662IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 4, APRIL 2006
A Robust Method for Estimating Respiratory Flow
Using Tracheal Sounds Entropy
Azadeh Yadollahi, Student Member, IEEE, and Zahra M. K. Moussavi, Senior Member, IEEE
is of great interest for researchers and physicians due to its diag-
nostic potentials. Due to difficulties and inaccuracy of most of the
flow measurement techniques, several researchers have attempted
to estimate flow from respiratory sounds. However, all of the pro-
posed methods heavily depend on the availability of different rates
of flow for calibrating the model, which makes their use limited by
a large degree. In this paper, a robust and novel method for esti-
matingflowusingentropy ofthebandpass filtered trachealsounds
is proposed. The proposed method is novel in terms of being inde-
pendent of the flow rate chosen for calibration; it requires only one
breath for calibration and can estimate any flow rate even out of
the range of calibration flow. After removing the effects of heart
sounds (which distort the low-frequency components of tracheal
sounds) on the calculated entropy of the tracheal sounds, the per-
formance of the method at different frequency ranges were inves-
tigated. Also, the performance of the proposed method was tested
using 6 different segment sizes for entropy calculation and the best
segment sizes during inspiration and expiration were found. The
method was tested on data of 10 healthy subjects at five different
2.8%forinspiration andexpiration phases,respectively.
Index Terms—Entropy, flow estimation, heart sounds, tracheal
nulae connected to a pressure transducer, heated thermistor or
anemometry. It can be also measured indirectly by means of
detecting chest or abdominal movements using respiratory in-
ductance plethysmography (RIP), strain gauges or magnetome-
method for measuring flow , it changes the breathing pattern
–. Also this device can not be used during swallowing as-
sessment. Therefore, the combined use of nasal cannulae con-
nected to a pressure transducer and the measurement of respira-
been recommended as the best approach in assessing respira-
tory and swallowing patterns . However, application of these
ESPIRATORY flow measurement is usually achieved by
spirometry devices such as pneumotachograph, nasal can-
Manuscript received February 15, 2005; revised August 5, 2005. This study
was supported by Natural Sciences and Engineering Research Council of
Canada (NSERC). Asterisk indicates corresponding author.
A. Yadollahi was with the Department of Electrical Engineering, Sharif uni-
versity of Technology, Tehran, Iran. She is now with the Department of Elec-
trical and Computer Engineering, University of Manitoba, Winnipeg, MB R3T
2N2, Canada (e-mail: email@example.com).
*Z. M. K. Moussavi is with the Department of Electrical and Computer En-
gineering, University of Manitoba, Winnipeg, MB R3T 2N2, Canada, and also
with the Department of Electrical and Computer Engineering, University of
Digital Object Identifier 10.1109/TBME.2006.870231
techniques has some disadvantages, especially when studying
children with neurological impairments, in whom the study of
swallowing is clinically important .
Many investigators have attempted to study the relationship
of flow with breath sounds using different features of tracheal
and lung sounds –. Tracheal sound mean amplitude, av-
erage power, mean power frequency and the multiplication of
tracheal sound mean frequency and mean amplitude in relation
to flow were investigated in , in which their best result for
flow estimation was from tracheal sounds average power with
6.6% error. Comparing different models for estimating
the exponential model is superior to the polynomial and power
models , . The method proposed in  achieved flow
estimation with a low error but it relied on a sophisticated cal-
ibration process requiring a copy of the flow and breath sound
signals at various flow rates for tuning the model parameters.
As a follow up of that study , two different methods for cal-
ibrating the model were suggested in  but at the cost of a
much higher error.
Although some of the above mentioned methods achieved a
reasonably low error in flow estimation, however all of them
are heavily dependent on the calibration part; they need to have
a copy of the breaths at every target flow during calibration.
