(wheeze source) (crackle source) (breath sounds source) Fig. 1. Schematic sketch of the AR model employed for lung sounds

Transcription

1 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. 11, NOVEMBER Communications Autoregressive Modeling of Lung Sounds: Characterization of Source and Transmission V. K. IYER, P. A. RAMAMOORTHY. ANDY. PLOYSONGSANG (wheeze source) Heart Sounds Muscle and Skin Noise Abstract-In this communication, we discuss the application of autoregressive modeling to lung sounds analysis. The lung sounds source in the airway is modeled as a white noise source, consisting of one or a combination of the following sources: sequence, periodic train of impulses, and impulsive bursts of energy. The acoustic transmission through the lung parenchyma and chest wall is modeled as an all-pole filter. Using this method, the source and transmission characteristics of lung sounds are estimated separately, based on the lung sounds at the chest wall. To illustrate the potential validity of the model, lung sound segments in known disease conditions were selected from teaching tapes and the source and transmission characteristics were estimated by applying the model. The estimated characteristics were found to he consistent with current knowledge of the generation and transmission of lung sounds in the known conditions. I. INTRODUCTION The simple, noninvasive, nonhazardous, and inexpensive nature of acquiring information about the condition of the lung using lung sounds makes it an attractive candidate for clinical diagnosis. However, the clinical application of this method is still limited owing to the inability to extract adequate information from lung sounds to make reliable, sensitive, and complete diagnosis based on them. Therefore, this is an area of research interest for pulmonary physicians. In this communication, we show how autoregressive modeling [ 11 techniques can be applied usefully to lung sounds analysis in order to separately estimate source and transmission characteristics. Some situations where separation of source and transmission characteristics would be useful like early detection of pneumonia and pulmonary edema, diagnosis of pleural effusion and emphesyma, etc., are discussed. Using examples from teaching tapes containing lung sounds in known lung diseases, we compared the expected subjective source and transmission characteristics of lung sounds during the known lung conditions to the estimated source and transmission characteristics from the modeling approach. 11. THE AUTOREGRESSIVE MODEL FOR LUNG SOUNDS The source and transmission of lung sounds are autoregressively modeled as follows (see Fig. 1). Lung sounds originate inside the airways of the lung (termed source in this paper) as one of the following white noise sequences. Breath sounds sources are modeled as sequences since they are believed to be produced by turbulent flow in a large range of airway dimensions. Wheeze and rhoncus sources are modeled as a sequence of periodic impulses since they have characteristic distinct pitches associated Manuscript received September 22, 1987: revised February 6, The work of V. K. Iyer was performed at the Department of Electrical and Computer Engineering, University of Cincinnati, Cincinnati, OH V. K. Iyer is with Criticare Systems, Inc., Waukesha, WI P. A. Ramamoorthy is with the Department of Electrical and Computer Engineering, University of Cincinnati, Cincinnati, OH Y. Ploysongsang is with the Department of Internal Medicine, Division of Pulmonary Diseases, University of Cincinnati, Cincinnati, OH IEEE Log Number (crackle source) (breath sounds source) I Instrumentation Noise Fig. 1. Schematic sketch of the AR model employed for lung sounds with them, and they are believed to be produced by periodic oscillations of the air and airway walls. Crackle sources are modeled as random intermittent impulses since they are believed to be produced by sudden opening/closing of airways or bubbling of air through extraneous liquids in the airways, both phenomena associated with sudden intermittent bursts of sounds energy. One or an