Hearing Screening through Frequency Analysis of Auditory Brainstem Response Using PhysioNet Data

Abstract

Objectives. Responding to the reality of neonate patients with delayed childhood development due to late diagnosis of and intervention on hearing impairment, this study aims to determine the features based on time-frequency domain of auditory brainstem response (ABR) signals and to test the protocol on ABR signals from PhysioNet.Methods. This is done by pre-processing, performing time-frequency analysis, and characterizing hearing impairment using the dominant features of the ABR. In this study, normal (N) and hearing impaired (HI) ABR adult human signals were acquired from Physionet.org, a publicly available database. Considering its high signal-to-noise ratio, numerous filters and transformations were applied to extract the ABR. Consequently, the features acquired — dominant frequency and bigrams, were used as data classifiers.Results. Initial results using only N classifiers, that is features from the Normal dataset, and bandpass Chebyshev filter with a lower cut-off frequency of 60 Hz show that the tests yielded low to middle sensitivity. Further tests were done to improve the sensitivity that incorporated the HI classifiers, used data filtered with a low cut-off frequency of 300 Hz, and data divided per stimulus intensity level.Conclusion. Conclusions made are 1) data with both N and HI classifiers have higher sensitivity than those using only N classifiers, 2) data with a Chebyshev cut-off frequency of 300 Hz have a higher sensitivity than those with 60 Hz, and 3) data divided per intensity level have a higher sensitivity than data analyzed as a whole, and that features with stimulus intensity in middle ranges have a better distinction between HI and N patients.

Full text

Hearing Screening through Frequency Analysis of
Auditory Brainstem Response Using PhysioNet Data
Catherine Manuela Lee-Ramos,1 Al Francis L. Bontogon,2 Angelica S. Collanto,2 Patrick John P. Labra, MD,3
Luis G. Sison, PhD4 and Charloe M. Chiong, MD, PhD3
1Manufacturing Engineering and Management, De La Salle University, Manila, Philippines
2Electronics, Communications and Computer Engineering Department, Ateneo de Manila University, Quezon City, Philippines
3Philippine National Ear Institute, National Institutes of Health, University of the Philippines Manila, Manila, Philippines
4Electrical and Electronics Engineering Institute, University of the Philippines Diliman, Quezon City, Philippines
ABSTRACT
Objecves. Responding to the reality of neonate paents with delayed childhood development due to late diagnosis
of and intervenon on hearing impairment, this study aims to determine the features based on me-frequency
domain of auditory brainstem response (ABR) signals and to test the protocol on ABR signals from PhysioNet.
Methods. This is done by pre-processing, performing me-frequency analysis, and characterizing hearing impairment
using the dominant features of the ABR. In this study, normal (N) and hearing impaired (HI) ABR adult human signals
were acquired from Physionet.org, a publicly available database. Considering its high signal-to-noise rao, numerous
lters and transformaons were applied to extract the ABR. Consequently, the features acquired — dominant
frequency and bigrams, were used as data classiers.
Results. Inial results using only N classiers, that is features from the Normal dataset, and bandpass Chebyshev
lter with a lower cut-o frequency of 60 Hz show that the tests yielded low to middle sensivity. Further tests were
done to improve the sensivity that incorporated the HI classiers, used data ltered with a low cut-o frequency of
300 Hz, and data divided per smulus intensity level.
Conclusion. Conclusions made are 1) data with both N and HI classiers have higher sensivity than those using only
N classiers, 2) data with a Chebyshev cut-o frequency of 300 Hz have a higher sensivity than those with 60 Hz,
and 3) data divided per intensity level have a higher sensivity than data analyzed as a whole, and that features
with smulus intensity in middle ranges have a beer disncon between HI and N paents.