In one study , flow rate was assumed to be constant and
half of the data was used for calibration. In the other studies
in which a variety of flow rates were considered, different sets
of parameters were used for estimating flow at different flow
rates with the assumption that a copy of every flow rate is
possible to capture respiratory sounds at different flow rates for
calibration, especially when assessing young children, patients
with neurological impairments and/or patients in emergency
conditions. Furthermore, the average error of these methods
was more than 10% , , , except in one  which
3.0% but at the cost of a much more complicated
The respiratory sound features used in previous studies to es-
timate flow were calculated from either the mean amplitude or
average power of the sound signal. Respiratory sounds are sto-
chastic signals and nonstationary in nature. Given the fact that
the mean amplitude and average power are only the first and
second order moments of the signal, they do not represent the
full statistical properties of respiratory sounds. Thus, we sought
a feature that represents more of the statistical properties of the
breath sounds. Entropy is a measure that involves calculation of
probability density function (PDF) of the signal and, therefore,
0018-9294/$20.00 © 2006 IEEE
YADOLLAHI AND MOUSSAVI: ROBUST METHOD FOR ESTIMATING RESPIRATORY FLOW USING TRACHEAL SOUNDS ENTROPY 663
this study was to investigate the relationship between flow and
resenting such relationship to estimate flow. The objective was
to find a robust flow estimation method that does not require
more than one breath cycle for calibration and can adjust itself
with flow variability.
In studying respiratory sounds, heart beat is an unavoidable
source of interference that changes both frequency and time
characteristics of the respiratory sounds. Most of the heart
sounds energy is concentrated in the frequency range of 20–200
Hz , which overlaps with the low-frequency components
of respiratory sounds. In most of the acoustical flow estima-
tion methods –, the tracheal sound was analyzed for
frequency range above 300 Hz, where it is almost free of the
heart sounds effect. However, we found that for very shallow
breathing where the tracheal sound has very low intensity, it
is important to consider the frequency range below 300 Hz for
flow estimation. On the other hand, heart sounds interfere with
the respiratory sounds in the frequency range below 300 Hz
and also our pilot studies showed a high inaccuracy in very low
flow estimation due to heart sounds. Therefore, we investigated
several methods to cancel the effect of heart sounds prior to
As the effect of heart sounds on lung sounds is much more
pronounced than that on tracheal sound, all of the heart sounds
cancellation techniques have been applied to lung sounds. Sev-
eral methods based on adaptive filtering –, wavelet de-
noising , , adaptive thresholding and 2-D interpolation
of lungsoundsinthetime-frequencydomain and removing
heart sounds-included segments from the wavelet coefficients
of lung sounds and then reconstructing the signal by auto re-
gressive or moving average models  have been proposed for
heart sounds reduction from lung sounds.
We employed some of the above mentioned techniques but
the results were not satisfactory for flow estimation purpose.
This is probably due to the fact that by applying most of these
techniques for heart sounds cancellation, some of the heart
sounds still remain in the respiratory sound signal. On the other
hand, some other techniques alter the respiratory sound signal.
Although these changes are not much noticeable by auditory
means, they increase the error of flow estimation at shallow
breathing. Therefore, instead of heart sounds cancellation from
the tracheal sound, we sought a new technique for canceling
the effect of heart sounds on the entropy of the tracheal sound,
which is explained below.
Flow is a changing signal and flow-sound relationship may
differ as the flow changes from zero to its maximum. In our
previous study , it was shown that the best results for es-
timating flow was achieved in the upper (lower) 40% values of
flow during inspiration (expiration), where the change of flow is
slower. Thus, in this study also, the upper and lower 40% of the
flow within a breath cycle were estimated by the models under
investigation. Then, the flow between the upper and lower 40%
of inspiration and expiration were found by linear interpolation.
Entropy is a term borrowed from the Thermodynamics repre-
senting the degree of uncertainty (complexity) of the state of
a system. In information theory, entropy is the measure of a
,, theShannonentropy isdefinedas 
In our study of respiratory sounds, we use neither of the two
exact descriptions of entropy as used in communication or ther-
modynamics. Here, we call the measure defined in (1) as the
entropy for its mathematical form. Since entropy is based on
the PDF of the signal, accurate estimation of the PDF is of great
fast way for estimating the PDF, its accuracy declines when the
number of samples is low. In this study, the PDF of the signal
was estimated using Normal kernel estimator. More details of
the method could be found in .
