NAL News
The Cortical Automatic Threshold
Estimation in Adults
By Fabrice Bardy, Bram Van Dun, Harvey Dillon, Mark Seeto, Humphry Qin, Teck Loi,
and Robert Cowan
This article brings together the work we have been doing at the National Acoustic Laboratories (NAL) to make cortical
threshold testing easy, fast, and accurate. Our aim is to improve techniques for estimating hearing thresholds on people
who, for whatever reason, are unable to give reliable behavioural responses to indicate that they have heard a sound. The
particular techniques described have general applicability to cortical testing, and have been evaluated within a practical
package that will become part of the HEARLab family of test modules.
Cortical testing nicely complements auditory brainstem testing both because it is best carried out while the patient is awake
rather than asleep, and it measures more of the auditory system compared with brainstem testing. Our eventual goal is to
have a test that can be used for babies with auditory neuropathy spectrum disorder – in whom cortical thresholds can often
be measured even though auditory brainstem responses are absent. We have not achieved that goal yet, but as the article
shows, we developed a test that automatically produces reliable threshold estimates in older children and adults.
The technique is highly automated – in principle, only one button press, and no human decision making or intervention, is
needed to obtain an audiogram in both ears. Our philosophy is that if a computer is better able than humans to examine the
consistency of the cortical responses to, say, 50 repeated stimulus presentations, and is able do this simultaneously for
eight different stimuli, then it should be left to perform tests uninterrupted by humans!
Clinicians should continue doing what they do best: deal with truly complex issues. These include understanding client
beliefs and behaviors, and developing inferences from the results of diverse auditory tests.
Harvey Dillon, PhD
hile use of cortical auditory evoked potentials
(CAEPs) has been shown to be accurate and
reliable for objective threshold estimation, the
procedure is in general conducted manually
(i.e., using only one stimulus frequency presented at one intensity level at a time). The consequence is a lengthy test duration and a significant number of manual operations for the
audiologist, which renders it impractical in clinical applications. To address this issue, the National Acoustic Laboratories (NAL), through the HEARing Cooperative Research Centre,
has come up with a combination of smarter technologies.
W
SUBJECTIVE RESPONSES
Standard audiometric threshold assessment requires the client to subjectively respond when they hear a sound, either by
pressing a button or giving a verbal indication. In the case of
children, subjective responses are age-appropriate actions
such as putting a toy into a jar or turning to a lighted puppet
in response to the stimulus. For people with significant cognitive impairment, or for infants, obtaining reliable subjective
responses may be difficult or impossible. In addition, the presence of other co-morbidities may further complicate the procedure.
To enable a reliable audiogram to be established in the socalled “difficult to test” client (i.e., those with severe dementia,
those who experienced a stroke, malingerers looking for legal
compensation, children with multiple disabilities, and children
too young to respond behaviorally), there is a clear need for
use of objective testing in addition to any reliance on voluntary
responses.
HEARING THRESHOLDS
Dr. Bardy is a research electrophysiologist at the National Acoustic Laboratories (NAL) in Sydney, Australia, where Dr. Van Dun is a senior research
electrophysiologist, Dr. Dillon serves as the NAL director, Mr. Seeto is a statistician, Mr. Qin is a software engineer, and Mr. Loi is a senior development
engineer. Prof. Cowan is a professor of audiology and clinical researcher at
The University of Melbourne and Macquarie University, Australia.
32
The Hearing Journal
Hearing thresholds can be estimated reliably in adults or children through electrophysiological assessment, without the
need to rely on their behavioral feedback. This is achieved by
using sensors attached to the scalp to record the brain’s response to different sound frequencies presented at a range of
intensities. An accurate estimate of a person’s actual hearing
threshold for any specific frequency can be determined based
on the intensities that generate time-locked response to the
stimulus.
June 2016
NAL News
Figure 1. Screenshot of the CATE software while recording CAEPs from a normal-hearing adult with hearing thresholds
in both ears at frequencies 500, 1000, 2000 and 4000 Hz. For each frequency, the cortical waveforms recorded at different intensities are displayed in each sub-windows. The estimated threshold is indicated in the audiogram on the right
part of the screen. The bottom right window displays the continuous EEG.
audiogram (i.e., hearing threshold estimation at four freThere are two common ways to do this. One is through
quencies in both ears) to be acquired using objective
auditory brainstem response (ABR), done by recording the
CAEP testing in approximately 40 minutes for hearing imresponses from the auditory brainstem (i.e., lower end of the
paired subjects.
auditory path). Another way is through the auditory cortex (i.e.,
higher end of the auditory path) or CAEP. For the
ABR, responses are preferably recorded when the
person is asleep; CAEP testing is best conducted
An accurate estimate of a person’s actual hearing
when the client is awake. In cases of auditory neuthreshold for any specific frequency can be
ropathy spectrum disorder (ANSD), however,
brainstem measurements are out of the question,
determined based on the intensities that generate
and CAEPs are the only objective method with the
potential to obtain reliable hearing thresholds.
time-locked response to the stimulus.
