Keywords
amplitude-integrated EEG - neonatal neurology - neonatal seizures - brain monitoring
Amplitude-integrated electroencephalography (aEEG) is an increasingly popular tool
to screen for neonatal seizures. aEEG uses a limited number of channels to record
raw electroencephalography (EEG) data which is filtered, processed, and displayed
on a semilogarithmic amplitude and time-compressed scale.[1] Seizures appear as abrupt increases in the voltage on the compressed tracing.[1]
[2] There are several challenges in the interpretation of aEEG. Due to the time compression
of a standard aEEG display, brief neonatal seizures may be missed, as a seizure lasting
90 seconds would only appear for 1.4 mm on the screen; seizures lasting less than
30 seconds may not be detected.[1]
[2] As aEEG filters out frequencies slower than 2 Hz, neonatal seizures which are of
very low frequency may be filtered out of the tracing.[1] Due to the limited electrode array focused on the central/parietal regions, focal
neonatal seizures which are outside of these regions may not be detected by aEEG.[3] Additionally, artifacts are very common in aEEG and it can be difficult to distinguish
seizure from artifacts such as patting, bedside nursing care, electrocardiographic
artifact, or electrode malfunction.[1]
While expert aEEG interpretation has been shown to have good sensitivity and specificity
for seizures, typical use may have lower accuracy. The utility of any screening test
depends upon how well the test result refines an estimate that a patient has a given
condition, moving the pretest probability to a more accurate post-test probability.
The potential utility of aEEG to screen for seizures is dependent on the risk of seizures
in the patient population undergoing aEEG (pretest probability) and accuracy of the
aEEG when interpreted by providers of varying skill levels (test characteristics).
For a neonate at risk for seizures, the pretest probability may be estimated based
on published evidence regarding the prevalence of seizures in a specific population.
A likelihood ratio (LR) incorporates sensitivity and specificity to assess test utility.
Reported sensitivities and specificities for either diagnosing or ruling out seizures
range widely, largely dependent on user skill in aEEG interpretation. This produces
varying evidence-based LRs for aEEG in different described contexts. To demonstrate
the importance of seizure prevalence and expertise in aEEG interpretation, we applied
statistical principles to model utility of aEEG to move the pretest probability of
seizures to a more accurate post-test probability across three common clinical scenarios.
Materials and Methods
The prevalence, sensitivity, and specificity of aEEG were derived from reported values
in the literature. The prevalence of seizures in three clinical conditions affecting
neonates (hypoxic ischemic encephalopathy [HIE] in term neonates ≥37 weeks GA,[4] bacterial meningitis in term and late preterm neonates ≥33 weeks GA,[5]
[6] and prematurity in neonates <30 weeks GA[7]
[8]
[9]) were used to derive the pretest probability of seizures for scenarios based on
each condition. The sensitivity and specificity of aEEG for seizure detection were
derived from the literature for three models of aEEG interpretation based on experience
level: ideal/expert interpretation,[10]
[11]
[12] intermediate interpretation,[10]
[13] and the lowest reported values with inexperienced users.[10]
[14]
[15] Expert interpretation of aEEG was performed by senior neonatologists with experience
interpreting aEEG and epileptologists with experience reading aEEG.[10]
[11]
[12] Intermediate interpretation was performed by epileptologists with experience reading
conventional EEG and neonatology fellows.[10]
[13]
[15] Inexperienced interpretation was performed by medical students and neonatologists
with no prior experience interpreting aEEG.[10]
[14]
[15] The utility of aEEG for seizure screening was calculated via post-test probabilities
from LRs.
Results
The first scenario considered an example of HIE, reported to confer a relatively high
risk of neonatal seizures. [Table 1] shows the reported sensitivity and specificity from the literature applied to model
differences in aEEG interpretation between the expert, intermediate, and inexperienced
levels. The reported prevalence of seizures in hypoxic-ischemic encephalopathy is
approximately 40%.[4] The best reported sensitivity for aEEG in seizure detection is 85%, specificity
is 90%.[10]
[11]
[12] Applying these reported values to model expert interpretation yields a positive
likelihood ratio (LR+) of 8.5 and a negative likelihood ratio (LR−) of 0.17 for this
test. Therefore, under this model a neonate with HIE starts with a general pretest
probability of 0.4 for seizures (based on 40% prevalence in the literature), in the
absence of any information from aEEG. If expert interpretation finds seizures on aEEG,
an individual patient then has a post-test probability of 0.85 of true seizures, reflecting
a considerable change in the estimated seizure risk for that neonate ([Fig. 1]). At the same time, a negative aEEG by expert interpretation gives a post-test probability
of 0.1 for seizures, which is meaningfully different from the pretest estimate of