B. Data Collection
Data of 10 healthy subjects were adopted from previous
studies , . The data were chosen from two age groups: 5
adults (all female) 29
8 years old and 5 children (3 female)
1.7 years old. Respiratory sounds were recorded using
Siemens accelerometers (EMT25C) placed over suprasternal
notch and the upper right lobe lung. The accelerometers were
attached to the subjects’ skin using doubles sided tapes. Res-
piratory flow was measured by a pneumotachograph (Fleisch
no. 3) connected to a differential pressure transducer (Validyne,
Northridge, CA). In order to reduce the ambient noise, the
subjects were seated in an acoustic chamber. Subjects were
instructed to breathe at 5 different flow rates with 5 breaths at
each target flow followed by a 10 s of breath hold at the end of
experiment. In this study, the shallow ( 6 ml/s/kg), low (6–9
ml/s/kg), medium (12–18 ml/s/kg), high (18–27 ml/s/kg), and
very high ( 27 ml/s/kg) target flow rates were chosen. Breath
sounds and flow signals were digitized simultaneously at a rate
of 10240 Hz. The signals were digitized at much higher rate
than practically needed because of the instrumentation set up
and a lack of constraints on the amount of data. However, the
flow signals were later decimated to 320 Hz. Tracheal sound
signals were used for flow estimation while the lung sound
signal in correspondence with tracheal sound signal were used
for respiratory phase detection using the method introduced in
 and .
C. Signal Processing
The entire signal processing of this study was performed
using MATLAB 6.5. In the first step, tracheal sounds were
bandpass filtered in the range of 300–1000 Hz to remove the
effects of heart sounds and high-frequency noise. Then the
bandpass filtered signal was sequestered into segments of
50 ms (512 samples) with 75% overlap between successive
segments. The PDF of each segment was estimated using the
Normal kernel estimator followed by computing the Shannon
entropy. Based on the results of the previous studies , ,
664IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 4, APRIL 2006
three different models were investigated for flow estimation
using entropy of the bandpass filtered tracheal sounds.
3 are the model parameters. Here we assumed that only one
able to calibrate the model and derive the model coefficients for
each subject. This breath is called the base during the rest of
this paper. According to the fact that the mechanisms of gen-
erating sound during inspiratory and expiratory phases are dif-
ferent ,  each phase was analyzed separately and the
model coefficients were derived for each phase.
We found that for the linear model (2) the slope,
at different flow rates resulting in an over/under estimation at
flow rates higher/lower than the base flow. Thus, the linear
model was modified to take into account the effects of
different flow rates. The modified linear model was defined as
is the entropy values of tracheal sounds in each res-
is the estimated flow and, , 2,
After calibrating each model the flow was estimated for inspira-
tory and expiratory phases separately. The error was defined as
is the upper 40% of the entropy values of the base.
where F and
estimated flow at each respiratory phase, respectively. The error
was averaged within and between the subjects. The model with
the least error in flow estimation was considered as the superior
Once the superior model for the most normal range of flow
was found, then it was tuned to improve its accuracy. First,
different frequency ranges for bandpass filtering the tracheal
sounds were investigated. The goal was to test not only the per-
formance of the method at different frequency ranges but also
to find the best range to improve the acquired results at shallow
Hz, 150–600 Hz, 150–1000 Hz, 300–450 Hz, 300–600 Hz, and
300–1000 Hz. However, as mentioned in Section I when con-
sidering the tracheal sound in the range below 300 Hz, heart
sounds interfere with the time and frequency characteristics of
tracheal sound. Thus, a technique was sought to cancel the ef-
fects of heart sounds on the tracheal sounds entropy.