Keywords: evoked potenals, auditory, brain stem, delayed diagnosis, hearing loss humans
INTRODUCTION
Early detection of hearing loss is crucial for mental
and social development of the child. is is primarily
addressed by Republic Act (RA) 9709 known as “Universal
Newborn Hearing Screening and Intervention Act” which
establishes a universal newborn hearing screening program
for the Prevention, Early Diagnosis and Intervention of
Hearing Loss.1
RA 9709 mandates all newborns in the Philippines to
go through hearing screening prior to hospital discharge
or within three months after birth for those born outside
hospitals. Prior to discharge, the baby’s hearing is tested by
either otoacoustic emission (OAE) or automated auditory
brainstem response (AABR). Depending on the screening
modality used, this may be a one-step test (AABR) or a
two-step test with a re-screening session.2
eISSN 2094-9278 (Online)
Published: September 28, 2023
hps://doi.org/10.47895/amp.v57i9.4336
Corresponding author: Catherine Manuela Lee-Ramos
Manufacturing Engineering and Management
De La Salle University
2401 Ta Ave., Malate, Manila, Philippines
Email: catherine.ramos@dlsu.edu.ph
ORCiD: hps://orcid.org/0000-0002-0099-3554
VOL. 57 NO. 9 202332
ORIGINAL ARTICLE

Since the implementation of the Universal Newborn
Hearing Screening and Intervention Program, only an
estimated 10% of the hearing-impaired Filipino babies
are being screened. is is due to the high cost and lack
of availability of hearing screening devices.1 One of the
current solutions is the Hearing for Life (HeLe) project, a
collaborative project between the University of California,
Davis, the University of California, Berkeley, and the
University of the Philippines that aims to develop a low-
cost telehealth device that will enable local health units to
implement newborn hearing screening. HeLe, an Auditory
Brainstem Response (ABR) machine for neonates is currently
being locally designed and fabricated. In conjunction with
the hardware, a software algorithm is being developed and
tested so that the entire system would have the capability
of detecting automatically whether the subject is normal or
hearing-impaired. Furthermore, the algorithm should be
able to display the results immediately after the test.
For the past two decades, there has been a limited number
of ABR research. Traditionally, most Automated Auditory
Brainstem Response (AABR) machines rely on time domain
analysis of the ABR signals. Current methods of detecting
hearing impairments measure the latency of Wave V as a
function of auditory stimulus. e latency is then measured
against a standard model.3,4 If the latency falls within the
normal range, then the patient is tagged by the machine as
normal hearing and the result of the test would be a PASS.
However, if the latency falls outside of the normal range of
the standard model, then the patient is tagged as hearing-
impaired and the result of the test would be a FAIL. In this
study, the researchers will try to develop a dierent algorithm
based on the frequencies present in the ABR signals.
e study was done in collaboration with Philippine
National Ear Institute (PNEI). Clinicians from PNEI
were also consulted with regards methodology such as
how individual sweeps for one stimulus are processed and
analyzed. e researchers aim to analyze and determine what
are the distinguishing features present in the ABR signals
of normal hearing patients (N) and those with hearing
impairments (HI). Once the distinguishing features have
been obtained, further testing could be done to determine if
these features could be used to automatically give a PASS
(normal hearing) or a REFER (hearing-impaired) result
immediately after the test. Unlike previous ABR algorithms,
which focus on detection of specic features such as Wave
V, the analysis of the entire short-latency auditory-evoked
response waveform will be performed to identify any feature
that may distinguish a waveform as “PASS” or “REFER”.
METHODS
Data
Using the ABR dataset from PhysioNet, an MIT
database, the authors employed methods to analyze the ABR
signals using the frequency components present in these
signals and determine what are the unique frequencies in the
normal and hearing-impaired subjects. e Evoked Auditory
Responses in Normal and Hearing-impaired databases were
both sampled at 48 kHz.5 As of 2021, only the PhysioNet
data were publicly available. e data from PhysioNet were
validated and used in submissions in peer reviewed journals.6,7
ere are two separate databases in PhysioNet, a Normal
Database and Hearing-impaired Database.6,7 As shown in
Figure 1, the data used in this study consists of eight normal
and seven hearing-impaired adult subjects. ere are a total
of eight hearing-impaired adult subjects in the PhysioNet
database but one was excluded because it used a dierent
stimulus frequency. For the hearing-impaired, they were
conrmed with clinical tests. e patients were tested using
dierent audio stimulus intensities: 10-55, 60, 65, 70, 75,
80, 85, 90, 95, and 100 dB.
Each subject would have a number of les per stimulus
intensity. is can be seen in Tables 4 and 5. For example,
at an intensity level of less than 60 dB, there would be 23
hearing-impaired les and 98 normal hearing les. Overall,
there were 224 les from normal subjects and 204 les for
hearing impaired. For this study, the les were considered
independent of each other. Each le contains about 989
sweeps which are then averaged and analyzed as a single ABR
signal. at is at each time point, the arithmetic mean of 989
data points is obtained. Each sweep is initiated by an audio
signal of 4-kHz tones with a duration of 1 ms. ese are then
padded with silence to get a total stimulus length of 41.7 ms.
