An adaptive noise reduction stethoscope for auscultation in high noise environments

Samir B. Patel
School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana 47907-1285

Thomas F. Callahan and Matthew G. Callahan

University Research Engineers & Associates, Inc., Acton, Massachusetts 01720

James T. Jones, George P. Graber, and Kirk S. Foster

Hillenbrand Biomedical Engineering Center, Purdue University, West Lafayette, Indiana 47907-1296

Kenneth Glifort
Armstrong Laboratory, U.S. Air Force, San Antonio, Texas 78201

George R. Wodickaa)
School of Electrical and Computer Engineering and Hillenbrand Biomedical Engineering Center, Purdue University, West Lafayette, Indiana 47907-1285

Received 17 January 1996; accepted for publication 23 January 1998 Auscultation of lung sounds in patient transport vehicles such as an ambulance or aircraft is unachievable because of high ambient noise levels. Aircraft noise levels of 90100 dB SPL are common, while lung sounds have been measured in the 2230 dB SPL range in free space and 6570 dB SPL within a stethoscope coupler. Also, the bandwidth of lung sounds and vehicle noise typically has signicant overlap, limiting the utility of traditional band-pass ltering. In this study, a passively shielded stethoscope coupler that contains one microphone to measure the noise-corrupted lung sounds and another to measure the ambient noise was constructed. Lung sound measurements were made on a healthy subject in a simulated USAF C-130 aircraft environment within an acoustic chamber at noise levels ranging from 80 to 100 dB SPL. Adaptive ltering schemes using a least-mean-squares LMS and a normalized least-mean-squares NLMS approach were employed to extract the lung sounds from the noise-corrupted signal. Approximately 15 dB of noise reduction over the 100600 Hz frequency range was achieved with the LMS algorithm, with the more complex NLMS algorithm providing faster convergence and up to 5 dB of additional noise reduction. These ndings indicate that a combination of active and passive noise reduction can be used to measure lung sounds in high noise environments. 1998 Acoustical Society of America. S0001-4966 98 02605-8 PACS numbers: 43.60.Qv, 43.80.Cs, 43.80.Ev, 43.80.Vj JLK

INTRODUCTION

Auscultation is the process of listening to sounds emanating from the body. Medical personnel frequently make diagnoses pertaining to the health of the respiratory system using a standard binaural stethoscope. However, this stethoscope is highly prone to interference from ambient noise and thus becomes clinically useless in high ambient noise environments. In a typical emergency, civil or military, an ambulance or aircraft helicopter or xed wing is used to evacuate critically ill patients. Several sources of noise exist in such vehicles, and their interaction with the interior space or fuselage is complex and time variant due to the wide range of conditions and maneuvers encountered during a typical evacuation. For example, ambient noise levels in the range of 90100 dB SPL sound pressure level exist in the cabin of C-130 aircraft used by the U.S. Air Force USAF for aeromedical evacuation. In contrast, normal breath sounds in an healthy adult are roughly 2230 dB SPL in free space exa

Electronic mail: wodicka@ecn.purdue.edu J. Acoust. Soc. Am. 103 (5), Pt. 1, May 1998

cluding airow generated noise radiating from the mouth and 6570 dB SPL within a stethoscope coupler Zenk, 1994 . Thus traditional auscultation in such a high noise environment is not feasible due to an inherently poor signal-tonoise ratio SNR . In fact, even moderate background noise levels such as those found during average conversation are sufcient to hamper auscultation in the clinic. In general, the SNR can be improved either by increasing the signal strength numerator or by decreasing the interfering noise denominator . However, the high noise levels and strong overlap of spectral content of lung sounds and the time-varying vehicle noise impose limitations on the successful employment of conventional passive noise reduction schemes. Thus in this study, a digital scheme for adaptive noise reduction ANR was implemented in addition to passive noise reduction PNR at the transducer in an integrated approach. In this manner, the time varying characteristics of the noise are estimated and exploited to effectively extract the desired lung sounds. This manuscript details the design and preliminary testing of a Digital Active Noise Attenuation DANA Stethoscope, with a focus on the function of the
1998 Acoustical Society of America 2483

2483

0001-4966/98/103(5)/2483/9/$10.00

FIG. 1. Diagram of the stethoscope coupler. Diaphragm is 4 cm in diameter.