Traditional electrophysiological assessment has
several potential drawbacks. Establishing thresholds using ABR and/or CAEP requires multiple test running at
Figure 1 shows the test screen with CAEPs at different
different frequencies and a sequence of intensities, producing
levels and derived CAEP thresholds for four audiometric fretraces that must then be subjectively evaluated by an experiquencies in two ears. The method is designed to be fully
enced electrophysiologist or audiologist trained to identify the
push-of-the-button automatic; it requires only minimal training
small peaks in the recorded waveforms that signify the presof the tester and focuses on maximizing the cortical response
ence of a response. This can be time-consuming, open to erwhile minimizing the recording time. This is made possible by
ror (it’s based on subjective judgement), and expensive, given
the use of:
the need for a skilled clinician.
novel stimuli that increase the size of the CAEP;
two EEG recording channels to weigh the response
more toward the channel with the better signal-to-noise
IMPROVEMENTS IN HEARING THRESHOLD
ratio;
ESTIMATION
the presentation of interleaved frequencies, ears and levels (to keep the stimuli unpredictable enough such that the
Recent innovations devised by NAL through the HEARbrain is kept interested);
ing CRC and described in this article enable a full adult
34
The Hearing Journal
June 2016
NAL News
statistical evaluation of the recorded responses to make
simpler the interpretation by the clinician; and
an automatic decision algorithm that decides which stimulus and level to test and when to stop testing.
We found that using two reference
electrodes results in an improved
detection sensitivity of 5 percent
as compared to a single channel
montage.
NOVEL STIMULI
The stimuli most commonly used to obtain frequency-specific
hearing thresholds using CAEPs are short tone-bursts with a
specific frequency (e.g., 1000 Hz) and a duration of around
50 ms. We recently have shown that by using multi-tone
complexes, the size of the resulting CAEPs can be increased by
32 percent for normal-hearers and 29 percent for listeners with
a hearing loss, corresponding to a potential recording time decrease of 40 percent for all audiometric frequencies except 500 Hz (Bardy.
Ear Hear 2015;36[6]:677-87; Bardy.
J Am Acad Audiol 2016;In press).
For example, the 1000 Hz pure
tone is replaced by a set of tones contained within 2/3 of an octave and
centered around 1000 Hz. We hypothesize that, despite the frequency
specificity of these sounds, the larger
cortical response is a result of increased processing in the auditory
cortex of these complex sounds (possibly inhibitory processing associated
with analyzing the multiple spectral
peaks in the signal) as compared with
processing of pure tones. The larger
amplitude is not a result of the multitone stimulus having a bigger bandwidth than the pure tone, as the effect
does not occur for bands of noise. In
the cortical automatic threshold estimation (CATE) adult protocol, each
stimulus is 50 ms long and presented
a maximum number of 120 times, with
an inter-stimulus interval varying between one and two seconds.
of interest from the auditory cortex. The common (or ground)
electrode is placed on the forehead, the active electrode on top
of the head, and the reference electrode on one of the mastoids.
This is generally referred to as a “one-channel setup,” which
shows the voltage at the active electrode relative to that at the
reference electrode. By using two reference electrodes instead
of one, and placing them on the two mastoids (and hence using
an additional channel), extra information can be simultaneously
collected, which can then be exploited by choosing or weighting
the channels considering their signal-to-noise ratio. In the extreme case—for example, when one mastoid electrode loses
contact—the other channel can be chosen without the need to
stop the current recording. On average, we found that using
two reference electrodes results in an improved detection sensitivity of 5 percent as compared to a single channel montage.
PRESENTATION OF INTERLEAVED
FREQUENCIES, EARS AND LEVELS
The brain habituates rapidly to similar stimuli. By randomizing the
frequency of the stimulus, the ear of presentation and the stimulus intensity, the brain may be kept “more interested” in the stimuli
being presented. Although the achieved benefit is usually rather
small, studies have shown larger responses can be elicited
(Bardy. Clin Neurophysiol 2014;125[4]:814-26). In particular,
TWO EEG RECORDING
CHANNELS
In most clinical applications involving
CAEPs, three electrodes are placed
on the scalp to record the responses
June 2016
The Hearing Journal
35
NAL News
Figure 2. Child tested with the new CATE approach. Although the threshold searching approach has been verified in
adults only, there is ongoing research in infants and young children.
when stimuli are being presented at an intensity level below
threshold at one frequency, it appears to the brain that the stimuli
immediately before and after it (at other frequencies and intensities) have a particularly long inter-stimulus interval. The brain “rewards” this long gap with a larger than usual cortical response.