Preprocessing
Processing and analysis of the raw data from Physionet.
org were done using Matlab (Mathworks, Inc., MATLAB).
Prior to analyzing the ABR signal, dierent processing
techniques were employed to determine which one would
perform best in removing signicant noise or unwanted
signals from the ABR signals. e noise present was due to
electromagnetic interference (60 Hz) as well as the native
EEG signals. Two cut-o frequencies were used, 60 Hz and
300 Hz. It was determined that the best cut-o frequency
for the lters which yielded higher sensitivity is 300 Hz.
All sets of ABR signals will rst undergo pre-processing
to remove noise and eliminate unwanted frequencies thus
isolating the signals needed. To further specify and explain
this step, Figure 2 is shown below. e rst step is to obtain
Figure 1. ABR Dataset from PhysioNet.
VOL. 57 NO. 9 2023 33
Frequency Analysis for Hearing Screening

the ABR data. Next, 900-1000 sweeps of the ABR signal
are averaged. ird, the audio signal is removed so that the
response waves may be isolated. Next, a moving average lter
is applied which is essentially a low-pass lter to remove
the high frequency noise. A high-pass lter with a cut-o
frequency of 1 Hz is then applied to remove the DC oset
or to center the ABR signal along the x-axis (time). Finally,
a Chebyshev bandpass lter is applied to isolate the relevant
ABR frequencies. ere were several lters tested, Chebyshev,
Bessel, and Butterworth lters. However, Chebyshev was
used as this was the lter that gave a result closest to the
“textbook” ABR signal in appearance.
e study used two types of bandpass lters, one with a
cut o of 60 Hz to 1500 Hz and another with a cut o of 300
Hz to 1500 Hz. e frequency spectra of each of the steps are
plotted for verication. Figure 3 shows the raw, unprocessed
ABR data for 42 ms. e spikes aside from immediately after
the stimulus cannot be dierentiated because of the various
noise present. On the other hand, a post-processed ABR
signal from a hearing-impaired subject and the corresponding
spectrum after the high pass lter is shown in Figure 4.
Frequency Analysis
To be able to analyze the frequency, the ABR signal’s
frequency spectrum was retrieved using the Short-Time
Fourier Transform (STFT). One sweep would correspond
to 2002 datapoints. For 989 sweeps per le, there would
be around 2 million datapoints. e averaged ABR signal
would be divided into 39 segments. e segmented 2002
points will have 100 points (2ms) window size with 50% (50
points) overlap. However, only the rst 10ms (9 segments)
was analyzed since this already contains waves I-V of the
ABR signal as shown in Figure 6. e top three dominant
frequencies per segment are obtained. To obtain the dominant
frequencies, a threshold was set so only those frequencies
with sucient power relevant to the analysis would be used.
Frequencies outside the threshold would be disregarded.
To determine the frequency components that are
present in the ABR signals, the frequency resolution was
determined as 480 Hz which is computed accordingly:
where, Fs is the sampling frequency (48 kHz) and Nd is
the number of datapoints. e STF T in Matlab would show
which frequency components are present in each segment.
e dominant frequencies per segment were obtained and
formed a frequency set. e top 20% of the frequency sets,
Figure 2. Preprocessing of ABR Data.
Figure 3. Raw ABR signal of normal hearing subject for 42 ms.
-1.5
-1
-0.5
0
0.5
1
1.5
Voltage (Normalized)
Time (sec)
Raw ABR (Normal)
Figure 4. Filtered ABR signal of hearing-impaired subject and
its corresponding frequency spectrum in Hz.
Frequency resoluon = Fs/Nd (1)
VOL. 57 NO. 9 202334
Frequency Analysis for Hearing Screening

following the Pareto principle, for both normal and hearing-
impaired were obtained. e frequency sets can be visualized
using the spectrogram shown in Figures 5 and 6. From both
gures, most of the high-power frequencies for N and HI
subjects were at frequencies below 2.4 kHz.
e bigram method is then used wherein it analyzes the
frequency sets or the dominant frequencies from one time
segment to the next and is considered a feature.8 Figures
7 and 8 illustrate the bigram method. e bigram of the
marked 10 ms box in Figure 7 is shown in Figure 8. In
Figure 8, the left most numbers would be the frequencies
used for each threshold. us, frequencies up to 480 would
correspond to bin 1, frequencies up to 960 would be bin 2,
and so on. For segment 1, the three frequency components
with the highest contribution during that time duration
would be 48, 96, and 1560 Hz. is would correspond to
bins 1, 1, and 4, respectively. Similarly, the same is done for
segments 2 to 5. Each unique set of frequencies would then
be assigned a symbol, in this case, the frequencies 48, 96,
and 1560 Hz would be assigned the symbol A and so on.