adaptive lters. The two algorithms that were implemented are the least-mean-squares LMS and normalized leastmean-squares NLMS .
I. STETHOSCOPE COUPLER

A passively shielded coupler with two omnidirectional electret microphones Telex ELM-22, Minneapolis, Minnesota was constructed Fig. 1 . The microphone d is placed in a conical air cavity behind a standard stethoscope diaphragm to detect the lung sounds Wodicka et al., 1994 . Such an arrangement allows signicant lung sound energy to reach microphone d, with a frequency response that mimics that of the standard stethoscope. The other microphone, x, measures the noise reference that drives the adaptive lter. These sensors are positioned in the coupler so that microphone x detects very little lung sounds, yet measures sound that is highly correlated to the noise contaminating the lung sounds measured by microphone d. The stethoscope coupler is passively shielded from high frequency ambient noise via a foam surround since the effectiveness of adaptive noise reduction decreases with increasing frequency.
II. ADAPTIVE NOISE CANCELLATION

FIG. 2. Adaptive noise canceler: a block diagram; b tapped delay line.

liott et al., 1987; Poole et al., 1984 . Figure 2 a depicts the basic construct for noise reduction in a linear system with additive noise. The primary input d consists of signal s and noise n 0 for convenience the time index j is omitted . The reference input, i.e., the input of the adaptive lter, consists of noise x. The signal s is assumed to be uncorrelated with both n 0 and x, while n 0 and x are assumed to be correlated with each other in some unknown way. n 0 is often referred to as the primary noise and x as the noise reference. In this approach, x is adaptively ltered to produce an output y which is then subtracted from the primary input d to produce the system output e the recovered signal where e s n 0 y. 1

The conventional method of signal estimation from its noise corrupted version is to pass it through a lter that removes the unwanted noise without signicantly affecting the signal. The lter employed can be xed or self-adjusting, i.e., adaptive. If the signal and noise characteristics are known a priori the optimal xed lter can be designed based on Wiener lter theory. In practice, a priori knowledge of noise characteristics is generally not available and hence a lter that adapts to the characteristics of the noise is desired. In the following, we present the basic theory of the adaptive lter algorithms considered in this study.
A. Least-Mean-Squares LMS algorithm

From Eq. 1 , it is observed that minimization of the meansquare error MSE E e 2 results from the minimization of E (n 0 y) 2 , i.e., min E e 2 E s2 min E n 0 y
2

The LMS algorithm, developed by Widrow and Hoff in 1959 Widrow et al., 1985; Haykin et al., 1986 , has been studied in great detail and has found a myriad of applications Widrow et al., 1975; Darlington et al., 1985; Rodriguez et al., 1987; Harrison et al., 1986; Pulsipher et al., 1979; El2484 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

and that smallest possible output power is E e 2 E s 2 when E (n 0 y) 2 0. In this case, the output y of the adaptive lter is a replica of the noise n 0 in the primary input and thus n 0 is completely eliminated and the output SNR is maximized. The adaptive lter in Fig. 2 a typically consists of a tapped delay line as depicted in Fig. 2 b . The output y j , where j is the time index, is equal to the inner product of noise reference input vector X j and lter coefcient vector W j , each of length M , i.e., y j W TX j j
Patel et al.: Adaptive noise reduction stethoscope

3
2484

and the general expression for the error signal as a function of weight vector is given by e j d j y j d j W TX j j 4

B. Normalized LMS NLMS algorithm

In the minimization of the MSE, when the true value of gradient derivative of j E e 2 with respect to lter coefj cient vector W j is equated to zero, the steepest descent method Widrow et al., 1985 yields the form of the socalled WidrowHoff LMS algorithm Wj
1

Wj

e jX j ,

where is the step size which controls the stability and rate of convergence. The larger the value of , the faster the algorithm converges and the more variable is the gradient estimate, and vice versa Widrow et al., 1976 . The magnitude-squared coherence function
2 pq

The LMS algorithm described in the previous section uses a constant step size . The stability, speed of convergence, and uctuation of the adaptation process are governed by the step size and the input power to the adaptive lter P x . The normalized LMS NLMS algorithm Treichler et al., 1987 represents a technique to improve the speed of convergence. This is accomplished while maintaining the steady-state performance independent of the input signal power. This NLMS algorithm uses a time variable convergence factor j ,
j