The first step in all of the methods for canceling heart sounds
from respiratory sounds except the wavelet de-noising method
is to localize the segments including heart sounds. Comparing
the results of different methods for localizing heart sounds in
the lung sound recordings, Shannon entropy was found to be
superior to the other methods with the average error of 0.1
0.4 and 1.0 0.7 at low and medium flow rates, respectively
; hence, we used that method in this study.
are the upper 40% values of the actual and the
different models at different flow rates and the overall error in the frequency
range of 300–1000 Hz and the segment size of 50 ms. (a) Inspiration. (b)
Mean ? standard deviation ?? ? ?? of flow estimation error (%) for
and Shannon entropy were calculated for the segments void of
heart sounds. Then, spline interpolation was applied to estimate
This technique effectively cancels the effect of heart sounds on
the entropy of the tracheal sound. Then, for each of the 11 fre-
timated with the best model chosen previously. The frequency
range for calculating entropy. After defining the best frequency
segment sizes to calculate the entropy of the tracheal sound on
the performance of the method was also investigated.
Once the model was tuned to perform the best, the entire flow
respiratory phases were detected using the method in , 
usingboth lungand tracheal sounds. It was shown thatthe accu-
racy of this method was 100% in all subjects . The computa-
tional cost of the flow estimation method in this study was also
considered and investigated compared to those of the previous
Fig. 1 shows the results of different models, (3)–(5), for
estimating flow from the entropy of the bandpass filtered
(300–1000 Hz) tracheal sounds using a 50-ms segment size
with 75% overlap between the adjacent segments. The error
was averaged within and between the subjects for inspiration
and expiration phases. Overall, the results showed that the mod-
ified linear model was superior to the power and second-order
model during both inspiration and expiration phases. Also, the
performance of the modified linear model was more consistent
YADOLLAHI AND MOUSSAVI: ROBUST METHOD FOR ESTIMATING RESPIRATORY FLOW USING TRACHEAL SOUNDS ENTROPY 665
(%) of the modified linear model in different frequency ranges and the segment
size of 50 ms. (a) Inspiration. (b) Expiration.
Mean? standarddeviation ????? oftheoverall flowestimationerror
at different flow rates than the other methods for both phases.
Therefore, the modified linear model was chosen for further
Having chosen the modified linear model as the superior
model for flow estimation, different frequency ranges were
investigated toimproveitsperformance.Fig.2depicts themean
and standard deviation of the overall errors (averaged over all
flow rates) for estimating flow over the 11 frequency ranges.
The results showed that the average errors of the method in the
frequency ranges of 75–600 Hz, 150–1000 Hz, and 150–600
Hz were better than those of the other frequency ranges during
inspiration [Fig. 2(a)]. On the other hand, the performance
of the method was better for the frequency range 75–1000
Hz, 75–600 Hz and 75–450 Hz during expiration [Fig. 2(b)].
However, for the frequency range of 75–600 Hz the method
performed much more consistent at different flow rates with
smaller error at shallow and low flow rates. Also, during inspi-
ration the difference between the performances of the method
in the frequency ranges of 150–1000 Hz and 75–600 Hz was
negligible. Therefore, for the sake of consistency the frequency
range of 75–600 Hz was chosen for bandpass filtering tracheal
sound signal during both inspiration and expiration phases
The next step was investigating the effect of changing the
segment size for calculating entropy on the estimated flow. The
results of different segment sizes of 5, 10, 20, 35, 50, 75, and
100 ms on the performance of the modified linear model are
illustrated in Fig. 3.
By comparing theresults shown in Fig. 3, it can be concluded
that during inspiration the results are better for the 10-, 20-, and
MEAN ? STANDARD DEVIATION ?? ? ?? OF FLOW ESTIMATION ERROR (%) IN
THE RANGE OF [75 600] Hz WITH THE SEGMENT SIZE OF 50 ms
(%) of the modified linear model for different segment sizes in the frequency
range of 75–600 Hz. (a) Inspiration. (b) Expiration.
Mean?standard deviation?????ofthe overallflowestimationerror
achieved with the segment size of 75 ms. However, during the
inspiration and for the lowflow rates, whichare ofmore interest
thanthe high flowrates, theresults of 10 ms were more accurate
with smaller standard deviation than those of the 20- and 50-ms
segment sizes. Thus, the segment sizes of 10 and 75 ms were
chosen for inspiration and expiration phases, respectively.
