted to NICU who required EEG monitoring for suspected seizure or therapeutic hypothermia.
Index tests	Four lead aEEG with raw traces interpreted retrospectively by a neonatologist with long‐standing experience in aEEG interpretation.
Target condition and reference standard(s)	Conventional nine lead EEG interpreted by staff neurologist with expertise in reading neonatal EEGs.
Flow and timing	Simultaneous recoding of aEEG and cEEG.
Comparative	cEEG interpreted by staff neurologist with expertise in reading neonatal EEGs
Notes	SGB (first author) received a Faculty of Medicine Student Research Award and Undergraduate Research Opportunity Program funding from the University of Ottawa to assist in completing the project.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Unclear
Could the conduct or interpretation of the index test have introduced bias?		Unclear risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Unclear
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Eaton 1994.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG interpretation
Patient characteristics and setting	Neonates 28‐40 wk GA without and with risk factors for seizures included. In practice, preterm infants without risk factors are not monitored for seizures with aEEG or cEEG.
Index tests	Two leads (C3, C4) aEEG without raw trace and interpreted by an investigator who also interpreted cEEG
Target condition and reference standard(s)	Thirteen lead cEEG interpreted by an investigator who also interpreted aEEG.
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by EEG specialist
Notes	The funding source was not reported.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	No
Could the conduct or interpretation of the index test have introduced bias?		High risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	No
Could the reference standard, its conduct, or its interpretation have introduced bias?		High risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Evans 2010.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective interpretation
Patient characteristics and setting	Infants requiring continuous EEG monitoring for clinical purposes.
Index tests	Four leads (C3, C4, O1, O2) aEEG without raw trace and interpreted retrospectively by paediatric neurologist with special training in neonatal EEG and certified in clinical neurophysiology
Target condition and reference standard(s)	Eleven lead cEEG interpreted by paediatric neurologist with special training in neonatal EEG and certified in clinical neurophysiology
Flow and timing	Simultaneous recording
Comparative	cEEG interpred by a paediatric neurologist with special training in neonatal EEG and certified in clinical neurophysiology.
Notes	EE (first author) was supported by NICHD T32 training grant no. 07549
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Unclear
Could the conduct or interpretation of the index test have introduced bias?		Unclear risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Unclear
Could the reference standard, its conduct, or its interpretation have introduced bias?		Unclear risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Frenkel 2011.

Study characteristics
Patient Sampling	Retrospective analysis of recorded aEEG/cEEG
Patient characteristics and setting	High‐risk infants with pathological or equivocal aEEG recording were included.
Index tests	Two lead aEEG with raw traces and interpreted retrospectively by a medical student, neonatal fellow, and senior neonatologist.
Target condition and reference standard(s)	cEEG interpreted by two paediatric neurologists and a senior neonatologist.
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by two paediatric neurologists and one senior neonatologist
Notes	The study was funded by the National Institute of Psychobiology of Israel.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	No
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Lawrence 2009.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG interpretation
Patient characteristics and setting	Neonatal encephalopathy, including suspected or confirmed clinical seizure events, pH ≤ 7 or base excel ≥ 12 by cord gas or within 1 hour of life, Apgar ≤ 5 at 5 min of life, severe clinical course requiring muscle relaxation for > 4 hours.
Index tests	Four lead aEEG with raw trace and seizure detection software and interpreted by consensus between “blinded” epileptologist and “unblinded” experienced aEEG reader
Target condition and reference standard(s)	Seventeen lead cEEG interpreted by epileptologist
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by epileptologist
Notes	Cure Kids Foundation provided financial support. BrainZ Instruments, Ltd, New Zealand, provided the monitors, but it had no role in the study or its conduct.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	No
Could the conduct or interpretation of the index test have introduced bias?		High risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Mastrangelo 2013.