P x, j

10

ej

S pq e j 2 S p p e j S qq e j

is often used to measure the correlation between two stationary random processes p and q where S pq is the cross-power spectral density of p and q, and S pp , S qq are the respective auto-power spectral densities. When the lter output signal, e, is orthogonal to the noise reference data, x, the steepest-descent method converges to the optimal Wiener lter coefcients W * and there are no further updates. On the other hand, in the steady state, the LMS algorithm continues to have updates. These uctuations, called weight jitters, are about the optimal Wiener lter coefcients W * and the LMS algorithm does not converge with unity probability. It can be shown that the mean of these updates converges to W * under the assumption of stationarity of the input to the adaptive lter Haykin, 1986 . However, the MSE, , converges to a larger value than the optimal minimum mean square error, min achieved with an optimal Wiener lter . With the assumptions made in independence theory Haykin, 1986 , the necessary and sufcient condition for the convergence of LMS algorithm is that step-size parameter satises 0 2
max

where is the normalized step size chosen to be between 0 and 2, while is a small positive term included to ensure that the update term does not become excessively large when the input average power at time j, P x, j , becomes small. The NLMS algorithm gains its stability by normalizing the weight vector update with an estimate of the signal power. A computationally inexpensive way of obtaining this normalization is via recursive power estimation P x, j
1

P x, j M x 2 , j

11

1 is a smoothing parameter and M is the lter where 0 length, with the NLMS algorithm weight vector update equation as Wj
1

Wj

je jX j

12

C. Criteria for performance measures of adaptive algorithms

A study of the relationship between the speed of adaptation and performance of adaptive systems is complex since different choices of performance measures often termed misadjustments yield different optimized step sizes. To optimize the step size of the LMS algorithm and study its performance, the root-mean-squared RMS misadjustment, a performance measure similar to the temporal root-meansquared MSE was used. This RMS misadjustment is dened as MRMS E sj ej E s2 j
2 1/2

13

where max is the maximum eigenvalue of the autocorrelation matrix of the input to the lter, R M M . In practice, the calculation of the individual eigenvalues of R M M to nd the given by Eq. 7 is computationally very upper limit of intensive. Under the stationarity assumption of the input to the lter, and positive deniteness of R M M ,
M max i 1 i

trace R M M

M Px ,

where P x is the total power input to the adaptive lter, a generally known quantity. Thus in practice, a tighter bound is placed on the upper limit of and Eq. 7 becomes 0
2485

2 . M Px
J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

where s j is the lung sounds signal component and e j is the output of the adaptive system the recovered lung sounds . MRMS is a normalized root-mean-squared MSE and reects the degree of the distortion of the desired signal introduced by the adaptive system. Obviously, the optimal step-size parameter and the * optimal lter length M that will yield a minimum MRMS is * desired. In general, faster adaptation leads to a more noisy adaptive process. In a stationary environment, enhanced steady-state performance smaller MRMS results from a smaller value of the step size and thus a slower convergence. In a nonstationary environment however, a compromise has to be made between fast convergence necessary to track the changes in the input signal and slow adaptation needed to reduce the gradient noise misadjustment . Thus
Patel et al.: Adaptive noise reduction stethoscope 2485

FIG. 3. Experimental apparatus.

the optimal value of is a compromise between a fast rate of convergence and a reduced misadjustment. Further, it can be shown that even for a stationary input, the misadjustment Haykin, 1986; due to gradient noise, G, is proportional to Widrow et al., 1976 , G M Px , 14

and this relationship is used in this study to calculate the value of for given values of G, M , and P x . A performance measure that reects the optimization of lter length M also must be chosen. As mentioned previously, the steady-state error signal e the recovered breath sounds and the reference data samples x used to obtain the minimum MSE MMSE estimate of the desired signal obeys orthogonality for the steepest descent method and even for the optimal Wiener lter . Expanding upon this relationship, the cross correlation between the output error signal e and the reference noise x should attain some minimum value close to zero for the optimal lter length M . To quantitate * this relationship, the correlation parameter COR was dened as COR 1 L
L j 1