additive combination of two or more of these sources pass through the acoustic transmission filter, consisting of the lung parenchyma and chest wall structure; this effect is modeled to be an all-pole filter. The output of this filter is considered to be the lung sounds at the chest wall. The lung sounds at the chest wall also contain heart sounds interference, the reduction of which has been addressed earlier by us in [2] and other pending publications. Using the model hypothesized above, if we were given a signal sequence of lung sounds at the chest wall, we can apply autoregressive analysis to compute the model parameters, and therefore the source and transmission filter characteristics can separately be estimated. This separate estimation could significantly enhance the usefulness of lung sounds in the clinic. The obvious and direct method to accurately verify this hypothesis is to measure the source and transmission characteristics and compare them to estimated source and transmission characteristics from the model parameters. Since such a measurement is difficult to achieve, an alternative approach was pursued to illustrate the potential validity of the model. By using segments of lung sounds during known lung conditions, it was first deduced, based on currently accepted mechanisms of lung sounds production and transmission (January 1985 issue of Seminars in Respiratory Medicine), what is expected in the source and transmission characteristics of the corresponding lung sound segments. The predicted characteristics of source and transmission from the model were then compared to the deduced characteristics. For example, lung sounds segments during pulmonary edema would be expected to have intermittent bursts of energy corresponding to airway opening and bubbling of airway fluids in the source characteristics. Also, it would be expected that a shift of transmission spectral response to higher frequencies may be noted due to consolidation of the lung by fluids. The estimated characteristics of source and transmission would, if the model is consistent, have these characteristics. Using this approach, the potential validity of the model was tested without extensive experimentation $1.OO 1989 IEEE

2 1134 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. 11, NOVEMBER 1989 Description of segment normal, inspiratory, bronchial Associated Phenomenon Estimated Source normal breathing Transmission Filter (central frequency) low pass ( Hz) normal, inspiratory, vesicular asthmatic (wheezing), expiratory, bronchial early pneumonia, expiratory, vesicular normal breathing airway obstruction excess fluids in periodic impulses low pass (1 Hz) low pass with resonance low pass (2 Hz) interstitial spacc and pulmonary blood vessels late pneumonia (crackles), expiratory, vesicular fluids enters airways impulsive bursts of energy low pass (2 Hz) &IS SECTOR 1, STARTING FRAME 8, 15 FRAMES, CONTEXT [- TR1 1 I I I 1 I I I 1 I I I BEG SEC MID SEC END SEC (a) I 111 //I1 I I 111' ' ' I SECTOR 1, STARTING FRAME 8, 15 FRAMES, CONTEXT VALUES-.-._ TR3 1 I I 9 1 ' 1 ' I /I U 1 I 1-48 ' i BEG dec/ MID = SEC 1 I 1 ENT SM: I (d) (C) Fig. 2. (a) An inspiratoly segment of bronchial breath sounds, (b) the frequency spectrum of the segment, (c) the estimated source waveform, (d) the spectrum of the estimated source METHODOLOGY Lung sound segments, each - s in duration, representing typical diseases affecting the airways, parenchyma, and chest wall, were selected from teaching tapes [3]. The segments were filtered in the range -1.4 khz and then digitized at a sampling rate of 3 khz using a 12 bit analog-to-digital converter. The signals were subjected to autoregressive analysis. A sixth-order all-pole model was used, amved at by the criteria of nonreducing mean-square prediction error with further increase in order. The size of the segments for short-time analysis was 256 samples (.853 s). IV. EXAMPLES AND DISCUSSION Typical segments of lung sounds (Table I) were chosen for analysis. The average estimated gain of the filter was around.22. Representative results are shown in Figs. 2-6.

