I
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2, FEBRUARY 1991
I26
Acoustic Transmission of the Respiratory System Using Speech Stimulation Amon Cohen, Member, IEEE, and Albert0 D. Berstein
Abstract-Two methods for the analysis of the acoustic transmission of the respiratory system are presented. Continuous speech utterance is used as acoustic stimulation. The transmitted acoustic signal is recorded from various sites over the chest wall. The AR method analyzes the power spectral density function of the transmitted sound, which heavily depends on the microphone assembly and the utterance. The method was applied to a screening problem and was tested on a small database that consisted of 19 normal and five abnormal patients. Using the first five AR coefficients and the prediction error of an AR(1O) model, as discriminating features, the system screened all ahnormals. An ARMA method is suggested, which eliminates the dependence on microphone and utterance. In this method, the generalized least squares identification algorithm is used to estimate the ARMA transfer function of the respiratory system. The normal transfer function demonstrates a peak at the range of 130-250 Hz and sharp decrease in gain for higher frequencies. A pulmonary fibrotic patient demonstrated a peak at the same frequency range, a much higher gain in the high frequency range with an additional peak at about 700 Hz.
I . INTRODUCTION
A
USCULTATION is the most popular method for the diagnosis of pulmonary dysfunctions [ ll-[3]. The method consists of stimulating the chest cavity with an acoustical input and the monitoring (usually by means of a stethoscope) of the resultant acoustical signal in several locations on the chest wall. The monitored signal depends on the thorax’s acoustic transfer function. In pathologies such as emphyzema, pneumothorax, consolidation or accumulation of pleural fluid, the acoustical characteristics of the thorax change drastically from normal conditions. Three types of stimulating signals are used: a) Breath sounds-Acoustical signals generated in normal and pathological conditions by the airflow through the airway b) Percussion-The tapping of body structures to provide sounds. The sound source is usually made by striking the chest with a finger. c) Speech sounds-acoustical signal generated by the vocal cords and vocal tract. This sound is radiated from the lips to produce the speech. It is also transmitted through the airway and lungs to the chest wall. Auscultation is an attractive diagnosis method, mainly because of its simplicity. However, it provides only qualitative, subjective information and requires physician’s experience and proficiency. Recently, computerized methods for aiding in the analysis of breath sounds have been published [4]-[6]. These methods yield quantitative as well as qualitative information. Manuscript received December 28, 1989; revised May 16, 1990. The authors are with the Department of Electrical and Computer Engineering, Biomedical Engineering Program, Ben Gurion University, BeerSheva, Israel. IEEE Log Number 9041295.
Breath sounds analysis suffer additional drawback due to the fact that the acoustical stimulation, generated along the bronchial tree [7], is not well defined, is uncontrollable, and its exact source location is unknown. It is well known that the analysis of breath sounds depends on the flow rate [8].When comparative and more accurate studies are to be made the air flow has to be measured yielding system complexity. Using speech sounds as the stimulating signal provides several advantages. The speech source is controllable-its intensity and spectral content can be varied (within physiological limitations). Its source location is known and defined. Speech stimulation is almost independent of air flow. Spectrographic studies show [9] that the normal chest acts as a low-pass filter, attenuating all higher formants [lo] of the speech. The speech heard over the chest wall is thus completely distorted and meaningless. When a pathology, such as consolidation of the lung, is present the acoustical transmission of high frequencies increases and the monitored speech becomes less distorted. Speech signals are currently used, to some extent, in the clinic. This type of diagnostic test is, however, only qualitative, highly subjective and requires physician’s proficiency. Several qualitative diagnostic sounds have been defined: a) Normal voice sounds-indistinct, muffled, and only partially intelligible voice sounds heard over most of the normal chest. b) Bronchophony-Distinct, clear, and loud voice sounds heard over areas of the lung in which the normal alveoli are filled with fluid or replaced by solid tissue. Bronchophony is usually tested by asking the patient to utter “ninety nine.” c) Whispered Pectoriloquy-The presence of a clear, distinct intelligible whispered sound over areas of the lung in which the normal alveoli are filled with fluid or replaced by solid tissue. The patient is asked to whisper the utterance “ninety nine” or “one, two, three”. Whispering voice does not include the vibration of vocal cords and does not contain the low frequency formants. Over the normal chest, the whispered sound is not heard at all. d) Egophony-The presence of loud, nasal voice sound heard above a pleural effusion and rarely over a consolidated lung. To check for egophony the patient is asked to utter the utterance lit, When egophony is present the utterance is heard over the chest wall as la/. It is therefore sometimes known as “til to /a/ egophony.” The goal of this paper is to present algorithms for the objective, quantitative analysis of the chest cavity using speech signals. A multichannel spectral analysis of the chest speech is presented. The acoustical transfer function of the thorax is estimated, using parameter identification methods, and automatic classification algorithms are applied to the chest speech signal.
0018-9294/91/0200-0126$01 .OO 0 1991 IEEE
COHEN A N D BERSTEIN: ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
11. EXPERIMENTAL METHODS The experimental setup was designed to allow the recording of the speech signal in various locations on the chest wall and above the vocal cords. Above the vocal cords (tracheal site) the speech signal suffers almost no attenuation. The signal recorded at the tracheal site is thus used as a reference or as the input signal stimulating the chest cavity. In the current system five channels are used, which are depicted in Fig. 1. The reference signal was recorded at the trachea. The right and left apical (RA, LA) locations are at the third intercostal space facing the upper lobes of the lungs. The right and left basilar (RB, LB) locations are at the eighth intercostal space, facing the lower lobes of the lungs. This measurement setup is proposed for the detection and analysis of abnormalities in the chest cavity. Electret microphones were used (Sony ECM-Sops). A small bell-shaped assembly was constructed [ 111, [ 121 to house the microphone. The assembly was attached to the skin by means of a standard double-sided glued tape. No attempts were made to compensate for the microphone and the assembly’s transfer function. In the experiments where the chest transfer function was estimated the microphone and assembly transfer function are automatically compensated since the ratio of two measurements is calculated. In all other measurements only relative results are required for the discrimination between normal and various pathological cases. The methods described in this paper were checked on a limited number of normal and pathological subjects. Nineteen normal subjects, all males 25-40 years old, served as a normal group. The pathological group consisted of single pulmonary fibrosis (both lungs), pneumonia (right lung), pleural effusion, obstructive lung disease, and cancer (right lung) patients. The speech signals were low pass filtered at 1 kHz and sampled at 2.5 kHz sampling frequency with 12 b resolution. The speech used was the continuous utterance lil where 2.4 s per recording site were processed. Analysis frames of 80 ms with Hamming window were used for feature extraction and spectra estimation. Each frame consisted of 200 samples, and 30 frames per site per subject were analyzed.