ject was seated in a soundproof room that contained an audio system which played the recordings of USAF C-130 aircraft fuselage noise to simulate this environment. The subject breathed through a pneumotachograph Fleisch #2 at target ow rates of 2 L/s. An experimenter held the stethoscope coupler on the subjects right anterior upper chest during the protocol. The signals from both microphones d and x were amplied by a factor of 32.5 and high-pass ltered using a fourth-order Butterworth lter at a cutoff frequency of 100 Hz. These signals and the pneumotachograph output were digitized at 5120 samples per second using a Tektronix Fourier Analyzer 2642A, with internal anti-aliasing lters and stored on an IBM-PC compatible computer. Data segments of 20-s duration were recorded during breathing and apnea at aircraft noise levels of 80, 90, and 100 dB SPL, and also in the quiet room. In addition, noise measurements were made with the coupler placed on the subjects thigh. The stored digital epochs were subsequently processed with the adaptive noise reduction LMS and NLMS algorithms via MATLAB, where a detailed performance analysis was performed. The signals d and x were also simultaneously input to a real-time implementation of one version of the LMS algorithm running on a digital signal processing board ATDSP2200, National Instruments with a second IBM-PC compatible computer. The output of this algorithm was fed to an audio amplier DRA-345R, Denon and presented to the experimenter within the chamber via a noise cancellation headset AH-BG, Bose . Thus the experimenter could monitor the extracted lung sounds from a specic realization of the algorithm during the recording process, yet the focus of this study was on the off-line processing of the stored acoustic signals.
IV. RESULTS AND DISCUSSION A. Characteristics of lung sounds and the interfering noise

e jx j .

15

As M increases, COR should decrease as relatively more and more uncorrelated data samples are taken into the estimation of the desired signal.
III. EXPERIMENTAL APPARATUS AND PROCEDURE

The primary goal of the experimentation was to record and process lung sounds in a simulated USAF C-130 aircraft environment. Figure 3 provides an overview of the experimental apparatus. The protocol was approved by the Purdue University Committee on the Use of Human Subjects. A sub2486 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

Figure 4 depicts representative data collected in the quiet chamber. The pneumotachometer signal Fig. 4 a is positive for inspiration and negative for expiration. Figure 4 b is the corresponding lung sounds detected at microphone d. For all the plots of microphone measurements, 1 V corresponds to 3.1 Pa. The most clear and audible lung sounds occur for a ow rates above roughly 1 L/s. Figure 4 c is the time record of the noise reference signal x, where some detection of breathing is observed during expiration due to airborne sounds from the mouth reaching the coupler. Figure 5 provides in the same format a representative data epoch in the presence of the high level noise. Here the measured unprocessed lung sounds are overwhelmed by the noise and hence are inaudible. The amplitude of the noise reference is less than that of the lung sounds detected in the presence of noise because the relatively high acoustic impedance of the air cavity in front of microphone d results in greater sound pressure levels measured as compared to microphone x which is essentially open to free space. An estimate of the power spectrum of both inspiratory lung sounds from the microphone d signal detected during breathing in a quiet chamber and aircraft noise from the microphone d
Patel et al.: Adaptive noise reduction stethoscope 2486

FIG. 4. Time record of a pneumotachometer, b microphone d lung sounds , c microphone x noise reference signals in a quiet chamber.

FIG. 6. Relative power spectra of aircraft noise, inspiratory lung sounds, and background noise in a quiet chamber.

signal during apnea in presence of 100 dB SPL of aircraft noise is shown in Fig. 6. The majority of the energy in both signals is concentrated in the 100600 Hz frequency range. This strong overlap in spectral content makes it particularly difcult to auscultate inside an aircraft. The coherence, as a measure of correlation between the noise reference x and the noise corrupting the lung sounds n 0 , is depicted in Fig. 7. The valley in the coherence at just above 800 Hz corresponds to that in the aircraft noise spectra.
B. Performance analysis of the adaptive algorithms

To optimize and preliminary evaluate the processing schemes, the lung sound measurements microphone d made in the quiet chamber signal, s were rst considered the pure desired signal. To this signal, s, the amplitude of the

aircraft noise detected by microphone d during nonbreathing with the stethoscope coupler on the right thigh was added to obtain a simulation of signal plus noise, s n 0 for each of the respective noise levels. These generated s n 0 data were input as the primary signal to the adaptive lter, with x as the noise reference signal, to analyze and optimize the performance of the algorithm. The resulting algorithms were then used to process actual experimental data. This initial data simulation stage was required since simultaneous recording of pure lung sounds and its corrupted version is not possible. Some measures that were used to assess the algorithm were: 1 Optimization of step size and lter length M : For the LMS algorithm, the values of were calculated using Eq. 14 for various values of gradient misadjustment G ranging from 0.01 to 1.0 and lter lengths M of 10, 20, and 40. With these values of M and the calculated values of ,