Study characteristics
Patient Sampling	Prospective. Infants with < 12 hrs of cEEG/aEEG simultaneous recording were excluded. Retrospective aEEG interpretation.
Patient characteristics and setting	Infants undergoing therapeutic hypothermia or with suspected clinical or non‐convulsive seizures with simultaneous aEEG and cEEG recording for more than 12 hours were included.
Index tests	Four leads (C3, T3, C4, T4) aEEG without raw trace interpreted by neonatologist.
Target condition and reference standard(s)	cEEG interpreted by a senior paediatric neurologist who had access to the aEEG
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by senior paediatric neurologist
Notes	The funding source was not reported.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		Unclear risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	No
Could the reference standard, its conduct, or its interpretation have introduced bias?		High risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Mendelsohn 2019.

Study characteristics
Patient Sampling	Prospective recruitment and interpretation of aEEG
Patient characteristics and setting	Neonates with suspected seizures
Index tests	Four lead aEEG with raw trace interpreted by bedside clinicians
Target condition and reference standard(s)	cEEG interpreted by on‐call neurologist (not blinded)
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by neurologist
Notes	No financial support was used for the study. We contacted the study's authors and received additional information regarding patients with seizures, individual seizures, designation, experience and aEEG interpretation training of aEEG interpreters, number of aEEG leads and use of raw trace.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			Low concern
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	No
Could the reference standard, its conduct, or its interpretation have introduced bias?		Unclear risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Ponten 2010.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective (offline) analysis
Patient characteristics and setting	Term neonates with hypoxic‐ischaemic encephalopathy
Index tests	aEEG without raw trace interpreted independently by an expert
Target condition and reference standard(s)	cEEG interpreted by "blinded" clinical neurophysiologist
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by clinical neurophysiologist
Notes	The funding source was not reported. We contacted the study authors; however, no additional information could be retrieved.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Unclear
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Rakshasbhuvankar 2017.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG analysis
Patient characteristics and setting	Infants admitted in NICU with GA of ≥ 35 weeks and ≤ 28 days of life with a diagnosis of clinical or suspected seizures, or moderate to severe HIE, or unexplained apnoea.
Index tests	Four leads (C3, C4, P3, P4) aEEG with raw trace and seizure‐detection algorithm interpreted by neonatal consultants with experience in interpreting aEEG
Target condition and reference standard(s)	cEEG interpreted by a senior paediatric neurologist with experience of reading neonatal EEG
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by a senior paediatric neurologist with expertise in reading neonatal EEG
Notes	The study was supported by Channel 7 Telethon Trust in the form of a fellowship awarded to AR (first author).
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	No
Could the patient flow have introduced bias?		Low risk

Rennie 2004.

Study characteristics
Patient Sampling	Case‐control design
Patient characteristics and setting	cEEG recordings with seizures and control cEEG recordings without seizures. Infants at high risk of seizure (birth depression, meningitis, clinically suspected seizure, known cerebral pathology) were enroled in original study of anti‐epileptic drugs.
Index tests	cEEG signals were converted to aEEG using Lectromed CFM, two leads (P3, P4) without raw trace and interpreted by neonatologists.
Target condition and reference standard(s)	cEEG interpreted by "blinded" interpreter
Flow and timing	Simultaneous recording
Comparative	cEEG
Notes	The study was funded by the National Lotteries Charity Board and a grant from the Bernard Sunley Foundation.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	No
Was a case‐control design avoided?	No
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Schmidt 2015.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective analysis
Patient characteristics and setting	Full‐term neonates with risk factors for seizures in the first week of life
Index tests	aEEG interpreted by a skilled neonatologist
Target condition and reference standard(s)	cEEG interpreted by skilled neurophysiologist
Flow and timing	Simultaneous recording
Comparative	cEEG interpreted by skilled neurophysiologist
Notes	The funding source was not reported. We tried to contact the study authors; however, attempts were unsuccessful.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Unclear
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Shah 2008.