Having chosen the frequency range of 75–600 Hz and seg-
ment size of 10 and 75 ms for inspiration and expiration phases,
respectively, the modified linear model (5) was used to estimate
for everysubjectand theoverallerror wasaveragedbetweenthe
subjects (Table II). A typical example of the estimated flow in
comparison with the actual flow is shown in Fig. 4. As it can be
observed, the estimated flow follows the actual flow very well
even when the flow rate changes.
The goal of this study was to investigate the relationship be-
tween flow and entropy of tracheal sound signals and choose
the best model representing such relationship to estimate flow.
The main objective was to find a robust flow estimation method
that does not require more than one breath for calibration and
can adjust itself to the flow variability. The estimated flow was
666IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 4, APRIL 2006
MEAN ? STANDARD DEVIATION ?? ? ?? OF FLOW ESTIMATION ERROR (%) IN
THE RANGE OF [75 600] Hz WITH THE SEGMENT SIZE OF 10 AND 75 ms FOR
INSPIRATORY AND EXPIRATORY PHASES, RESPECTIVELY
flow (dashed black line), (a) entire data and (b) focused on a few seconds of
data. The positive flow is inspiration.
Typical example of the actual flow (solid gray line) and the estimated
compared with the actual flow recorded by a mouth-piece pneu-
As the results shown in Fig. 1 suggests, the modified linear
model was the superior model representing the relationship
between tracheal sounds entropy and flow. After finding the
best model in the frequency range of 300–1000 Hz, different
frequency ranges were considered. Investigating the tracheal
sounds it was found that most of the tracheal sounds energy
exists below 300 Hz, especially at shallow and low flow rates.
Thus, different frequency ranges, including the frequencies
below 300 Hz, were investigated to improve the performance
of the method at shallow and low flow rates. However, at fre-
quencies lower than 300 Hz, heart sounds interfere with breath
sounds and change the time and frequency characteristics of
the breath sounds.
In order to cancel the effect of heart sounds on tracheal
sounds, several methods of heart sounds reduction from res-
piratory sounds were examined and the method proposed in
 was chosen for more investigation. Although this method
preserves the frequency characteristics of tracheal sounds, it
changes the time characteristics of tracheal sounds in the ad-
jacent segments of the heart sounds-included segments. Thus,
a new method was sought in this study to cancel the effects
of heart sounds on the tracheal sounds entropy, which was
based on heart sounds localization and entropy interpolation.
Note that the heart sounds cancellation technique of this study
does not claim to cancel the effect of heart sounds on breath
sounds as it works only on the entropy of the tracheal sound.
We believe this is a more effective way to cancel the effect of
heart sounds on tracheal sound for flow estimation purpose than
applying any of the developed heart sounds cancellation/reduc-
tion methods in the past. The premise of this claim is the fact
that the entropy of tracheal sound is a smoother signal than the
original breath sound signal and interpolation on a smoother
signal would produce better results. Hence, this method is
expected to give better results for flow estimation than the heart
sounds cancellation methods using interpolation of the original
breath sound signal.
Comparing the results of different frequency ranges for band
pass filtering the tracheal sounds, the range of 75–600 Hz was
found to have less error. In a previous study of flow estimation
, out of the nine investigated frequency ranges, the range of
150–450 Hz was found to perform the best. It should be noted
that all of the nine investigated frequency bands in  were
chosen above 100 Hz due to the significant presence of heart
sounds in the frequencies below 100 Hz. Investigating the re-
sults presented in Fig. 2 shows while the 150–450 Hz range is
acceptable during inspiration in terms of flow estimation error,
it is not a good choice during expiration due to the noticeably
higher error compared to those of the other frequency ranges.
Furthermore, in the frequency range of 150–450 Hz the stan-
dard deviation of error at low and shallow flow rates were high
during both inspiratory and expiratory phases.