FIG. 5. Time record of a pneumotachometer, b microphone d lung sounds , c microphone x noise reference signals in 100 dB SPL of aircraft noise. 2487 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

FIG. 7. Coherence estimate between the primary microphone d and the noise reference microphone x signals at 100 dB SPL of aircraft noise. Patel et al.: Adaptive noise reduction stethoscope 2487

FIG. 8. For the LMS algorithm: Performance measure MRMS versus step size for lter length M 10 (), M 20 ( ), and M 40 ( ) in a 80, b 90, and c 100 dB SPL of aircraft noise.

FIG. 10. For the NLMS algorithm: Performance measure MRMS versus step size for lter length M 20 ( ), M 40 ( ), and M 80 (). a 80, b 90, and c 100 dB SPL of aircraft noise.

the LMS algorithm was implemented. For each case the performance measures of MRMS and COR were calculated. Figure 8 shows plots of MRMS versus for 80, 90, and 100 dB SPL noise levels which are generally consistent with theoretical predictions Haykin, 1986; Widrow et al., 1976 . MRMS for 80 and 90 dB SPL of noise Fig. 8 a , b was a relatively weak function of , with increasing slightly increasing the MRMS as a reection of G. The respective curves for 100 dB SPL of noise Fig. 8 c exhibit the exisfor which tence of optimal step-size parameter * MRMS is a minimum. In the optimization of lter length M , the COR parameter was also taken into consideration. Plots of COR versus M Fig. 9 show that in general a lter with a larger number of coefcients has a slightly lower COR.

When the lter length tends to an optimal value M , the * value of COR does not change signicantly thereafter. Thus from the perspective of implementation, the lowest value of M whereafter COR remains almost constant could be selected as the optimal value M . A similar evaluation of * the NLMS algorithm was carried out for both the simulated and actual data for lter lengths M of 20, 40, and 80 with normalized step sizes of 0.2, 0.8, 1.2, and 1.6. Plots of the performance measures for this case Figs. 10 and 11 depict the existence of potentially optimal parameter settings. The MRMS for the NLMS algorithm Fig. 10 is relatively lower than for the LMS algorithm because the use of a normalized dynamic step size drastically improves the tracking capability of the NLMS algorithm. Further, COR for the NLMS algorithm Fig. 11 shows M 40 as an optimal lter length from an implementation perspective. This observation is similar to that made with the LMS algorithm.

FIG. 9. For the LMS algorithm: Performance measure COR versus lter length M for step size 0.0073 (), 0.0146 ( ), and 0.0437 ( * ) in a 80, b 90, and c 100 dB SPL of aircraft noise. 2488 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

FIG. 11. For the NLMS algorithm: Performance measure COR versus lter length M for step size 0.2 ( ), 0.8 (), and 1.2 ( ) in (a) 80, (b) 90, and (c) 100 dB SPL of aircraft noise. Patel et al.: Adaptive noise reduction stethoscope 2488

TABLE I. SNR with the LMS algorithm M algorithm M 40 and 1.2 . Noise level dB SPL Quiet 80 90 100 Noise level dB SPL 80 90 100

40 and

0.02 , and NLMS

Simulated data Input SNR 3.0 : 1.0 1.1 : 1.0 1.0 : 2.1 1.0 : 7.7 Output SNR LMS 3.0 : 1 1.7 : 1 1.7 : 1 1.8 : 1 Actual data Input SNR 1.2 : 1.0 1.0 : 2.0 1.0 : 7.7 Output SNR LMS 1.8 : 1 2.0 : 1 2.1 : 1 NLMS 2.6 : 1 2.6 : 1 2.3 : 1 NLMS 3.0 : 1 2.8 : 1 2.9 : 1 2.7 : 1