111. AR SPECTRAL ANALYSIS First, the relative energy and spectral contents of the chest speech signals, at the various recording sites, were investigated. Note that the results have only relative meaning since the recorded signal is distorted by the microphone and the bell assembly. Let the signal at the ith recording site at time k be y , ( k ) . Define the mth frame energy E, ( m )by
E,(m) =
yz(k); k
E
mth Frame,
i
=
RA, RB, LA, LB
Fig. 1 , Recording sites: T-tracheal; LA, RA-left and right apical; LB, RB-left and right bahilar. TABLE I NORMALIZED SPEECH ENERGY FOR NORMAL AND FIBROTIC SUBJECTS u-TTERANCE-/i/) Normalized Energy [dB] Position Right Apex Left Apex Right Base Left Base
0.0
-6.6 ~
-7.5 10.0
(3) i=I
where G is the filter’s gain, p is the order of the all-pole system, and a,; i = 1, 2 , . * * , p are the AR or linear prediction coefficients (LPC). The signal may thus be represented by I’
y(k) = -
where M is the number of analysis frames (in our case M = 30). Table I depicts an example of the normalized speech energy measured from a normal and a fibrotic subject. The measurements show that the right-side energy is higher than that of the left side. This phenomena is known in the medical literature and is explained by the nonhomogeneity of the
0.0
-2.6 -3.2 -3.8
acoustical transfer of the chest 171. In the case of the fibrotic patient, tbe reference energy ERAwas higher than in the normal case (not shown in the table). The difference between right and left bases was 2 . 5 dB as compared to 0.6 dB in normals. This was due to the right fibrotic lung with improved high frequency transmission. The power spectral density (PSD) function of the signal was estimated by means of the well-known AR method [ 3 ] . We assume the signal { y ( k ) } to be the output of a hypothetical AR( p ) all poles filter, H ( z ) , driven by a zero mean and unit variance white noise sequence { ~ ( k ) } The . AR filter is given by
I
and the normalized speech energy, referred to the RA site by
Fibrosis
Normal
,=I
a,y(k
-
i)
+ Gw(k).
(4)
The LPC and the gain G can efficiently be estimated 131 from a finite signal sequence { y ( k ) ; k = 1, 2 , . . . , N } using the Durbin algorithm. An estimation of the PSD function S ( w ) of the signal is given in terms of the LPC’s by
I
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2. FEBRUARY 1991
128
PA
Position
Position
PSD
RA p o s i t i o n
I
freq.(hz)
Fig. 2 . Time and frequency presentations of the utterance iil in the various recording sites-normal subject with pitch of 135 Hz. (Reference and site signals not the same scale.) Upper trace: signal. 0-80 ms. Lower trace: PSD, 0-1.25 kHz.
Fig. 3 . Time and frequency presentations of the utterance iil in the various recording sites-subject with fibrosis (pitch of 230 Hz). (Traces same as in Fig. 2 . )
where r(i) is the signal’s autocorrelation sequence and
where Q ( z ) describes the speech generation mechanism, R(z) is the microphone and microphone assembly transfer function, and H ( z ) is the required chest transfer function. T is the sampling interval (in our case T = 0.4 ms). The main drawback of the AR method is its dependence on the quantity Q(z)R(z).Q(z) may vary drastically between subjects or even in the same subject over time. R(z) may be estimated. The estimation, however, is difficult and often inaccurate. The estimated PSD depends therefore on both the subject’s pitch and the microphone used in the measurement.
P-1
p(i) =
c a,,a,,+,
,,
=0
a.
= 1;
i
= 0, 1 ,
2,
...
,p.
(6)
The AR model order was first chosen as p = 20. This relatively large order was determined using the AIC criterion [3]. The PSD estimated by this order depicts the fine details due to the pitch frequency. Figs. 2 and 3 depict the AR estimated PSD of the utterances /i/ from a normal subject and from a patient with fibrosis. In Fig. 2, the reference signal, recorded over the trachea, shows a typical speech signal attenuated by the 1 kHz antialiasing filter. The pitch frequency of the subject is 135 Hz, shown by the large repetitive peaks in the PSD function. The recordings at the LA and RA and especially LB and RB sites demonstrate the normal low pass filtering effects of the chest. The signals at the LB and RB sites are distorted sinewaves since only the first few “harmonics” are transmitted. Fig. 3 depicts the increase of high-frequency transmission of the fibrotic case. Note that the fibrotic subject had a higher pitch of about 230 Hz. The utterance could be heard intelligibly over the chest bronchophony. Although the data in Figs. 2 and 3 cannot directly be compared due to the different input signals, it clearly demonstrate that the PSD of the speech signal recorded over the chest contains information about the acoustical transmission of the chest. The PSD of the signal is given by [ 101
IV. ARMAIDENTIFICATION In order to overcome the main drawback of the AR analysis method, it is required to estimate H ( z ) directly from the given signals. Consider the multichannel model described in Fig. 4. Here we assume that the reference signal X serves as a stimulating input to four linear systems representing the acoustical transfer function between trachea and the LA, RA, LB, and RB sites. We shall assume a pole-zero model (auto regressive moving average ARMA ( p , q ) model). U
I =
I
~
I29
COHEN A N D BERSTEIN, ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
REFERENCE SIGNL
b, ;'+b2i2+b,i3+b4z-4
H(z-l)
=
l+a1z'+a,r2+a,i3+a4z-4
Fig. 5 . ARMA (4,4) model of chest transfer function. DE
t
Fig. 4. Multichannel model of the chest
with noisy output sequences, { z , ( k ) } , given by
z,(Q
+ n,(4
= Y,(k) P
= -
c ajl'y,(k
I =
--
I
-
i)
+
Y
bj"x,(k I =
I
- i)
+ n,(k)
(9)
in (9) the coefficients U ; ' ' ; bj" are the ARMA ( p , q ) coefficients representing the acoustical transfer function of t h e j th recording site. Given the finite reference sequence { x ( k ) } (which is estimated by the signal recorded at the trachea) and the noisy output sequences { z , ( k ) } , k = 1, 2 , . . N , the ARMA ( p , q) coefficients may be estimated by means of the generalized least squares (GLS) algorithm [ 131. Preliminary experiments have demonstrated that the order of p = q = 4 is sufficient. Each one of the four channels was thus represented by an ARMA (4, 4) model given in Fig. 5 . The frequency transfer function of thejth site of the chest is given by
Fig. 6 . Chest frequency transfer function-normal (solid line, subject AF; broken line. subject BC).
where T i s the sampling interval ( T = 0.4 ms). Note that in (10) the chest transfer function is directly given. The speech and microphone effects have been eliminated in this method since the ratio between input and output signals was used. The frequency transfer functions of two normal subjects are shown in Fig. 6. Since the right and left sides functions were very similar in shape, only the RA and RB positions are shown. Fig. 7 depicts the chest frequency transfer function of a fibrotic subject. The estimated poles and zeroes maps of one of the normal subjects and the fibrotic subject are shown in Fig. 8. The identified system is stable with all its poles located inside the unit circle. It is, however, not minimum phase, with some of the zeroes located outside the unit circle. Zeroes with abso-
lute value larger than 1.5 were not plotted; these zeroes do not contribute much to the shape of the function. The transfer function of the normal chest, as depicted in Fig. 6, contains a peak at the frequencies of about 250-320 Hz (RA) and 130-160 Hz (RB). The RB transfer contains also a zero at 590-680 Hz. The peaks in the frequency transfer function correspond to the two conjugate dominant poles at about 22" and magnitude of 0.94 while the dip corresponds to the two zeroes at 88" and magnitude of 0.99 [Fig. 8(a)]. The fibrotic subject demonstrates two peaks at about 3 18 and 682 Hz (RA) and two almost indistinguishable peaks at 236 and 590 Hz (RB). The latter correspond to the poles [Fig. 8(b)] at 32" and a magnitude of 0.83 and at 81" and a magnitude of 0.73. These poles are not as close to the unit circle as those of the normal chest, hence the less distinct peaks in the frequency plot.