In all previous studies except  and , a constant flow
rate was considered. In , three different flow rates were in-
vestigated but only the total error over all flow rates was re-
ported. In this study, the performance of the method at five dif-
were investigated separately and the model was tuned such that
it performed most consistently with variable flow. In , four
different flow rates were considered. However, the method pro-
posed in this study results in a much smaller error especially in
shallow and low flow rates compared to those reported in 
without the need of calibrating the model for every flow rate.
YADOLLAHI AND MOUSSAVI: ROBUST METHOD FOR ESTIMATING RESPIRATORY FLOW USING TRACHEAL SOUNDS ENTROPY 667
None of the previous studies (except ) attempted to esti-
portance in studies of sleep apnea and/or any study that requires
flow estimation during sleep. In this study, we investigated the
minimum critical flow at which the tracheal sounds have some
detectable power. We used average power of tracheal sounds in
the frequency range of 75–600 Hz and defined the minimum
was at least 3 dB higher than the background noise. Averaging
the results between the subjects, the minimum critical flow for
tracheal sounds was found to be 2.7
ml/s/kg for children and adults, respectively. Thus, in the case
of shallow flow rate the tracheal sounds with the values of flow
between the critical flow and 6 ml/s/kg were investigated.
Segment size was the last parameter investigated to improve
the performance of the method. Increasing the segment size
has a smoothing effect on the entropy calculation because the
entropy is calculated over a larger number of data samples;
becomes too small, the number of input samples contributing
to entropy calculation is smaller which may increase the error
in the PDF estimation. By comparing the results of different
segment sizes (Fig. 3), it is concluded that the best results
were achieved for the segment sizes of 10 and 75 ms during
inspiration and expiration, respectively. In addition, the results
suggest that the model is more robust to the changes in segment
size during the inspiratory phase compared to that of expiratory
phase. The reason for such difference is probably due to the
inherent differences in tracheal sound in each respiratory phase
as the resistance at the vocal cord changes with respiratory
phase. Furthermore, the mechanism of sound production in
each respiratory phase is different , .
Previous studies are heavily dependent on the calibration
process and they need to tune the parameters of the model at
In other words, the previously developed methods need to have
a copy of every flow rate to be estimated during the calibration
process. This is the major drawback of all previous methods
in flow estimation as it is not always possible to capture res-
piratory sounds at all of the possible flow rates, especially
when assessing young children, patients with neurological
impairments and/or patients in emergency conditions. The
method proposed in this study is novel in the sense that it only
requires one breath cycle for model calibration and then can
follow all variable flows with a high and consistent accuracy.
The average errors of the proposed modified linear model in
this study for inspiratory and expiratory phases were 8.3
2.8% and 9.6 2.8%, respectively which is smaller than those
of previous methods. Albeit the method presented in  has
smaller average errors; however, as mentioned before it was
based on a very sophisticated calibration process.
The subjects in this study were both children and adults and
also all the adult subjects were happened to females. However,
according the findings reported in  and , there is no sig-
nificant difference in flow-sound relationship as a result of age
or gender. Furthermore, it should be noted that the model for
flow estimation is being calibrated for each subject. Therefore,
it gets adjusted for any possible variability as a result of age,
1.7 ml/s/kg and 3.7 1.7
ating a data bank of flow and respiratory sounds sorted based on
body mass index, gender and age. As a future study we would
like to investigate whether the calibration process could be re-
placed by using the sorted information of the data bank.
This paper presents a novel and robust method for estimating
flow from tracheal breath sounds. The proposed method in this
paper eliminates the dependency of flow estimation methods on
calibration by using a new feature of the tracheal sounds that
follows the flow variation very closely. The method uses the
entropy of the band pass filtered tracheal sounds to estimate
flow. Three different models were investigated and the modi-
fied linear model was found to be superior to the power and
second-order models. The effects of different segment sizes and
frequency ranges were also investigated. The overall average
error of the proposed method was found to be 8.3
2.8%, for inspiratory and expiratory phases, respectively,
for calibration; it requires only one breath for calibration and
can estimate any flow rate with a reasonable accuracy (less than
The first author wishes to acknowledge the input and support
of A. Abadpour.