0.4. The usefulness of the test changes if there is lower accuracy in aEEG interpretation.
Using moderate reported values of sensitivity (65%) and specificity (70%)[10]
[13] to model intermediate interpretation skill, aEEG has LR+ of 2.17 and LR− 0.5. This
means that for a neonate with HIE, again starting with a pretest probability of 0.4
for seizures, an aEEG positive for seizures by intermediate interpretation translates
to a post-test probability of only 0.59 for that patient having true seizures. Likewise,
a neonate with HIE having an aEEG negative for seizures by intermediate interpretation
still has a post-test probability of 0.25 that true seizures are present. Finally,
the least accurate, inexperienced interpretation was modeled using the lowest reported
test characteristics among users with limited aEEG accuracy (sensitivity = 40%, specificity = 50%),[10]
[14]
[15] with LR+ 0.8 and LR− 1.2. This again was applied to the neonate with HIE who is
known to have a risk of seizures of 0.4 without aEEG information—if the least accurate,
inexperienced user finds that patient's aEEG positive for seizures, the post-test
probability of true seizures is only 0.35. Similarly, if the inexperienced user finds
the aEEG negative for seizures, there remains a 0.44 post-test probability of true
seizures. Under the poor test characteristics for inexperienced aEEG interpretation,
based on the least accurate sensitivity and specificity previously reported, the test
paradoxically makes the estimate of seizure risk less accurate than it was prior to
the test results.
Table 1
Models of aEEG interpretation based on reported sensitivity and specificity for expert
level users, intermediate level interpretation, and the lowest reported values
|
Expert interpretation
|
Intermediate interpretation
|
Inexperienced interpretation
|
|
Sensitivity
|
85%
|
65%
|
40%
|
|
Specificity
|
90%
|
70%
|
50%
|
|
Positive likelihood ratio
|
8.5
|
2.17
|
0.80
|
|
Negative likelihood ratio
|
0.17
|
0.5
|
1.2
|
Abbreviation: aEEG, amplitude-integrated electroencephalography.
Fig. 1 Model of aEEG interpretation for users with varied experience levels in neonates
with HIE, with reported seizure prevalence of 40%. aEEG, amplitude-integrated electroencephalography;
HIE, hypoxic ischemic encephalopathy.
Another clinical scenario with risk of seizures is neonatal bacterial meningitis.
The reported prevalence of seizures in neonates with bacterial meningitis ranges from
20 to 34%.[5]
[6] Based on this, a prevalence of 27% can be used to model aEEG utility based on experience
level, using the same test characteristics detailed above ([Fig. 2]). From a pretest probability of 0.27 for seizures in neonatal bacterial meningitis
without aEEG information available, an aEEG positive for seizures by expert interpretation
leads to a more accurate post-test probability of 0.76 for true seizures. Conversely,
if aEEG is negative for seizures by expert interpretation, the post-test probability
of true seizures is only 0.06. Using intermediate interpretation, an aEEG positive
for seizures increases the post-test probability of true seizures to 0.45. While this
is a meaningfully higher probability of seizures than was estimated before the test,
it is notable that an aEEG positive for seizures in this situation still only means
there is a 45% chance an individual patient truly has seizures. Conversely, an aEEG
negative for seizures by intermediate interpretation creates a post-test probability
of 0.16 for true seizures. Finally, in the model of lower reported sensitivity and
specificity as above, the post-test probability of true seizures after positive aEEG
is only 0.23 and is 0.31 even after negative aEEG.
Fig. 2 Model of aEEG interpretation for users with varied experience levels in neonates
with bacterial meningitis, with reported seizure prevalence of 27%. aEEG, amplitude-integrated
electroencephalography.
Finally, we can model the use of aEEG for neonates with lower risk for seizures (low
pretest probability), using the example of reported prevalence of EEG confirmed seizures
in extremely and very preterm (<30 weeks gestational age) neonates of 5%.[7]
[8]
[9] Because the post-test probability of a condition depends both on the test characteristics
and the pretest prevalence of the condition, aEEG used to screen for seizures in patients
at lower risk for seizures yields lower overall post-test probabilities for true seizures,
even if aEEG is positive for seizures. Expert interpretation (sensitivity 85%, specificity
90%) was applied to the scenario of the extremely preterm neonate using the LR and
prevalence data above ([Fig. 3]). Starting with a pretest probability of 0.05 for seizure, an aEEG positive for
seizure by expert interpretation results in a post-test probability of 0.31 for true
seizures, meaning the patient is more likely to have a false positive aEEG than true
seizures. The post-test probability for true seizures is 0.01 after negative aEEG
by expert interpretation. Intermediate interpretation (sensitivity = 65%, specificity = 70%)
carries even lower utility. This gives post-test probabilities of 0.1 for true seizures
after positive aEEG, and 0.03 of seizures after negative aEEG. For the least accurate
users (sensitivity = 40%, specificity = 50%), post-test probabilities of 0.04 of seizures
after positive aEEG and 0.06 of seizures after negative aEEG reflect no gain of any
useful information.