T
L
I10
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38. NO. 2, FEBRUARY 1991
7 FREQUENCY
( Hz )
DE
-=
t
I
The AR analysis is much more effective, in terms of computations, than the ARMA method. To be clinically useful, the analysis will have to be implemented on a small, inexpensive, hardware preferably in “near” real time. It was therefore decided to try the AR inferior method as a base for the automatic classification. A set of six features was arbitrarily selected to represent the data. The features were the first five linear prediction coefficients, and the normalized prediction error [3] of an AR(10) model of the RB site. No attempt was made to find an optimal set of features [3]. Note that an AR order of p = IO is used here since the fine pitch peaks are of no interest. A recognition system was developed to recognize normal RB PSD and reject abnormals. The training set consisted of ten healthy male subjects (25-40 years old). From each ith, 80 ms segment of recorded RB signal of the j t h training subject, a feature vector p:’) was estimated.
P?
-
:
:
!
:
500
!
:
! , o w : FREQUENCY ( H z )
Fig. 7 . Chest frequency transfer function-fibrosis
I
“
(a
= (al, a 2 . a3. a4, as. eld’
(1 1 )
where a,, k = 1 , . * , 5 is an LPC coefficient of an AR(1O) model and e,o is the normalized prediction error of that model. The mean “healthy” feature vector was estimated by averaging (1 1) over all the training data
2 PLANE
where J is the total number of subjects in the training set (= 10) and I is the number of 80 ms segments for each subject (=30). The “healthy” covariance matrix was estimated by
1
I
“ (L)
Z PLANE
B
Fig. 8 . Poles-zeroes maps, site RB. (a) Normal subject (AF). (b) fibrosis.
V . AUTOMATIC CLASSIFICATION Automatic classification of signals have been successfully applied to chest diagnosis by breath sounds [4]. Similar methods are suggested in order to automatically classify normal and pathological subjects, using speech stimulated chest features.
The estimated mean feature vector (12) and the inverse of the covariance matrix (13) were used as a template of the healthy chest. In order to test the recognition system, a test set of 14 subjects was used. Nine subjects were healthy males (not included in the training group). One of the healthy subjects was instructed to say the testing utterance at different pitch frequencies and intensities in order to include the change of parameters not related to physiological state of the chest. The other five subjects had various pulmonary dysfunctions-pleural effusion, fibrosis, pneumonia, obstructive lung disease, and severe lung tumor. The mahalanobis distance [3] was used for the recognition. The distance d,,Hbetween the signal E‘‘) acquired from the RB site of a subject under test and the template of healthy subjects is given by
In an automatic recognition system one would estimate a template for each one of the pathologies of interest, calculate the distance of the signal under test, to each one of these templates, and apply the minimum distance classification method [3] to determine the classification decision. This procedure requires sufficient training data for each one of the pathologies. Since such a database is not yet available, the procedure was checked for screening rather than recognition application. In screening one has to classify the data into one of two classes: normal (healthy) and abnormal. The problem may be posed as follows: given a template of the “healthy” random process, accept or reject the hypothesis that the given sample function,
I
COHEN A N D BERSTEIN: ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
Distance +
Fig. 9. Distance distribution (RB). Distances are calculated from the tested subject to the “healthy” template.
acquired from the subject under test, belongs to the “healthy” process. This hypothesis testing may be performed using the distance given by(14). We accept the hypothesis, and classify the tested subject as “healthy” if dr,His less than some empirical threshold T . If the distribution of the distance of “healthy” subjects from the healthy template was known, the threshold T that yielded some allowed mean false accept error could be theoretically determined. Fig. 9 depicts results of distance measurements taken from the limited test group. The distances are between the “healthy” template and signals taken from the test subjects. The right side of Fig. 9 (Normal) is a histogram of the distribution of dx,H given x is “healthy”. The relatively large number of “healthy” segments in the test group allowed the formation of the histogram. The distances, dx,Hgiven x belongs to one of the pathologies, are marked by arrows representing the mean over the few segments available. The data shown in Fig. 9 suggest that the normal (healthy) and abnormal classes are well separable by the suggested method. A threshold level of 70 may be used for screening applications.
VI. DISCUSSION AND CONCLUSIONS Two methods for the acoustical analysis of the chest have been suggested. Both methods use the speech utterance as the stimulating function. This type of stimulation has several advantages: it is controllable, repeatable, and its source location is known and defined. One of the disadvantages is the fact that it cannot be easily used with noncooperative subjects such as infants (the transfer function method may however be applied, using the infant’s cry [ 141 or an external acoustical source [ 151 as stimulus). Since the chest acts as a low-pass filter, it is advisable to choose a test utterance which contains sufficient power in the lower region of the frequency range. A continuous, almost stationary, vowel utterance (such as the continuous vowel /i/) is used in the system. The vowels having low frequency energies are /U/ and /i/. The vowel /U/ has a first format (Fl-the first peak in the spectrum) range of 200-400 Hz and a second format ( F 2 ) range of about 500-1000 Hz. The vowel ti/ has about the same F1 range and a much higher F 2 range of 2-4 kHz. Both these continuous vowels are suitable for stimulating the chest. For qualitative analysis, speech utterances like “ninety-nine’’ may be used. Using such utterances for a more detailed, quantitative, analysis will require sophisticated time warping [ 101 algorithms. The AR method estimates the power spectral density (PSD) function of the speech at the recording site of the chest. This
-
-7
131
PSD depends on the microphone assembly’s transfer function and the utterance. As seen in Figs. 2 and 3, it is heavily influenced by the pitch frequency (Fo), which may largely be changed from subject to subject. The estimated PSD may also depend on the time at which the signal is taken along the continuous utterance since the lungs are deflated during speech. The order of the AR estimator used in Figs. 2 and 3 was chosen a s p = 20, so that the detailed PSD, with its pitch peaks is displayed. A lower order AR estimator will provide a smoothed PSD estimate, reducing the dependence on the pitch. In the classification experiments a lower order of p = 10 was used with only the first five AR coefficients and the prediction error participating as features. Even with all its disadvantages the AR method separated well the abnormals from the normals, as demonstrated by the data of Figure 9. The data presented here is based on a very limited database that consisted of only 19 normals and five abnormal subjects. A comprehensive database is in the process of being established. Such a database will allow the estimation of the distance distribution and the optimal set of features to be used. The smoothed PSD has ony one peak, it is therefore logical to assume that an AR order o f p < 10 should suffice. The features used in this work were arbitrarily chosen. With a larger database one may apply optimization techniques [3] to establish a set of optimal features to be used in (1 1). The larger database will also allow the evaluation of the system as a recognition, rather than a screening system. The fact that the AR algorithm is simple motivated (in this work) the use of AR rather than ARMA method. The ARMA method used here (or other recursive or block methods availability in the literature) though more complex than the AR algorithm may also be implemented on modem microprocessors. The ARMA method provides the chest transfer function independent of the microphone assembly and the subject’s pitch. It thus yields information which is directly related to the problem under investigation. The ARMA ( p , q ) coefficients may be used as the features in ( I 1) to provide probably a better classification process. Since the goal of this work was to prove feasibility of the suggested methods, only the RB site was utilized. A clinical system will use multichannel data since various pathologies may be best distinguished by data taken from specific sites. In minimum distance classifiers, it is beneficial to work with a set of orthonormal gaussian features. The distance distributions (measured from the true process expectation) are then chisquare distributions with known degrees of freedom. It is possible to orthogonalize the LPC (or ARMA) features using a (not too simple) orthogonalization algorithm. The speech LPC features amplitude distributions may be considered gaussian. This, however, has not been shown for the type of utterances used here and for a speech that was transformed by the chest. It may thus very well be that a theoretically calculated distance distribution to be used for a theoretical determination of threshold levels will not be practical. It is, however, our intention to investigate it when a larger database is available. Several attempts have been made to model the acoustic impedance and transfer of the respiratory system, which includes the vocal tract, the trachea, the network of branching airways, the lung parenchyma, and the chest wall [ 161-[20]. The model by Wodicka et al. [ 181 predicts a peak in the transfer function at approximately 250 Hz, which nicely matches the data in Fig. 6 (RA position). The model, however, does not explain the shift of the peak to lower frequency of about 150
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2, FEBRUARY 1991
I32
Hz and 0 at higher frequencies depicted in Fig. 6 (RB position), nor does it explain the second peak at about 700 Hz in the fibrotic transfer function (Fig. 7). These may be explained by the interactions between the waves generated by different segments of the respiratory system [ 181. It is concluded that speech stimulated chest analysis systems have the potential of yielding important clinical information. The feasibility has been demonstrated (on a small size population) in this work. The ARMA method is especially attractive due to its independence of pitch and microphone. Additional results, using a larger database, are required in order to establish the clinical use of the suggested method.