 S. Tarrant, R. Ellis, F. Flack, and W. Selley, “Comparative review of
techniques for recording respiratory events at rest and during degluti-
tion,” Dysphagia, vol. 12, pp. 24–38, 1997.
 R. Gilbert, J. Auchincloss, J. Brodsky, and W. Boden, “Changes in tidal
volume frequency and ventilation induced by their measurement,” J.
Appl. Physiol., vol. 33, pp. 252–254, 1972.
 J. Ashkanazi, P. Silverberg, R. Foster, A. Hyman, J. Milic-Emili, and J.
Kinney, “Effects of respiratory apparatus on breathing pattern,” J. Appl.
Physiol., vol. 48, pp. 577–580, 1980.
 J. Hirsch and B. Bishop, “Human breathing patterns on mouthpiece or
facemask during air, co2, or low o2,” J. Appl. Physiol., vol. 53, pp.
 C. Weissman, J. Ashkanazi, J. Milic-Emili, and J. Kinney, “Effect of
respiratory apparatus on respiration,” J. Appl. Physiol., vol. 57, pp.
 Z. Moussavi, M. Leopando, H. Pasterkamp, and G. Rempel, “Comput-
erized acoustical respiratory phase detection without airflow measure-
ment,” Inst. Elect. Eng. J. Med. Biolog Eng. comp., vol. 38, no. 2, pp.
 C. Lessard and W. Wong, “Correlation of constant flow rate with fre-
quency spectrum of respiratory sounds when measured at the trachea,”
IEEE Trans. Biomed. Eng., vol. BME-33, pp. 461–463, 1986.
and C. Gaultier, “Interaction between tracheal sounds and flow rare: a
comparison of some different flow evaluations from lung sounds,” IEEE
Trans. Biomed. Eng., vol. 37, no. 4, pp. 384–391, Apr. 1990.
 M. Mussell and Y. Miyamoto, “Comparison of normal respiratory
sounds recorded from the chest and trachea at various respiratory air
flow levels,” Front. Med. Biol. Eng., vol. 4, no. 2, pp. 73–85, 1992.
 N. Gavriely and D. Cugell, “Airflow effects on amplitude and spectral
content of normal breath sounds,” J. Appl. Physiol., vol. 80, pp. 5–13,
 Y. Yap and Z. Moussavi, “Acoustic airflow estimation from tracheal
sound power,” in Proc. IEEE Canadian Conf. Elec. Comp. Eng.
(CCECE), 2002, pp. 1073–1076.
668IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 4, APRIL 2006 Download full-text
 I. Hossain and Z. Moussavi, “Respiratory airflow estimation by acous-
tical means,”in Proc.2nd joint EMBS/BMES Conf., Houston, TX, 2002,
 M. Golabbakhsh, “Tracheal breath sound relationship with respiratory
flow: modeling, the effect of age and airflow estimation,” M.Sc. thesis,
Elec.Comp. Eng.Dept.,Univ. Manitoba,Winnipeg,MB, Canada,2004.
 C. Que, C. Kolmaga, L. Durand, S. Kelly, and P. Macklem,
methodology and preliminary results,” J. Appl. Physical., vol. 93, pp.
 A. Sovijarvi, L. Malmberg, G. Charbonneau, J. Vanderschoot, F. Dal-
and adventitious respiratory sounds,” Eur. Resp. Rev. J, pp. 591–596,
 V. Iyer, P. Ramamorthy, H. Fan, and Y. Ploysongsang, “Reduction of
Biomed. Eng., vol. BME-33, no. 12, pp. 1141–1148, Dec. 1986.
 M. Kompis and E. Russi, “Adaptive heart-noise reduction of respiratory
sounds recorded by a single microphone,” in Proc. 14th Ann. Int. Conf.
IEEE EMBS, 1992, pp. 691–692.
 L. Hadjileontiadis and S. Panas, “Adaptive reduction of heart sounds
from lung sounds using forth-order statistics,” IEEE Trans. Biomed.