A bigram would be a pair of symbols, so 48-96-1560 Hz
followed by 240-876-1240 Hz would have the bigram AB.
us, the formed bigrams for Figure 8 would be as follows,
AB, BC, CD and DC. e bigrams would be tested for
uniqueness and checked for occurrence among the les. e
top 20% unique bigrams, again using the Pareto principle
would be used as classiers for the les. However, the
sensitivity and specicity is increased if bigrams from both
normal and hearing-impaired are used, that is the bigrams
with the highest dierence of occurrence between N and HI
are used as classiers as shown in Figure 9.
Figure 5. Spectrogram of averaged ABR signal (80dB) of
hearing-impaired subject.
Figure 7. STFT Data.
Figure 6. Spectrogram of averaged ABR signal (80dB) of
normal hearing subject.
Figure 8. Sample of Bigrams.
VOL. 57 NO. 9 2023 35
Frequency Analysis for Hearing Screening

RESULTS
In analyzing the data, the frequency sets and bigram
method were used. Instead of analyzing the frequency
components of an ensemble ABR signal, the frequency sets
of each time segment were obtained. ere were 36 unique
frequency sets for the N and HI ABR. However, only 11
frequency sets were obtained by getting those with highest
dierence in occurrence between N and HI as shown in
Figure 9. is procedure was replicated for the bigram test.
Figure 9 (bottom pane) shows the multiple top bigrams
with highest dierence in occurrence between N and HI.
A total of 377 unique bigrams were extracted from the
dataset and Figure 9 which showed eighty-two (82) were
used as classiers.
Five iterations of the Bigram test were performed
depending on whether either or both Normal (N) or
Hearing-Impaired (HI) classiers were used, whether or
not the bigrams were analyzed per intensity level, and the
lower cuto frequency set for the lter. e results from these
iterations are discussed throughout the section.
Table 1 shows the sensitivity (SN) and specicity (SP)
using 224 les from normal subjects and 204 les for hearing-
impaired. ey were obtained from the test conducted with
only N classiers and using a lter with a lower cuto
frequency of 60 Hz. Table 2, on the other hand, shows the
SN and SP when both N and HI classiers were used. It can
be observed from the values in the two tables that the second
iteration yielded results that are considerably better than
the rst, which means that using both N and HI classiers
achieves better performance than using only N classiers.
e third iteration also uses both N and HI features
since it was previously determined that doing so yields better
performance rates than using only N classiers. Unlike in the
rst two, however, the lower cuto frequency of the lter in
this iteration is set to 300 Hz. e results from this test are
shown in Table 3. Comparing the values in Tables 2 and 3,
it can be seen that the lter allows for signicantly better
sensitivity and slightly improved specicity if the lower
cuto frequency is set to 300 Hz instead of 60 Hz.
In the fourth and fth iterations, the unique bigrams
for N and HI signals were then classied and analyzed per
intensity level. A lower cuto frequency of 60 Hz was set
for the lter in the fourth iteration, the results of which are
shown in Table 4. From the table, true positive rates ranging
from 60.87 to as high as 100% were achieved. As for the fth
and nal iteration, with a cut-o frequency of 300 Hz, the
bigram features were able to distinguish HI from 83.33 to
100% of the time based on the values presented in Table 5.
DISCUSSION
e bigram test, an algorithm used in a previous study to
classify electrocardiogram signals with normal sinus rhythm
and those with atrial brillation was also used in this study
to classify whether a signal was from a normal or hearing-
impaired subject.8 e bigram method involves pairing up
frequency sets of ABR signals from both N and HI patients
to produce even more distinct features that may be used to
classify the ABR signals. In this study, ve iterations of the
Figure 9. Frequency sets (A) and Bigrams (B) with highest dierence in occurrence between N and HI (frequency is in Hz).