REFERENCES [ 11 S. Lehrer, Understanding Lung Sounds.
Philadelphia, PA: W.B. Saunders, 1989. [2] G. Druger, The Chest: Its Signs and Sounds. Los Angeles, CA: Humetrics Corp., 1973. [3] A. Cohen, Biomedical Signal Processing. Boca Raton, FL: CRC Press, 1986. [4] A. Cohen and D. Landsberg, “Analysis and automatic classification of breath sounds,” IEEE Trans. Biomed. Eng., vol. BME31, pp. 585-590, 1984. 151 R. B. Urquhart, J. McGee, J. E. S . Macleod, S . W. Banham, and F. Moran, “The diagnostic value of pulmonary sounds: A preliminary study by computer aided analysis,” Comput. Biol. Med., vol. 11, pp. 129-139, 1981. [6] A. Berstein and A. Cohen, “Speech processing as a chest diagnosis assist device,” in Proc. IEEE 14th Conrsent. Elect. Electron. Eng., Tel-Aviv, Israel, 1985. 171 S. S. Kraman and D. Austrheim, “Comparison of lung sounds and transmitted sound amplitude in normal men,” Amer. Res. Respir. Dis., vol. 128, pp. 451-454, 1983. [SI T. R. Fenton, H. Pasterkamp, A. Tal, and V . Chernick, “Automated spectral characterization of wheezing in asthmatic children,” IEEE Trans. Biomed. Eng., vol. BME-32, pp. 50-55, 1985. (91 V . A. McKusik, J. T. Jenkins, and G. N. Webb, “The acoustic basis of the chest examination studies by means of sound spectrography,” Amer. Res. Tuberc. Pul. Dis., vol. 72, pp. 12-34, 1955. [ 101 L. R. Rabiner and R . W. Schafer, Digital Processing of Speech Signals. Englewood Cliffs, NJ: Prentice Hall, 1978. [ I l l C. K. Druzgalski, R. L. Donnenberg, and R. M. Campbell, “Techniques of recording respiratory sounds,’’ J . Clini. Eng., vol. 5 , pp. 321-330, 1980. [ 121 N . Sugumara and K. Ikegaya, “Characteristics of the air cavities of phono-cardiographic microphones, ” Med. Biol. Eng. Comput.. vol. 15, pp. 240-247, 1977. [13] M. S . Ahmed, “Fast GLS algorithms for parameter estimation,” Automatica, vol. 20, pp. 231-236, 1984. [14] A. Cohen and E. Zmora, “Automatic classification of infant’s hunger and pain cry,” in Proc. Int. Conf Dig. Sig. Proc., Florence, Italy, 1984, pp. 667-672.
[15] M. Miyakawa, K. Yamamoto, and T. Mikami, “Acoustic measurement of the respiratory system. An acoustic pneumograph,” Med. Biol. Eng., vol. 14, pp. 653-659. 1976. [16] J. J . Fredberg and A. Hoenig, “Mechanical response of the lungs at high frequencies,” J . Biomech. Eng., vol. 100, pp. 57-66, 1987. [17] K. Ishizaka, M. Matsudaira, and T. Kaneko, “Input acoustic impedance measurement of the subglottal system,” J . Acoust. Soc. Amer., vol. 60, pp. 190-197, 1976. [18] G. R. Wodicka, K. N . Stevens, H. L. Golub, E. G. Cravalho, and D. C. Shannon, “A model of acoustic transmission in the respiratory system,” IEEE Trans. Biomed. Eng., vol. BME-36, pp. 925-934, 1989. [19] R. W. Guelke and A. E. Bum, “Transmission line theory applied to sound wave propagation in tubes with compliant walls,” Acoustica, vol. 48, pp. 101-106, 1981. [20] W. L. Capper, “Acoustic input impedance measurements on elastic walled tubes and large airways in man,” Ph.D. Thesis, Dep. Biomed. Eng., Univ. Cape Town, South Africa, 1988.
Arnon Cohen (M’69-M’80) was born in Haifa, Israel, in 1938. He received the B.Sc. and M.Sc. degrees from the Technion-Israel Institute of Technology in 1964 and 1966, respectively, and the Ph.D. degree from CarnegieMellon University, Pittsburgh, PA, in 1969. Since 1972 he has been with the Department of Electrical and Computer Engineering and the Biomedical Engineering Program, Ben Gurion University, Beer-Sheva, Israel where he is a Professor of Electrical and Biomedical Engineering. He was with the Department of Electrical Engineering and Bioengineering, University of Connecticut, Storrs, from 1967 to 1972, the Colorado State University, Fort Collins, during 1976-1977, and the Department of Biomedical Engineering. University of Cape Town Medical School, Cape Town, South Africa, during 1989-1990. His research interests are in signal processing, mainly with biomedical applications. He is the author of the book Biomedical Signal Processing (Boca Raton, FL: CRC).
Albert0 D. Berstein was born in Argentina in 1954. He received the B.Sc. and M.Sc. degrees in electrical and computer engineering in 1981 and 1985, respectively, from the Ben Gurion University, Beer-Sheva, Israel. Since 1985 he has been engaged in Speech Enhancement and Speaker Recognition research work. He is now with the DSP Group, Israel. He is also a graduate student in the Electrical and Computer Engineering Department, Ben Gurion University, working towards the Ph.D. degree.