Study characteristics
Patient Sampling	Prospective recruitment, retrospective aEEG analysis
Patient characteristics and setting	Infants with clinical seizures referred to NICU
Index tests	Four leads (C3, P3, C4, P4) aEEG with raw traces interpreted by consensus between two neonatal raters.
Target condition and reference standard(s)	cEEG interpreted by two neurologists with expertise in EEG
Flow and timing	Simultaneous recording
Comparative
Notes	The study was funded by a research grant from BrainZ Instruments (industry funded).
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	Yes
Could the selection of patients have introduced bias?		Low risk
Are there concerns that the included patients and setting do not match the review question?			Low concern
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Unclear
Could the reference standard, its conduct, or its interpretation have introduced bias?		Unclear risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Shellhaas 2007a.

Study characteristics
Patient Sampling	Case‐control design
Patient characteristics and setting	cEEGs with seizures and without seizures in near‐term infants
Index tests	Two lead (C3, C4) aEEG derived from cEEG interpreted by six neonatologists with varying aEEG interpretation expertise. All had ≥ I year experience with aEEG and 2 were internationally recognised experts.
Target condition and reference standard(s)	cEEG interpreted by two paediatric electroencephalographers
Flow and timing	Simultaneous recording
Comparative
Notes	No funding was sought for the study.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	No
Was a case‐control design avoided?	No
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Yamamoto 2015.

Study characteristics
Patient Sampling	Retrospective, neonatal cEEG with seizures were included
Patient characteristics and setting	Only cEEGs with seizures were included rather than infants at risk of seizures
Index tests	aEEG derived from cEEG and interpreted by neurology consultant
Target condition and reference standard(s)	cEEG interpreted by board‐certified child neurologists and epileptologist
Flow and timing	Simultaneous recording
Comparative
Notes	The funding source was not reported. We contacted the study authors and received additional information regarding patients with seizures, aEEG interpreter qualification, experience and training in aEEG interpretation, and the gestational age of infants.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Unclear
Was a case‐control design avoided?	No
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Unclear
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Unclear risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

Zhang 2011.

Study characteristics
Patient Sampling	Prospective enrolment, retrospective aEEG analysis
Patient characteristics and setting	Full‐term infants with electrographic seizures on cEEG were included.
Index tests	aEEG derived using fast Fourier transform (FFT). Raw trace was used. Interpreted by two electroencephalographers with 5‐year experience in aEEG + raw trace interpretation
Target condition and reference standard(s)	cEEG interpreted by two expert electroencephalographers
Flow and timing	Simultaneous recording
Comparative
Notes	The study was funded by the National Natural Science Foundation of China.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	Yes
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			High
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Was there an appropriate interval between index test and reference standard?	Yes
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk

aEEG: amplitude‐integrated electroencephalography; cEEG: conventional video‐electroencephalography; GA: gestational age; HIE: hypoxic‐ischaemic encephalopathy; NICU: neonatal intensive care unit

Characteristics of excluded studies [ordered by study ID]