AB
Table 2. Bigram Analysis Results using both Normal and
Hearing-impaired Classiers (Lower Cuto – 60 Hz)
Aribute Percent
Specicity 87.50 (95% CI: 82.44 - 91.53)
Sensivity 79.31 (95% CI: 73.21 - 84.74)
Table 3. Bigram Analysis Results using both Normal and
Hearing-impaired Classiers (Lower Cuto – 300 Hz)
Aribute Percent
Specicity 88.39 (95% CI: 83.46 - 92.28)
Sensivity 96.55 (95% CI: 93.06 - 98.61)
Table 1. Bigram Test using only Normal Classiers (Lower
Cuto – 60 Hz)
Aribute Percent
Specicity 85.27 (95% CI: 79.94 - 89.64)
Sensivity 59.61 (95% CI: 52.73 - 66.59)
VOL. 57 NO. 9 202336
Frequency Analysis for Hearing Screening

bigram test were performed. A total of 82 bigrams were used
for classifying the signals from the normal and hearing-
impaired subjects. However, this may have to be readjusted
when a dierent hardware is used. Raw data from Physionet.
org indicates that the noise present is signicant. us, care
should be taken to pre-process the data. One iteration of the
bigram test was performed on the dierent stimulus intensity
shown on Tables 4 and 5. is was done to see if the stimulus
intensity would aect the test. Eventually, this could be of
benet if the testing time could be reduced by using fewer
stimulus intensities.
ree factors that signicantly improved the classication
performance were identied. First, the overall accuracy was
found to be higher when both N and HI classiers were used
than when only the N classiers were considered. e results
have shown a 20% true positive rate improvement. Second,
using a Chebyshev lter with a cut-o frequency of 300 Hz
demonstrated better rates than using one with a 60-Hz cut-
o. Analysis of the results revealed that the percentage of
false negatives reduced from around 20% to 3%. Lastly, the
results further improved when the data had been partitioned
according to their intensity levels. Data divided per intensity
level have more accuracy than data analyzed as a whole, and
that middle level ranges have a better accuracy than at the
extremes. All the true negatives have percentages of 90% and
above. All intensities, except 60 Hz, have a false negative rate
of less than 10%.
e raw data from PhysioNet were successfully processed
to acquire the required ABR signal for analysis. Various signal
processing techniques were properly implemented wherein
the ABR signal to be analyzed resembles an ABR signal from
that of a commercial device. Although the PhysioNet data
can still be further processed as there are some parameters
that can still be modied to produce new ndings.
e entire procedure from the processing of data to the
results of frequency analysis were executed in MATLAB.
Instead of hours, the whole process was done in 8 - 30
seconds. is test time proves vital to ease up the newborn
hearing screening process. ough this time might still vary
depending on the capabilities of the HeLe hardware. It is also
recommended to use on a hardware with a higher sampling
rate to enable ner distinction in the frequencies present in
the ABR signals.
Currently, the study is only limited with the usage of data
from PhysioNet. For further trials, it would be best to test the
algorithm using data of newborn subjects from commercial
devices. Moreover, the number of sweeps of a certain data
might be a factor in the analysis of frequency. Testing the
current algorithm using a lower number of sweeps (600 or
less) might result in new ndings. ough it demands more
time, another possibility is to test the algorithm for each
sweep as there might be some features to consider.
With the current study focused on newborn hearing
screening, chances are that the frequency analysis algorithm
Table 5. Bigram Analysis Results for Varied Smulus Intensies using both Normal and Hearing-impaired
Classiers (Lower Cuto – 300 Hz)
dB # HI les # N les Specicity Sensivity
Lower 60 23 98 95.92 (95% CI: 89.88 - 98.88) 100.00 (95% CI: 85.18 - 100.00)
60 12 13 92.31 (95% CI: 63.97 - 99.81) 83.30 (95% CI: 51.59 - 97.91)
65 12 17 88.23 (95% CI: 63.56 - 98.54) 100.00 (95% CI: 73.54 - 100.00)
70 20 13 92.31 (95% CI: 63.97 - 99.81) 100.00 (95% CI: 83.16 - 100.00)