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2, FEBRUARY 1991
I26
Acoustic Transmission of the Respiratory System Using Speech Stimulation Amon Cohen, Member, IEEE, and Albert0 D. Berstein
Abstract-Two methods for the analysis of the acoustic transmission of the respiratory system are presented. Continuous speech utterance is used as acoustic stimulation. The transmitted acoustic signal is recorded from various sites over the chest wall. The AR method analyzes the power spectral density function of the transmitted sound, which heavily depends on the microphone assembly and the utterance. The method was applied to a screening problem and was tested on a small database that consisted of 19 normal and five abnormal patients. Using the first five AR coefficients and the prediction error of an AR(1O) model, as discriminating features, the system screened all ahnormals. An ARMA method is suggested, which eliminates the dependence on microphone and utterance. In this method, the generalized least squares identification algorithm is used to estimate the ARMA transfer function of the respiratory system. The normal transfer function demonstrates a peak at the range of 130-250 Hz and sharp decrease in gain for higher frequencies. A pulmonary fibrotic patient demonstrated a peak at the same frequency range, a much higher gain in the high frequency range with an additional peak at about 700 Hz.
I . INTRODUCTION
A
USCULTATION is the most popular method for the diagnosis of pulmonary dysfunctions [ ll-[3]. The method consists of stimulating the chest cavity with an acoustical input and the monitoring (usually by means of a stethoscope) of the resultant acoustical signal in several locations on the chest wall. The monitored signal depends on the thorax’s acoustic transfer function. In pathologies such as emphyzema, pneumothorax, consolidation or accumulation of pleural fluid, the acoustical characteristics of the thorax change drastically from normal conditions. Three types of stimulating signals are used: a) Breath sounds-Acoustical signals generated in normal and pathological conditions by the airflow through the airway b) Percussion-The tapping of body structures to provide sounds. The sound source is usually made by striking the chest with a finger. c) Speech sounds-acoustical signal generated by the vocal cords and vocal tract. This sound is radiated from the lips to produce the speech. It is also transmitted through the airway and lungs to the chest wall. Auscultation is an attractive diagnosis method, mainly because of its simplicity. However, it provides only qualitative, subjective information and requires physician’s experience and proficiency. Recently, computerized methods for aiding in the analysis of breath sounds have been published [4]-[6]. These methods yield quantitative as well as qualitative information. Manuscript received December 28, 1989; revised May 16, 1990. The authors are with the Department of Electrical and Computer Engineering, Biomedical Engineering Program, Ben Gurion University, BeerSheva, Israel. IEEE Log Number 9041295.
Breath sounds analysis suffer additional drawback due to the fact that the acoustical stimulation, generated along the bronchial tree [7], is not well defined, is uncontrollable, and its exact source location is unknown. It is well known that the analysis of breath sounds depends on the flow rate [8].When comparative and more accurate studies are to be made the air flow has to be measured yielding system complexity. Using speech sounds as the stimulating signal provides several advantages. The speech source is controllable-its intensity and spectral content can be varied (within physiological limitations). Its source location is known and defined. Speech stimulation is almost independent of air flow. Spectrographic studies show [9] that the normal chest acts as a low-pass filter, attenuating all higher formants [lo] of the speech. The speech heard over the chest wall is thus completely distorted and meaningless. When a pathology, such as consolidation of the lung, is present the acoustical transmission of high frequencies increases and the monitored speech becomes less distorted. Speech signals are currently used, to some extent, in the clinic. This type of diagnostic test is, however, only qualitative, highly subjective and requires physician’s proficiency. Several qualitative diagnostic sounds have been defined: a) Normal voice sounds-indistinct, muffled, and only partially intelligible voice sounds heard over most of the normal chest. b) Bronchophony-Distinct, clear, and loud voice sounds heard over areas of the lung in which the normal alveoli are filled with fluid or replaced by solid tissue. Bronchophony is usually tested by asking the patient to utter “ninety nine.” c) Whispered Pectoriloquy-The presence of a clear, distinct intelligible whispered sound over areas of the lung in which the normal alveoli are filled with fluid or replaced by solid tissue. The patient is asked to whisper the utterance “ninety nine” or “one, two, three”. Whispering voice does not include the vibration of vocal cords and does not contain the low frequency formants. Over the normal chest, the whispered sound is not heard at all. d) Egophony-The presence of loud, nasal voice sound heard above a pleural effusion and rarely over a consolidated lung. To check for egophony the patient is asked to utter the utterance lit, When egophony is present the utterance is heard over the chest wall as la/. It is therefore sometimes known as “til to /a/ egophony.” The goal of this paper is to present algorithms for the objective, quantitative analysis of the chest cavity using speech signals. A multichannel spectral analysis of the chest speech is presented. The acoustical transfer function of the thorax is estimated, using parameter identification methods, and automatic classification algorithms are applied to the chest speech signal.
0018-9294/91/0200-0126$01 .OO 0 1991 IEEE
COHEN A N D BERSTEIN: ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
11. EXPERIMENTAL METHODS The experimental setup was designed to allow the recording of the speech signal in various locations on the chest wall and above the vocal cords. Above the vocal cords (tracheal site) the speech signal suffers almost no attenuation. The signal recorded at the tracheal site is thus used as a reference or as the input signal stimulating the chest cavity. In the current system five channels are used, which are depicted in Fig. 1. The reference signal was recorded at the trachea. The right and left apical (RA, LA) locations are at the third intercostal space facing the upper lobes of the lungs. The right and left basilar (RB, LB) locations are at the eighth intercostal space, facing the lower lobes of the lungs. This measurement setup is proposed for the detection and analysis of abnormalities in the chest cavity. Electret microphones were used (Sony ECM-Sops). A small bell-shaped assembly was constructed [ 111, [ 121 to house the microphone. The assembly was attached to the skin by means of a standard double-sided glued tape. No attempts were made to compensate for the microphone and the assembly’s transfer function. In the experiments where the chest transfer function was estimated the microphone and assembly transfer function are automatically compensated since the ratio of two measurements is calculated. In all other measurements only relative results are required for the discrimination between normal and various pathological cases. The methods described in this paper were checked on a limited number of normal and pathological subjects. Nineteen normal subjects, all males 25-40 years old, served as a normal group. The pathological group consisted of single pulmonary fibrosis (both lungs), pneumonia (right lung), pleural effusion, obstructive lung disease, and cancer (right lung) patients. The speech signals were low pass filtered at 1 kHz and sampled at 2.5 kHz sampling frequency with 12 b resolution. The speech used was the continuous utterance lil where 2.4 s per recording site were processed. Analysis frames of 80 ms with Hamming window were used for feature extraction and spectra estimation. Each frame consisted of 200 samples, and 30 frames per site per subject were analyzed.