Fig. 3 Model of aEEG interpretation for users with varied experience levels in premature
neonates with gestational age <30 weeks, with reported seizure prevalence of 5%. aEEG,
amplitude-integrated electroencephalography.
Discussion
With sufficiently high pretest probability of seizures and expert level accuracy in
interpretation, corresponding to high sensitivity and specificity, aEEG is a useful
screening tool for the detection of neonatal seizures. In neonates with HIE and expert
level interpretation, post-test probabilities of 0.85 for true seizures after positive
aEEG and 0.1 for seizures after negative aEEG make this a helpful and informative
tool. However, aEEG was moderately useful in the setting of HIE (high pretest probability)
with a nonexpert, intermediate level of interpretation. Worrisomely, when the lowest
reported sensitivity and specificity was applied to the scenario of HIE, there was
no meaningful information gained from the test, with post-test probabilities of 0.35
of having seizures after a positive aEEG, and 0.44 of seizures after a negative aEEG.
As the pretest probability of seizures gets lower for different clinical scenarios,
the utility of aEEG decreases as well. In neonates with bacterial meningitis with
a seizure prevalence of approximately 27%, aEEG remained a useful tool under conditions
of expert interpretation, with post-test probabilities of 0.76 for true seizures after
positive aEEG and 0.06 of seizures after negative aEEG. However, with intermediate
level interpretation, the post-test probability of true seizures was below 0.5 after
positive aEEG. This suggests that while a positive aEEG might identify those neonates
with a relatively higher risk of seizures to undergo more definitive testing, caution
should be used before initiating seizure treatment based on a positive aEEG in such
a scenario. With the lowest reported sensitivity and specificity, there was again
no meaningful information added from the test.
The scenario of the extremely preterm neonate illustrates a case that is very common,
and also reflects the need for caution before acting on aEEG results in a neonate
at low risk for seizures. Extremely preterm neonates have been reported to have an
overall seizure prevalence of 5% when the gold standard of full array continuous EEG
monitoring is applied prospectively.[7] Because this is a relatively low pretest probability of seizures, even with the
most expert level of aEEG interpretation, a positive test is not necessarily diagnostic.
In our model, even with expert interpretation, the post-test probability of true seizures
after positive aEEG was only 0.31. Again, while this may be high enough to justify
more intensive diagnostics or monitoring, it may not be appropriate to initiate treatment
based on this post-test probability alone.
This study has important limitations. These theoretical models, while based on evidence
reported in the literature, cannot capture the nuances of every potential clinical
scenario. We do not mean to claim that the post-test probabilities calculated here
are accurate for actual individual patients. Rather, we chose these cases to illustrate
examples of where aEEG could be most useful and when it might have lower utility.
The pretest probabilities used are based on estimates of seizure prevalence for the
selected conditions reported in the literature from population-based studies of large
cohorts of neonates. The clinical scenarios outlined were selected to offer conditions
commonly encountered in the neonatal intensive care unit with differing seizure prevalence
to highlight the impact this has on test utility. Our estimates of the test characteristics
sensitivity and specificity are based on those published across multiple studies of
aEEG accuracy. Some of these studies used older equipment or methods of review, such
as single-channel aEEG, that are less relevant to current practice at many centers.
The estimates here for “expert,” “intermediate,” and “inexperienced” accuracy are
only estimates; individual users may fall in between these categories or vary in accuracy
in different situations. At the same time, these examples clearly illustrate the impact
of varying skill in interpretation—with lower accuracy, aEEG gives much less useful
information. This suggests that those using aEEG for seizure detection would benefit
from having insight into their own skill level, thus recognizing their approximate
sensitivity and specificity in interpretation. It also creates an opportunity to improve
aEEG utility as increased experience and training to improve interpretation skill
will improve test characteristics for a given user. Finally, we do not offer primary
evidence from actual patient populations regarding how aEEG may improve care or change
outcomes. This work uses only statistical approaches to illustrate examples of how
patient selection and skill in interpretation impact aEEG utility, and should not
be considered as definitive recommendations for which patients should or should not
be monitored with aEEG.
Conclusion
aEEG is very useful when appropriately used to screen for neonatal seizures. However,
under some circumstances it can be much less helpful and even negligible. Providers
using aEEG must recognize the importance of patient selection for aEEG based on evidence
of sufficiently high pretest probability. Similarly, insight into skill level of aEEG
interpretations at one's own center can inform test utility in a specific context.
Continuing training and education in aEEG interpretation will allow users to improve
the test characteristics at their sites, while further research regarding prevalence
of neonatal seizures in different populations is needed to inform pretest probability
estimates.