3 ~~ IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. II, NOVEMBER ~~ STARTING FRAME - 8, NUMBER OF FRAMES TR2 SECTOR 1, STARTING FRAME 48, 1 FRAMES, CONTEXT 256 '.. (f) Fig. 2 (Continued.)(e) The transmission filter frequency response as a function of time (short time analysis), (f) typical transmission filter frequency response. (DB) I I I, 9 11 I I I Fig. 3. Typical transmission filter frequency response for a segment of vesicular breath sounds. 8o (b) Fig. 4. (a) 256 points of the source waveform of a segment of an asthmatic expiratory wheeze (bronchial). (b) The corresponding segment of the transmission filter frequency response during wheezing. Fig. 2 shows the results of analysis of an inspiratory segment of bronchial breath sounds. The source [Fig. 2(c) and (d)] was found to be. The estimated filter response [Fig. 2(e)] was noted to be essentially the same over the breath segment, indicating a primarily stationary filter response over the breath cycle. The typical response [Fig. 2(f)] was found to have a low-pass characteristic, with a frequency band range around - Hz (central frequency around Hz). Also, the lack of energy noted in the low-frequency range (- Hz) was attributed to analog highpass filters used prior to the recording of these sounds to avoid heart sounds, muscle, and skin noise. Fig. 3 shows the results for an inspiratory portion of vesicular breath sounds. The source characteristics were found to be indistinguishable from the bronchial breath sounds, but the typical transmission filter frequency response (Fig. 3) showed a distinctly smaller and lower band range (- Hz, central frequency around 1 Hz). This difference was expected since vesicular sounds are known to have a longer transmission path with inertial components than the bronchial sounds, which leads to their getting filtered to a greater extent. Fig. 4 shows the results for a segment containing an asthmatic expiratory wheeze. The estimated source [Fig. 4(a)] was found to

4 1136 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. II, NOVEMBER 1989 SbCTOR 1, STARTING FRAME 12, FRAMES, CONTEXT 256 YALUtS CR Fig. 5. A typical filter frequency response in an expiratory segment during early pneumonia (vesicular). be a periodic train of impulses (with period approximately.65 s) and the estimated source spectrum was essentially white. Only one frame (frame 48 with 256 points) of the estimated time waveform of the source is shown to illustrate its nature clearly. The estimated filter response [Fig. 4(b)] was found to be low pass, with a distinct resonance noted in frames 43-. The resonance frequency was at 166 Hz. This result is consistent with the accepted theory that wheezes are produced by periodic sources and have distinct resonances in the transmission path, probably at the airway walls. Fig. 5 shows the results for an expiratory segment of vesicular lung sounds during early pneumonia, before crackles appear. No change was noted in the estimated source waveform or spectrum. The transmission filter response (Fig. 5) (central frequency around 2 Hz) shifted to higher frequencies compared to normal vesicular breath sounds (central frequency of 2 Hz in Fig. 5 versus 1 Hz in Fig. 3). These results are consistent with the understanding of early pneumonia being caused by congestion of pulmonary blood vessels and excess fluid in the interstitial space, which probably causes higher frequency sounds to be transmitted due to consolidation of the lung. Fig. 6 shows the results for an expiratory segment of lung sounds during late pneumonia, containing coarse crackles. The estimated source waveform [Fig. 6(a)] were noted to have impulsive bursts corresponding to the crackles. This may be explained by the phenomenon of bubbling of fluids that find their way into the airways. The estimated source spectrum was found to be white. The transmission filter response [Fig. 6(b)] shifted to higher frequencies (central frequency around 2 Hz). Hence, in pneumonia, both the source and transmission filter are affected, the filter response starting to be affected first during early (probably interstitial) stages of the disease. The above results show that there is significant correlation between the expected associated phenomenon of the diseases and the affected source and transmission filter characteristics of the computed model. Diseases affecting the airway like advanced pulmonary edema, fibrosis, asthma, etc., may be diagnosed and categorized using the estimated source characteristics. Diseases affecting the pleural cavity and interstitial space like early edema, pneumonia, or pleural effusion may be diagnosed using the transmission filter characteristics. Since the source and transmission characteristics can be separated, the distinguishability between specific diseased conditions may become possible. For example, both pleural effusion and emphysema lead to attenuated lung sounds I - (b) Fig. 6. (a) The estimated source of an expiratory segment during late pneumonia. (b) The corresponding transmission filter frequency response (typical). at the chest wall. But, while the first affects the transmission path of lung sounds, the second affects the source significantly. Another useful situation is the detection of edema and pneumonia at an early stage using the transmission filter response changes, which occur before advanced stages [4] are reached (more work confirming this is pending publication). The ability to separate the characteristics should increase the sensitivity of detecting such changes. Such early detection could lead to more effective treatment of the situation, and also avoid any drastic measures and morbidity at advanced stages. While some of these applications still need to be clinically proven, the potential advantages make this ability to separately estimate source and transmission characteristics considerably significant. V. CONCLUSIONS In this work, the application of autoregressive analysis to separate the source and transmission characteristics of lung sounds is described. The hypothesized model is supported by using lung sounds from known lung conditions, and comparing the obtained source and transmission characteristics to expected characteristics in these conditions of the lung, and pointing out the similarity.