111. AR SPECTRAL ANALYSIS First, the relative energy and spectral contents of the chest speech signals, at the various recording sites, were investigated. Note that the results have only relative meaning since the recorded signal is distorted by the microphone and the bell assembly. Let the signal at the ith recording site at time k be y , ( k ) . Define the mth frame energy E, ( m )by
E,(m) =
yz(k); k
E
mth Frame,
i
=
RA, RB, LA, LB
Fig. 1 , Recording sites: T-tracheal; LA, RA-left and right apical; LB, RB-left and right bahilar. TABLE I NORMALIZED SPEECH ENERGY FOR NORMAL AND FIBROTIC SUBJECTS u-TTERANCE-/i/) Normalized Energy [dB] Position Right Apex Left Apex Right Base Left Base
0.0
-6.6 ~
-7.5 10.0
(3) i=I
where G is the filter’s gain, p is the order of the all-pole system, and a,; i = 1, 2 , . * * , p are the AR or linear prediction coefficients (LPC). The signal may thus be represented by I’
y(k) = -
where M is the number of analysis frames (in our case M = 30). Table I depicts an example of the normalized speech energy measured from a normal and a fibrotic subject. The measurements show that the right-side energy is higher than that of the left side. This phenomena is known in the medical literature and is explained by the nonhomogeneity of the
0.0
-2.6 -3.2 -3.8
acoustical transfer of the chest 171. In the case of the fibrotic patient, tbe reference energy ERAwas higher than in the normal case (not shown in the table). The difference between right and left bases was 2 . 5 dB as compared to 0.6 dB in normals. This was due to the right fibrotic lung with improved high frequency transmission. The power spectral density (PSD) function of the signal was estimated by means of the well-known AR method [ 3 ] . We assume the signal { y ( k ) } to be the output of a hypothetical AR( p ) all poles filter, H ( z ) , driven by a zero mean and unit variance white noise sequence { ~ ( k ) } The . AR filter is given by
I
and the normalized speech energy, referred to the RA site by
Fibrosis
Normal
,=I
a,y(k
-
i)
+ Gw(k).
(4)
The LPC and the gain G can efficiently be estimated 131 from a finite signal sequence { y ( k ) ; k = 1, 2 , . . . , N } using the Durbin algorithm. An estimation of the PSD function S ( w ) of the signal is given in terms of the LPC’s by
I
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2. FEBRUARY 1991
128
PA
Position
Position
PSD
RA p o s i t i o n
I
freq.(hz)
Fig. 2 . Time and frequency presentations of the utterance iil in the various recording sites-normal subject with pitch of 135 Hz. (Reference and site signals not the same scale.) Upper trace: signal. 0-80 ms. Lower trace: PSD, 0-1.25 kHz.
Fig. 3 . Time and frequency presentations of the utterance iil in the various recording sites-subject with fibrosis (pitch of 230 Hz). (Traces same as in Fig. 2 . )
where r(i) is the signal’s autocorrelation sequence and
where Q ( z ) describes the speech generation mechanism, R(z) is the microphone and microphone assembly transfer function, and H ( z ) is the required chest transfer function. T is the sampling interval (in our case T = 0.4 ms). The main drawback of the AR method is its dependence on the quantity Q(z)R(z).Q(z) may vary drastically between subjects or even in the same subject over time. R(z) may be estimated. The estimation, however, is difficult and often inaccurate. The estimated PSD depends therefore on both the subject’s pitch and the microphone used in the measurement.
P-1
p(i) =
c a,,a,,+,
,,
=0
a.
= 1;
i
= 0, 1 ,
2,
...
,p.
(6)
The AR model order was first chosen as p = 20. This relatively large order was determined using the AIC criterion [3]. The PSD estimated by this order depicts the fine details due to the pitch frequency. Figs. 2 and 3 depict the AR estimated PSD of the utterances /i/ from a normal subject and from a patient with fibrosis. In Fig. 2, the reference signal, recorded over the trachea, shows a typical speech signal attenuated by the 1 kHz antialiasing filter. The pitch frequency of the subject is 135 Hz, shown by the large repetitive peaks in the PSD function. The recordings at the LA and RA and especially LB and RB sites demonstrate the normal low pass filtering effects of the chest. The signals at the LB and RB sites are distorted sinewaves since only the first few “harmonics” are transmitted. Fig. 3 depicts the increase of high-frequency transmission of the fibrotic case. Note that the fibrotic subject had a higher pitch of about 230 Hz. The utterance could be heard intelligibly over the chest bronchophony. Although the data in Figs. 2 and 3 cannot directly be compared due to the different input signals, it clearly demonstrate that the PSD of the speech signal recorded over the chest contains information about the acoustical transmission of the chest. The PSD of the signal is given by [ 101
IV. ARMAIDENTIFICATION In order to overcome the main drawback of the AR analysis method, it is required to estimate H ( z ) directly from the given signals. Consider the multichannel model described in Fig. 4. Here we assume that the reference signal X serves as a stimulating input to four linear systems representing the acoustical transfer function between trachea and the LA, RA, LB, and RB sites. We shall assume a pole-zero model (auto regressive moving average ARMA ( p , q ) model). U
I =
I
~
I29
COHEN A N D BERSTEIN, ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
REFERENCE SIGNL
b, ;'+b2i2+b,i3+b4z-4
H(z-l)
=
l+a1z'+a,r2+a,i3+a4z-4
Fig. 5 . ARMA (4,4) model of chest transfer function. DE
t
Fig. 4. Multichannel model of the chest
with noisy output sequences, { z , ( k ) } , given by
z,(Q
+ n,(4
= Y,(k) P
= -
c ajl'y,(k
I =
--
I
-
i)
+
Y
bj"x,(k I =
I
- i)
+ n,(k)
(9)
in (9) the coefficients U ; ' ' ; bj" are the ARMA ( p , q ) coefficients representing the acoustical transfer function of t h e j th recording site. Given the finite reference sequence { x ( k ) } (which is estimated by the signal recorded at the trachea) and the noisy output sequences { z , ( k ) } , k = 1, 2 , . . N , the ARMA ( p , q) coefficients may be estimated by means of the generalized least squares (GLS) algorithm [ 131. Preliminary experiments have demonstrated that the order of p = q = 4 is sufficient. Each one of the four channels was thus represented by an ARMA (4, 4) model given in Fig. 5 . The frequency transfer function of thejth site of the chest is given by
Fig. 6 . Chest frequency transfer function-normal (solid line, subject AF; broken line. subject BC).
where T i s the sampling interval ( T = 0.4 ms). Note that in (10) the chest transfer function is directly given. The speech and microphone effects have been eliminated in this method since the ratio between input and output signals was used. The frequency transfer functions of two normal subjects are shown in Fig. 6. Since the right and left sides functions were very similar in shape, only the RA and RB positions are shown. Fig. 7 depicts the chest frequency transfer function of a fibrotic subject. The estimated poles and zeroes maps of one of the normal subjects and the fibrotic subject are shown in Fig. 8. The identified system is stable with all its poles located inside the unit circle. It is, however, not minimum phase, with some of the zeroes located outside the unit circle. Zeroes with abso-
lute value larger than 1.5 were not plotted; these zeroes do not contribute much to the shape of the function. The transfer function of the normal chest, as depicted in Fig. 6, contains a peak at the frequencies of about 250-320 Hz (RA) and 130-160 Hz (RB). The RB transfer contains also a zero at 590-680 Hz. The peaks in the frequency transfer function correspond to the two conjugate dominant poles at about 22" and magnitude of 0.94 while the dip corresponds to the two zeroes at 88" and magnitude of 0.99 [Fig. 8(a)]. The fibrotic subject demonstrates two peaks at about 3 18 and 682 Hz (RA) and two almost indistinguishable peaks at 236 and 590 Hz (RB). The latter correspond to the poles [Fig. 8(b)] at 32" and a magnitude of 0.83 and at 81" and a magnitude of 0.73. These poles are not as close to the unit circle as those of the normal chest, hence the less distinct peaks in the frequency plot.