Study	Reason for exclusion
Alfonso 2001	Study did not investigate accuracy of aEEG for the detection of seizures.
Al Naqeeb 1999	Simultaneous cEEG was not recorded.
Bogaarts 2014	Study did not investigate accuracy of aEEG for the detection of seizures.
Bourgoin 2020	Study did not investigate accuracy of aEEG for the detection of neonatal seizures.
Bridgers 1986	Study did not investigate accuracy of aEEG for the detection of seizures.
Cagdas 2011	No relevant data could be retrieved.
Chalak 2021	Simultaneous cEEG was not recorded.
Clancy 2011	Did not investigate diagnostic accuracy of aEEG for seizure detection.
Eken 1995	Study did not investigate accuracy of aEEG for the detection of seizures.
Falsaperla 2022	Review article
Fenichel 1987	Study did not investigate accuracy of aEEG for the detection of seizures.
Griesmaier 2011	Simultaneous cEEG was not recorded.
Hellstrom‐Westas 1992	Inadequate reference standard (cEEG was recorded using only 5 scalp electrodes)
Hirasawa 2009	Review article
Jalisatgi 2015	Simultaneous cEEG was not recorded.
Kadivar 2019	Review article
Kidokoro 2013	Study did not investigate accuracy of aEEG for the detection of seizures.
Klebermass 2001	Simultaneous cEEG was not recorded.
Lee 2021	Study did not investigate the accuracy of aEEG for the detection of seizures.
Lloyd 2017	Study did not aim to investigate the diagnostic accuracy of aEEG for seizure detection.
Meledin 2017	Study did not investigate the accuracy of aEEG for the detection of seizures.
Osredkar 2006	Simultaneous cEEG was not recorded.
Quigg 2009	Review article
Rakshasbhuvankar 2015	Review article
Ray 2011	Review article
Shellhaas 2007b	aEEG was not investigated. Only raw traces were studied.
Sisman 2005	Study did not investigate the accuracy of aEEG for the detection of seizures.
Toet 2002	Definition of seizure was different in the study (epileptiform activity > 5 sec compared with 10 or more seconds in the review).
Toso 2014	Study did not investigate the accuracy of aEEG for the detection of seizures.
Van Rooij 2005	Study did not investigate accuracy of aEEG for the detection of seizures.
Wusthoff 2009	aEEG was not investigated. Only raw trace was studied.
Wusthoff 2017	Review article
Zimbric 2011	aEEG was not investigated. Only raw trace was studied.

aEEG: amplitude‐integrated electroencephalography; cEEG: conventional video‐electroencephalography

Characteristics of ongoing studies [ordered by study ID]

NCT05079971.

Study name	EEG vs aEEG to Improve the Diagnosis of neonataL Seizures and Epilepsy (EAGLET)
Target condition and reference standard(s)	Neonatal seizures
Index and comparator tests	aEEG (index test) and cEEG (comparator test)
Starting date	January 2022
Contact information	Ronit M Pressler (r.pressler@ucl.ac.uk)
Notes

Differences between protocol and review

In the protocol, we planned to present diagnostic odds ratios for the outcome of 'neonates with seizures'. However, we decided not to provide diagnostic odds ratios in the review as they are not clinically intuitive. We planned to use the I² statistic to investigate heterogeneity. However, in the final review, we did not use the I² statistic due to its limitations in investigating heterogeneity. We made these changes in response to suggestions from peer reviewers; they were not driven by the findings of the review.

Contributions of authors

SCR conceptualised the review. SCR and ZZ contributed to the statistical methodology. LN contributed as a content expert to prepare the protocol. AAR prepared the protocol draft and registered the title with Cochrane. AAR and SCR independently reviewed the titles and abstracts of all articles obtained on the broad initial screening to identify potential studies. AAR and SCR independently collected data from the included studies in a prespecified form.

AAR and ZZ assessed all the included studies using the QUADAS‐2 (Quality Assessment of Diagnostic Accuracy Studies) tool. SCR and LN helped resolve any disagreements.

AAR entered data into Review Manager, and ZZ checked the data entry. SCR and ZZ performed statistical analysis. AAR, SCR, LN, and ZZ contributed to the interpretation of data. AAR prepared the first draft of the manuscript. All authors contributed to the final draft of the manuscript.

Sources of support

Internal sources

• New Source of support, Other

  No internal sources of support were sought for the review.

External sources

• National Institute for Health Research (NIHR), UK

  For author ZZ

Declarations of interest

AAR: the review included one study co‐authored by AAR (Rakshasbhuvankar 2017).

LN: the review included one study co‐authored by LN (Rakshasbhuvankar 2017).

ZZ: has declared that they have no conflict of interest.

SCR: the review included one study co‐authored by SCR (Rakshasbhuvankar 2017).

New