2 Signal-to-noise ratio SNR : In this study, the SNR in linear form at the input of the adaptive lter was conservatively dened as the square root of the ratio of variance of the microphone d signal when there is no aircraft noise for inspiratory ow rates 1 L/s, to that when the aircraft noise level is present at inspiratory ow rates 1 L/s. The SNR at the output of the adaptive lter was calculated as the square root of the ratio of variances of the recovered signal e during inspiratory ow rates 1 L/s to that during inspiratory ow rates 1 L/s. The difference of SNR at the input and output reects a measure of the noise reduction capability of the conguration. Table I shows that higher SNR values are achieved at the output of the adaptive lter both for the simulated and actual data with NLMS algorithm than with the LMS algorithm, for essentially all cases.
C. Detailed performance comparison of LMS and NLMS algorithms

FIG. 12. For simulated data: Time record of a typical lung sounds in a quiet chamber, b noise corrupted lung sounds microphone d at 100 dB SPL of aircraft noise, c noise reference microphone x and recovered lung sounds with d LMS algorithm M 40 and 0.02 , and e NLMS algorithm M 40 and 1.2 (1 V 3.1 Pa).

bandwidth with a small amount of added computational complexity, and hence it is the algorithm of choice for the reduction of corrupting C-130 aircraft noise. The recovered lung sounds for actual data are shown in Figs. 14 and 15, noting that the SNR is lower than that for the simulated case, as expected. This observation highlights the signicance of the need for the reference noise microphone x to be highly correlated with the noise in the primary input microphone d to enhance the performance of the adaptive noise canceler. For the simulated data, the primary

Figure 12 a presents time records of typical lung sounds detected in a quiet chamber at microphone d. The lung sounds detected with microphone d and the reference noise detected at microphone x in the presence of 100 dB SPL aircraft noise are shown in Fig. 12 b and c , respectively. The recovered time series from the NLMS algorithm Fig. 12 e is very similar to that of the pure lung sounds, with the misadjustment less than that of the LMS algorithm Fig. 12 d due to the enhanced tracking ability of the NLMS algorithm. The speed of convergence is also much faster for the NLMS algorithm since it employs a normalized dynamic step size which renders the rate of convergence independent of the input signal power. The spectral features Fig. 13 of the recovered lung sounds from the algorithms nearly overlap up to approximately 450 Hz, where the NLMS algorithms recovered lung sounds more closely match the spectra of the pure lung sounds at higher frequencies. It can be inferred from Fig. 13 that roughly 15 dB of noise reduction is attained in 100 dB SPL of aircraft noise in the 100600 Hz frequency range with LMS algorithm, while the NLMS algorithm provides up to 5 dB of additional noise reduction at frequencies above 450 Hz. The NLMS algorithm therefore affords higher noise reduction over wider
2489 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

FIG. 13. For simulated data: Estimated relative power spectra of inspiratory lung sounds as measured in a quiet chamber , noise corrupted lung sounds at 100 dB SPL of aircraft noise ( ), and recovered lung sounds from the LMS algorithm M 40 and 0.02 , and NLMS algorithm M 40 and 1.2 . Patel et al.: Adaptive noise reduction stethoscope 2489

FIG. 14. For actual data: Time record of recovered lung sounds with a LMS algorithm M

40 and

0.02 , b NLMS algorithm M

40 and

1.2 .

input was composed by adding the microphone d signal during apnea with the coupler placed on the thigh in presence of the known level of the aircraft noise to the pure lung sound signal i.e., microphone d signal during breathing in a quiet chamber . Further, for processing of these simulated data, the same microphone d signal obtained during apnea in presence of the known level of the aircraft noise is used as the reference noise. Hence, for the simulated data, the correlation between the reference noise and the noise in the primary channel is perfect, whereas for the actual data, the primary input microphone d and reference noise microphone x signals are obtained using two different microphones. Here, the correlation between the primary input and the reference noise is governed by the transfer function between the physical locations of microphones d and x inside the stethoscope coupler. The trade-off in acoustic isolation of these

microphones is to collect a reference noise signal that is least corrupted by the lung sounds and highly correlated with the noise component in the primary input. Thus the adaptive lter algorithm for the simulated data provides a better estimate of the noise in the primary input and hence results in a recovered lung sounds of higher delity as compared to those obtained for the actual data.
V. CONCLUSIONS