Eng., vol. 44, no. 7, pp. 642–648, Jul. 1997.
 J. Gnitecki, Z. Moussavi, and H. Pasterkamp, “Recursive least square
adaptive noise cancellation filtering for heart sound in lung sounds
recording,” in Proc. IEEE Eng. Med. Biol. Soc., 2003, pp. 2416–2419.
 L. Hadjileontiadis and S. Panas, “A wavelet based reduction of heart
sound noise from lung sounds,” Int. J. Med. Informatics, vol. 52, pp.
 I. Hossain and Z. Moussavi, “An overview of heart-noise reduction of
lung sound using wavelet transform based filter,” in Proc. IEEE EMBS,
2003, pp. 458–461.
 M. Pourazad, Z. Moussavi, and G. Thomas, “Heart sound cancellation
fromlungsound recordingusingadaptivethreshold and2dinterpolation
in time-frequency domain,” in Proc. IEEE EMBS, 2003, pp. 2586–2589.
on multi-scale products and linear prediction,” in Proc. IEEE EMBS,
2004, pp. 3840–3843.
 A. Papoulis, Probability, Random Variables and Stochastic Pro-
cesses. New York: McGraw-Hill, 1991.
 A. Yadollahi and Z. Moussavi, “A robust method for heart sounds local-
ization using lung sounds entropy,” IEEE Trans. Biomed. Eng., vol. 52,
no. 3, pp. 497–502, Mar 2006.
 J. S. Chuah and Z. Moussavi, “Automated respiratory phase detection
Conf., Jul. 2000, pp. 228–231.
 M. Mahagnah and N. Gavriely, “Repeatability of measurements of
normal lung sounds,” Am J Respir. Crit. Care Med., vol. 149, pp.
 M. J. Mussel, Y. Nakazano, Y. Miyamoto, S. Okabe, and T. Takishima,
cipal component analysis,” Jpn. J. Physiol., vol. 40, pp. 713–721, 1990.
 V. Gross, A. Dittmar, T. Penzel, F. Schuttler, and P. V. Wichert, “The re-
lationship between normal lung sounds, age, and gender,” Am. J. Respir.
Critic. Care Med., vol. 162, no. 3, pp. 905–909, 2000.
 N. Gavriely, M. Nissan, A. Rubin, and D. W. Cugell, “Spectral charac-
teristics of chest wall breath sound in normal subjects,” Thorax, vol. 50,
pp. 1292–1300, 1995.
Azadeh Yadollahi (M’05) received both the B.Sc.
and M.Sc. degrees in electrical engineering from
Sharif University of Technology, Tehran, Iran,
in 2003 and 2005, respectively. She is currently
working towards the Ph.D. degree in the Department
of Electrical and Computer Engineering, University
Her current research interests include respiratory
sounds analysis and motor learning.
Mrs. Yadollahi is a member of the IEEE Engi-
neering in Medicine and Biology Society (EMBS)
and the Canadian Medical and Biological Engineering Society (CMBES).
Zahra M. K. Moussavi (M’98) received the B.Sc.
degree from University of Manitoba, Winnipeg, MB,
Canada, in 1997, all in Electrical Engineering.
She then joined the respiratory research group
of the Winnipeg Children’s Hospital and worked
as a Research Associate for one and half years.
In 1999, she joined the Biomedical Engineering
Department at Johns Hopkins University, Baltimore,
MD, and worked there as a Postdoctoral Fellow for one year. Following that,
she joined the University of Manitoba, Department of Electrical and Computer
Engineering asa faculty member,whereshe iscurrently an AssociateProfessor.
She is also an adjunct professor at the TR lab of Winnipeg. Her current research
includes respiratory and swallowing sound analysis, postural control and
balance, rehabilitation, and human motor learning.
Dr. Moussavi is a member of the IEEE Engineering in Medicine and Biology
Society (EMBS), the Canadian Medical and Biological Engineering Society
(CMBES), and the International Lung Sound (ILSA) Association. She is also
currently the EMBS Chapter Chair, Winnipeg Section.