75 20 10 90.00 (95% CI: 55.50 - 99.75) 100.00 (95% CI: 83.16 - 100.00)
80 24 13 92.31 (95% CI: 63.97 - 99.81) 95.83 (95% CI: 78.88 - 99.89)
85 20 15 93.33 (95% CI: 68.05 - 99.83) 95.00 (95% CI: 75.13 - 99.87)
90 24 11 90.91 (95% CI: 58.72 - 99.77) 100.00 (95% CI: 85.75 - 100.00)
95 24 18 94.44 (95% CI: 72.71 - 99.86) 100.00 (95% CI: 85.75 - 100.00)
100 24 16 93.75 (95% CI: 69.77 - 99.84) 91.67 (95% CI: 73.00 - 98.97)
Table 4. Bigram Analysis Results for Varied Smulus Intensies using both Normal and Hearing-impaired
Classiers (Lower Cuto – 60 Hz)
dB # HI les # N les Specicity Sensivity
Lower 60 23 98 66.32 (95% CI: 56.07 - 75.56) 60.87 (95% CI: 38.54 - 80.29)
60 12 13 92.31 (95% CI: 63.97 - 99.81) 100.00 (95% CI: 73.54 - 100.00)
65 12 17 94.12 (95% CI: 71.31 - 99.85) 91.67 (95% CI: 61.52 - 99.79)
70 20 13 92.31 (95% CI: 63.97- 99.81) 90.00 (95% CI: 68.30 - 98.77)
75 20 10 90.00 (95% CI: 55.50 - 99.75) 80.00 (95% CI: 56.34 - 94.27)
80 24 13 92.31 (95% CI: 63.97 - 99.81) 87.50 (95% CI: 67.64 - 97.34)
85 20 15 73.33 (95% CI: 44.90 - 92.21) 90.00 (95% CI: 68.30 - 98.77)
90 24 11 90.91 (95% CI: 58.72 - 99.77) 83.33 (95% CI: 62.62 - 95.26)
95 24 18 94.44 (95% CI: 72.71 - 99.86) 83.33 (95% CI: 62.62 - 95.26)
100 24 16 93.75 (95% CI: 69.77 - 99.84) 95.83 (95% CI: 78.88 - 99.89)
VOL. 57 NO. 9 2023 37
Frequency Analysis for Hearing Screening

can also be explored for other auditory problems. Furthermore,
since frequency is the center feature, the algorithm could be
used to model and trace the changes in frequency along the
auditory pathway. It would be interesting, for example, to see
the frequency components present in the dierent waveforms
(I to V), particularly Wave V, as this might give a clue on
the conditions of the pathway. is is the recommendation as
well of Paulraj, a closer look at Wave V as this is instrumental
in classifying normal and hearing-impaired subjects.9 is
will supplement studies based on the traveling wave delay of
our auditory system.
CONCLUSION
is study shows that there is merit in studying the
frequency components of the ABR signal. It might lead to
new ways of classifying normal or hearing-impaired subjects
other than looking at the latencies of the dierent waveforms.
By combining both time and frequency domain in the bigram
method, new features can be extracted from the ABR signals.
Statement of Authorship
All authors certied fulllment of ICMJE authorship
criteria.
Author Disclosure
All authors declared no conicts of interest.
Funding Source
e study was funded by the Philippine-California
Advanced Research Institutes (PCARI).
REFERENCES
1. Hele.ph, Hearing for Life [Internet]. 2018 [cited 2018 Oct]. Available
from: http://hele.ph/.
2. Republic Act 9288 An Act Promulgating A Comprehensive Policy
And A National System For Ensuring Newborn Screening, Vol. 2.
2018 (Phil.).
3. Starr A, Achor J. Auditory brain stem responses in neurological
disease. Arch Neurol. 1975 Nov;32(11):761-8. doi: 10.1001/archneur.
1975.00490530083009.
4. Hecox K, Galambos R. Brain stem auditory evoked responses in
human infants and adults. Arch Otolaryngol. 1974 Jan;99(1):30-3.
doi: 10.1001/archotol.1974.00780030034006.
5. Goldberger AL, Amaral LA, Glass L, Hausdor JM, Ivanov PC, Mark
RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: Components of
a new research resource for complex physiologic signals. Circulation.
2000 Jun;101(23):e215-20. doi: 10.1161/01.cir.101.23.e215.
6. Silva I, Epstein M. Estimating loudness growth from tone-burst
evoked responses. J Acoust Soc Am. 2010 Jun;127(6):3629-42.
doi: 10.1121/1.3397457.
7. Silva I, Epstein M. Objective estimation of loudness growth in
hearing-impaired listeners. J Acoust Soc Am. 2012 Jan;131(1):
353-62. doi:10.1121/1.3666024.
8. Santos M. Electrocardiogram Analysis Using Constellation Maps
and Bigrams. [MS esis, Ateneo de Manila University]. 2011.
9. Paulraj MP, Subramaniam K, Bin Yaccob S, Bin Adom AH,
Hema CR. Auditory evoked potential response and hearing loss:
a review. Open Biomed Eng J. 2015 Feb;9:17-24. doi: 10.2174/
1874120701509010017.
VOL. 57 NO. 9 202338
Frequency Analysis for Hearing Screening