In patient transport vehicles, high ambient noise has signicant spectral overlap with lung sounds, making lung sounds inaudible through a standard stethoscope and limiting the utility of conventional amplication and ltering techniques. Passive noise reduction techniques such as a coupler housing help to shield the lung sounds from high-frequency noise corruption. An adaptive noise reduction system implemented using a digital adaptive lter employing the LMS or NLMS algorithms signicantly reduces the corruptive effects of low frequency noise. The proposed performance criteria of MRMS and COR are useful to determine algorithm parameters, such as convergence factor and lter length for the LMS approach. Approximately 15 dB of noise reduction over a 100600 Hz frequency range was achieved by the LMS approach, with the somewhat more complex NLMS algorithm yielding faster convergence and up to an additional 5 dB of noise reduction. Further studies on patients during actual medical evacuation are required to rene this technology and determine the scope of its clinical utility under various breathing and noise conditions.
ACKNOWLEDGMENTS

FIG. 15. For actual data: Estimated relative power spectra of noise corrupted lung sounds at 100 dB SPL of aircraft noise ( ), and recovered lung sounds from the LMS algorithm M 40 and 0.02 , and NLMS algorithm M 40 and 1.2 . 2490 J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

This research was performed for the U.S. Air Force, Armstrong Laboratory under a phase II Small Business Innovation Research subcontract to University Research Engineers & Associates, Inc. The authors thank Saul B. Gelfand for his advice concerning the adaptive lter design and evaluation.
Patel et al.: Adaptive noise reduction stethoscope 2490

Darlington, P., Wheeler, P. D., and Powell, G. A. 1985 . Adaptive noise reduction in aircraft communication systems, Proceedings of IEEE Conference on Acoustics, Speech and Signal Processing, Vol. 2, pp. 716719. Elliott, S. J., Stothers, I. M., and Nelson, P. A. 1987 . A multiple error lms algorithm and its application to the active control of sound and vibration, IEEE Trans. Acoust., Speech, Signal Proc. ASSP-35, 14231434. Harrison, W. A., Lim, J. E., and Elliott, S. J. 1986 . A new application of adaptive noise cancellation, IEEE Trans. Acoust., Speech, Signal Proc. ASSP-34, 2127. Haykin, S. 1986 . Adaptive Filter Theory Prentice-Hall, Englewood Cliffs, NJ . Poole, L. A., Warnaka, G. E., and Cutter, R. C. 1984 . The implementation of digital lters using a modied widrow-hoff algorithm for the adaptive cancelation of acoustic noise, Proceedings of IEEE Conference on Acoustics, Speech and Signal Processing, 21.7.121.7.4. Pulsipher, D. C., Boll, S. F., Rushforth, C., and Timothy, J. 1979 . Reduction of nonstationary acoustic noise in speech using lms adaptive noise cancelling, Proceedings of IEEE Conference on Acoustics, Speech and Signal Processing, pp. 204207.

Rodriguez, J. J., Lim, J. S., and Elliott, S. J. 1987 . Adaptive noise reduction in aircraft communication systems, Proceedings of IEEE Conference on Acoustics, Speech and Signal Processing, pp. 169172. Treichler, J. R., Johnson, Jr., C. R., and Larimore, M. G. 1987 . Theory and Design of Adaptive Filters Wiley, New York . Widrow, B., Glover, Jr., J. R., McCool, J. M., Kaunitz, J., Williams, C. S., Hearn, R. H., Zeidler, J. R., Dong, Jr., E., and Goodlin, R. C. 1975 . Adaptive noise cancelling: Principles and applications, Proc. IEEE 63, 16921716. Widrow, B., McCool, J., Larimore, M. G., and Johnson, Jr., C. R. 1976 . Stationary and nonstationary learning characteristics of the lms adaptive lter, Proc. IEEE 64, 11511162. Widrow, B., and Stearns, S. D. 1985 . Adaptive Signal Processing Prentice-Hall, Englewood Cliffs, NJ . Wodicka, G. R., Kraman, S. S., Zenk, G. M., and Pasterkamp, H. 1994 . Measurement of respiratory acoustic signalsEffect of microphone air cavity depth, Chest 106, 11401144. Zenk, G. M. 1994 . Stethoscopic detection of lung sounds in high noise environments, M.S. thesis, Purdue University, West Lafayette, IN.

2491

J. Acoust. Soc. Am., Vol. 103, No. 5, Pt. 1, May 1998

Patel et al.: Adaptive noise reduction stethoscope

2491