T
L
I10
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38. NO. 2, FEBRUARY 1991
7 FREQUENCY
( Hz )
DE
-=
t
I
The AR analysis is much more effective, in terms of computations, than the ARMA method. To be clinically useful, the analysis will have to be implemented on a small, inexpensive, hardware preferably in “near” real time. It was therefore decided to try the AR inferior method as a base for the automatic classification. A set of six features was arbitrarily selected to represent the data. The features were the first five linear prediction coefficients, and the normalized prediction error [3] of an AR(10) model of the RB site. No attempt was made to find an optimal set of features [3]. Note that an AR order of p = IO is used here since the fine pitch peaks are of no interest. A recognition system was developed to recognize normal RB PSD and reject abnormals. The training set consisted of ten healthy male subjects (25-40 years old). From each ith, 80 ms segment of recorded RB signal of the j t h training subject, a feature vector p:’) was estimated.
P?
-
:
:
!
:
500
!
:
! , o w : FREQUENCY ( H z )
Fig. 7 . Chest frequency transfer function-fibrosis
I
“
(a
= (al, a 2 . a3. a4, as. eld’
(1 1 )
where a,, k = 1 , . * , 5 is an LPC coefficient of an AR(1O) model and e,o is the normalized prediction error of that model. The mean “healthy” feature vector was estimated by averaging (1 1) over all the training data
2 PLANE
where J is the total number of subjects in the training set (= 10) and I is the number of 80 ms segments for each subject (=30). The “healthy” covariance matrix was estimated by
1
I
“ (L)
Z PLANE
B
Fig. 8 . Poles-zeroes maps, site RB. (a) Normal subject (AF). (b) fibrosis.
V . AUTOMATIC CLASSIFICATION Automatic classification of signals have been successfully applied to chest diagnosis by breath sounds [4]. Similar methods are suggested in order to automatically classify normal and pathological subjects, using speech stimulated chest features.
The estimated mean feature vector (12) and the inverse of the covariance matrix (13) were used as a template of the healthy chest. In order to test the recognition system, a test set of 14 subjects was used. Nine subjects were healthy males (not included in the training group). One of the healthy subjects was instructed to say the testing utterance at different pitch frequencies and intensities in order to include the change of parameters not related to physiological state of the chest. The other five subjects had various pulmonary dysfunctions-pleural effusion, fibrosis, pneumonia, obstructive lung disease, and severe lung tumor. The mahalanobis distance [3] was used for the recognition. The distance d,,Hbetween the signal E‘‘) acquired from the RB site of a subject under test and the template of healthy subjects is given by
In an automatic recognition system one would estimate a template for each one of the pathologies of interest, calculate the distance of the signal under test, to each one of these templates, and apply the minimum distance classification method [3] to determine the classification decision. This procedure requires sufficient training data for each one of the pathologies. Since such a database is not yet available, the procedure was checked for screening rather than recognition application. In screening one has to classify the data into one of two classes: normal (healthy) and abnormal. The problem may be posed as follows: given a template of the “healthy” random process, accept or reject the hypothesis that the given sample function,
I
COHEN A N D BERSTEIN: ACOUSTIC TRANSMISSION OF RESPIRATORY SYSTEM
Distance +
Fig. 9. Distance distribution (RB). Distances are calculated from the tested subject to the “healthy” template.
acquired from the subject under test, belongs to the “healthy” process. This hypothesis testing may be performed using the distance given by(14). We accept the hypothesis, and classify the tested subject as “healthy” if dr,His less than some empirical threshold T . If the distribution of the distance of “healthy” subjects from the healthy template was known, the threshold T that yielded some allowed mean false accept error could be theoretically determined. Fig. 9 depicts results of distance measurements taken from the limited test group. The distances are between the “healthy” template and signals taken from the test subjects. The right side of Fig. 9 (Normal) is a histogram of the distribution of dx,H given x is “healthy”. The relatively large number of “healthy” segments in the test group allowed the formation of the histogram. The distances, dx,Hgiven x belongs to one of the pathologies, are marked by arrows representing the mean over the few segments available. The data shown in Fig. 9 suggest that the normal (healthy) and abnormal classes are well separable by the suggested method. A threshold level of 70 may be used for screening applications.
VI. DISCUSSION AND CONCLUSIONS Two methods for the acoustical analysis of the chest have been suggested. Both methods use the speech utterance as the stimulating function. This type of stimulation has several advantages: it is controllable, repeatable, and its source location is known and defined. One of the disadvantages is the fact that it cannot be easily used with noncooperative subjects such as infants (the transfer function method may however be applied, using the infant’s cry [ 141 or an external acoustical source [ 151 as stimulus). Since the chest acts as a low-pass filter, it is advisable to choose a test utterance which contains sufficient power in the lower region of the frequency range. A continuous, almost stationary, vowel utterance (such as the continuous vowel /i/) is used in the system. The vowels having low frequency energies are /U/ and /i/. The vowel /U/ has a first format (Fl-the first peak in the spectrum) range of 200-400 Hz and a second format ( F 2 ) range of about 500-1000 Hz. The vowel ti/ has about the same F1 range and a much higher F 2 range of 2-4 kHz. Both these continuous vowels are suitable for stimulating the chest. For qualitative analysis, speech utterances like “ninety-nine’’ may be used. Using such utterances for a more detailed, quantitative, analysis will require sophisticated time warping [ 101 algorithms. The AR method estimates the power spectral density (PSD) function of the speech at the recording site of the chest. This
-
-7
131
PSD depends on the microphone assembly’s transfer function and the utterance. As seen in Figs. 2 and 3, it is heavily influenced by the pitch frequency (Fo), which may largely be changed from subject to subject. The estimated PSD may also depend on the time at which the signal is taken along the continuous utterance since the lungs are deflated during speech. The order of the AR estimator used in Figs. 2 and 3 was chosen a s p = 20, so that the detailed PSD, with its pitch peaks is displayed. A lower order AR estimator will provide a smoothed PSD estimate, reducing the dependence on the pitch. In the classification experiments a lower order of p = 10 was used with only the first five AR coefficients and the prediction error participating as features. Even with all its disadvantages the AR method separated well the abnormals from the normals, as demonstrated by the data of Figure 9. The data presented here is based on a very limited database that consisted of only 19 normals and five abnormal subjects. A comprehensive database is in the process of being established. Such a database will allow the estimation of the distance distribution and the optimal set of features to be used. The smoothed PSD has ony one peak, it is therefore logical to assume that an AR order o f p < 10 should suffice. The features used in this work were arbitrarily chosen. With a larger database one may apply optimization techniques [3] to establish a set of optimal features to be used in (1 1). The larger database will also allow the evaluation of the system as a recognition, rather than a screening system. The fact that the AR algorithm is simple motivated (in this work) the use of AR rather than ARMA method. The ARMA method used here (or other recursive or block methods availability in the literature) though more complex than the AR algorithm may also be implemented on modem microprocessors. The ARMA method provides the chest transfer function independent of the microphone assembly and the subject’s pitch. It thus yields information which is directly related to the problem under investigation. The ARMA ( p , q ) coefficients may be used as the features in ( I 1) to provide probably a better classification process. Since the goal of this work was to prove feasibility of the suggested methods, only the RB site was utilized. A clinical system will use multichannel data since various pathologies may be best distinguished by data taken from specific sites. In minimum distance classifiers, it is beneficial to work with a set of orthonormal gaussian features. The distance distributions (measured from the true process expectation) are then chisquare distributions with known degrees of freedom. It is possible to orthogonalize the LPC (or ARMA) features using a (not too simple) orthogonalization algorithm. The speech LPC features amplitude distributions may be considered gaussian. This, however, has not been shown for the type of utterances used here and for a speech that was transformed by the chest. It may thus very well be that a theoretically calculated distance distribution to be used for a theoretical determination of threshold levels will not be practical. It is, however, our intention to investigate it when a larger database is available. Several attempts have been made to model the acoustic impedance and transfer of the respiratory system, which includes the vocal tract, the trachea, the network of branching airways, the lung parenchyma, and the chest wall [ 161-[20]. The model by Wodicka et al. [ 181 predicts a peak in the transfer function at approximately 250 Hz, which nicely matches the data in Fig. 6 (RA position). The model, however, does not explain the shift of the peak to lower frequency of about 150
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38. NO. 2, FEBRUARY 1991
I32
Hz and 0 at higher frequencies depicted in Fig. 6 (RB position), nor does it explain the second peak at about 700 Hz in the fibrotic transfer function (Fig. 7). These may be explained by the interactions between the waves generated by different segments of the respiratory system [ 181. It is concluded that speech stimulated chest analysis systems have the potential of yielding important clinical information. The feasibility has been demonstrated (on a small size population) in this work. The ARMA method is especially attractive due to its independence of pitch and microphone. Additional results, using a larger database, are required in order to establish the clinical use of the suggested method.