STATISTICAL EVALUATION OF THE
RECORDED RESPONSES
The use of automatic interpretation of the response waveform
obviates the need for subjective visual interpretation by the
clinician. The software provides for an automatic interpretation of the response waveform by cutting the waveform into
predefined time slices. It subsequently measures whether any
combination of the average values within all of these slices is
significantly different from zero by taking into account the repeatability over epochs within each slice. This can be achieved
using the Hotelling’s T2 statistic, which returns a p-value expressing the level of confidence in response presence. That
is, it can be used to decide whether a response is present or
not (e.g., using a cut-off point of p = 0.01 to correct for multiple testing) (Golding. Int J Audiol 2009;48[12]:833-42).
For this statistical approach, no template is required, which is
beneficial as CAEP shapes are quite variable across ages or
even across different states of attention (Sharma. J Am Acad
Audiol 2005;16[8]:564-73). The Hotelling’s T2 generally is at
least as good as an expert human interpreter (Golding; Carter.
J Am Acad Audiol 2010;21[5]:347-56).
Table 1. Corrections and standard deviations (SDs) obtained from 20 normal-hearers and 27 listeners with
a hearing loss. These corrections can be subtracted from the thresholds found from a fully automated
CAEP audiometer to estimate behavioral thresholds.
Frequency
(Hz)
36
With normal hearing
(Mean ± SD, in dB)
With a hearing loss
(Mean ± SD, in dB)
500
17.2 ± 7.4
9.8 ± 8.9
1000
15.5 ± 6.0
5.8 ± 8.3
2000
16.8 ± 7.8
4.6 ± 10.6
4000
16.0 ± 8.4
1.9 ± 8.2
The Hearing Journal
June 2016
NAL News
In addition, statistical tests are calculated at specific residual noise values instead of a fixed interval, which is the common
approach. This allows optimization of the (variable) time intervals and the number of statistical tests during one recording
(the fewer the better). Moreover, the test can be stopped as
soon as a significant response is detected whilst keeping the
specificity (false positive rate) at an acceptable level.
AUTOMATIC DECISION ALGORITHM FOR
THE NEXT LEVEL
estimates the client’s CAEP thresholds, simple corrections or
a regression function can be applied to estimate the client’s
actual audiogram. Applications include the testing of adults
who cannot provide reliable feedback. Child evaluation has so
far tested children aged 8 to 9 years, but there is no reason
why it shouldn’t be possible to apply the technique down to
the age of 6. In fact, our current research is investigating the
feasibility of hearing threshold estimation using CAEPs in children younger than 6 years of age (Figure 2). Such a solution
would be especially valuable for infants with ANSD, for whom
hearing thresholds cannot be reliably obtained at present.
Starting from an initial intensity level of
60 dB HL and using a set of rules, the
algorithm automatically searches for the
CAEP threshold using an adaptive step
size. When the statistical detection algorithm decides that a response is present,
the presentation level is lowered (in
steps of 10 or 20 dB, depending on the
stage of testing). When the response is
absent, presentation level is increased
(in steps of 20 dB). In both cases, it is
also possible that the software decides
to stop testing the ear-specific stimulus
altogether. For example, a response detected at 20 dB HL is considered enough
evidence that the subject is normal-hearing
at that frequency. Similarly, if a response
has been detected at a specific level within
the tested range, but no response at a level
one step lower, the cortical threshold is
placed at the lowest level with a significant
response.
HEARING THRESHOLD
ESTIMATION
Table 1 shows the corrections and standard deviations obtained from 20 normalhearing adults and 27 adult listeners with
known hearing loss, using CATE to derive
auditory thresholds. The corrections are
determined as the difference between
CAEP thresholds, obtained via CATE, and
behavioral thresholds obtained through an
automatic computerized audiometry with
feedback from the listener using a push
button. These corrections and their standard deviations are similar to those
achieved with ABR (Stapells. J Speech
Lang Pathol Audiol 2000;24:74–83). A
more sophisticated correction strategy can
be used by applying a non-linear regression to the observed CAEP thresholds,
with the regressions based on the same
data that gave rise to Table 1. The CATE
approach can be used to determine hearing thresholds in a time-efficient and reliable way in adults. After the algorithm
June 2016
The Hearing Journal
37