REFERENCES [ 11 S. Lehrer, Understanding Lung Sounds.
Philadelphia, PA: W.B. Saunders, 1989. [2] G. Druger, The Chest: Its Signs and Sounds. Los Angeles, CA: Humetrics Corp., 1973. [3] A. Cohen, Biomedical Signal Processing. Boca Raton, FL: CRC Press, 1986. [4] A. Cohen and D. Landsberg, “Analysis and automatic classification of breath sounds,” IEEE Trans. Biomed. Eng., vol. BME31, pp. 585-590, 1984. 151 R. B. Urquhart, J. McGee, J. E. S . Macleod, S . W. Banham, and F. Moran, “The diagnostic value of pulmonary sounds: A preliminary study by computer aided analysis,” Comput. Biol. Med., vol. 11, pp. 129-139, 1981. [6] A. Berstein and A. Cohen, “Speech processing as a chest diagnosis assist device,” in Proc. IEEE 14th Conrsent. Elect. Electron. Eng., Tel-Aviv, Israel, 1985. 171 S. S. Kraman and D. Austrheim, “Comparison of lung sounds and transmitted sound amplitude in normal men,” Amer. Res. Respir. Dis., vol. 128, pp. 451-454, 1983. [SI T. R. Fenton, H. Pasterkamp, A. Tal, and V . Chernick, “Automated spectral characterization of wheezing in asthmatic children,” IEEE Trans. Biomed. Eng., vol. BME-32, pp. 50-55, 1985. (91 V . A. McKusik, J. T. Jenkins, and G. N. Webb, “The acoustic basis of the chest examination studies by means of sound spectrography,” Amer. Res. Tuberc. Pul. Dis., vol. 72, pp. 12-34, 1955. [ 101 L. R. Rabiner and R . W. Schafer, Digital Processing of Speech Signals. Englewood Cliffs, NJ: Prentice Hall, 1978. [ I l l C. K. Druzgalski, R. L. Donnenberg, and R. M. Campbell, “Techniques of recording respiratory sounds,’’ J . Clini. Eng., vol. 5 , pp. 321-330, 1980. [ 121 N . Sugumara and K. Ikegaya, “Characteristics of the air cavities of phono-cardiographic microphones, ” Med. Biol. Eng. Comput.. vol. 15, pp. 240-247, 1977. [13] M. S . Ahmed, “Fast GLS algorithms for parameter estimation,” Automatica, vol. 20, pp. 231-236, 1984. [14] A. Cohen and E. Zmora, “Automatic classification of infant’s hunger and pain cry,” in Proc. Int. Conf Dig. Sig. Proc., Florence, Italy, 1984, pp. 667-672.
[15] M. Miyakawa, K. Yamamoto, and T. Mikami, “Acoustic measurement of the respiratory system. An acoustic pneumograph,” Med. Biol. Eng., vol. 14, pp. 653-659. 1976. [16] J. J . Fredberg and A. Hoenig, “Mechanical response of the lungs at high frequencies,” J . Biomech. Eng., vol. 100, pp. 57-66, 1987. [17] K. Ishizaka, M. Matsudaira, and T. Kaneko, “Input acoustic impedance measurement of the subglottal system,” J . Acoust. Soc. Amer., vol. 60, pp. 190-197, 1976. [18] G. R. Wodicka, K. N . Stevens, H. L. Golub, E. G. Cravalho, and D. C. Shannon, “A model of acoustic transmission in the respiratory system,” IEEE Trans. Biomed. Eng., vol. BME-36, pp. 925-934, 1989. [19] R. W. Guelke and A. E. Bum, “Transmission line theory applied to sound wave propagation in tubes with compliant walls,” Acoustica, vol. 48, pp. 101-106, 1981. [20] W. L. Capper, “Acoustic input impedance measurements on elastic walled tubes and large airways in man,” Ph.D. Thesis, Dep. Biomed. Eng., Univ. Cape Town, South Africa, 1988.
Arnon Cohen (M’69-M’80) was born in Haifa, Israel, in 1938. He received the B.Sc. and M.Sc. degrees from the Technion-Israel Institute of Technology in 1964 and 1966, respectively, and the Ph.D. degree from CarnegieMellon University, Pittsburgh, PA, in 1969. Since 1972 he has been with the Department of Electrical and Computer Engineering and the Biomedical Engineering Program, Ben Gurion University, Beer-Sheva, Israel where he is a Professor of Electrical and Biomedical Engineering. He was with the Department of Electrical Engineering and Bioengineering, University of Connecticut, Storrs, from 1967 to 1972, the Colorado State University, Fort Collins, during 1976-1977, and the Department of Biomedical Engineering. University of Cape Town Medical School, Cape Town, South Africa, during 1989-1990. His research interests are in signal processing, mainly with biomedical applications. He is the author of the book Biomedical Signal Processing (Boca Raton, FL: CRC).
Albert0 D. Berstein was born in Argentina in 1954. He received the B.Sc. and M.Sc. degrees in electrical and computer engineering in 1981 and 1985, respectively, from the Ben Gurion University, Beer-Sheva, Israel. Since 1985 he has been engaged in Speech Enhancement and Speaker Recognition research work. He is now with the DSP Group, Israel. He is also a graduate student in the Electrical and Computer Engineering Department, Ben Gurion University, working towards the Ph.D. degree.