5 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 36, NO. 11, NOVEMBER Application of the AR analysis approach in order to separate the source and transmission characteristics is the specific contribution of this work and is distinct from other approaches like those by Cohen et al. [5]. More extensive experimentation is now needed to further study this model and its clinical application. REFERENCES [I] L. R. Rabiner and R. W. Schafer, Digirul Processing of Speech Signals. Englewood Cliffs, NJ: Prentice-Hall, [2] V. K. Iyer, P. A. Ramamoorthy, Y. Ploysongsang, and H. Fan, Reduction of heart sounds from lung sounds by adaptive filtering, IEEE Trans. Biomed. Eng., vol. BME-33, pp , Dec [3] R. L. H. Murphy, Jr., A simplified introduction to lung sounds, teaching audio tape cassette. [4] V. K. Iyer, P. A. Ramamoorthy, and Y. Ploysongsang, Analysis of lung sounds during pulmonary edema in dogs, in Proc. 8th IEEEI EMBS Con$, vol. 2, 1986, pp [5] A. Cohen and D. Landsberg, Analysis and automatic classification of breath sounds, IEEE Trans. Biomed. Eng., vol. BME-31, pp , Sept A Correction Procedure for the Asymmetry of Differential Pressure Transducers in Respiratory Impedance Measurements RAMON FARRE, DANIEL NAVAJAS, RENE PESLIN, MAR ROTGER, AND CLAUDE DUVIVIER Abstract-The usual setup for measuring respiratory input impedance requires a differential pressure transducer attached to a pneumotachograph. As, up to now, no data correction procedure has been devised to account for transducer asymmetry, a highly symmetrical transducer is required to obtain reliable impedance dala. In this communication, a general model for the measuring system is presented. Its main feature is that differential pressure transducers are modeled as two input-one output systems. From the theoretical model, we defined a dynamic calibration and data correction procedure. This was tested using highly asymmetrical transducers (common-mode rejection ratio between 45 and 27 db) to measure the impedance of two respiratory analogs. The latter were linear resistance (R), inertance (I), compliance (C) series models simulating a normal subject ( R = 3.47 hpa. s. I-, I = 1.45 Pa. s2. I-, C = 18.6 ml. hpa- ) and an obstructive patient ( R = hpa. s. I-, I = 1.28 Pa. sz. I-, C = 18.5 ml. hpa-i). Results obtained applying the devised procedure (errors in R, I, and C always less than 4 percent) show that respiratory input impedance can be adequately measured if data are corrected for transducer asymmetry. INTRODUCTION Respiratory mechanical impedance ( ZRs) is commonly obtained by connecting the subject s airway to a pressure generator, and by Manuscript received May 31, 1988; revised April 17, This work was supported in part by Hispano-French Integrated Action Grant 18/8 and by CAICYT Grants 3132/83 and The work of R. Farr6 at Unit6 14 INSERM, Nancy, France was supported in part by CIRIT (Generalitat de Catalunya). R. Farrk, D. Navajas, and M. Rotger are with the Laboratori Biofisica i Bioenginyeria, Facultat de Medicina, Universitat de Barcelona, 28 Barcelona, Spain. R. Peslin and C. Duvivier are with Unit6 14 de Physiopathologie Respiratoire (INSERM), Vandoeuvre-lts-Nancy, France. IEEE Log Number relating airway pressure to the pressure drop (A P ) across a reference impedance (pneumotachograph ZpN) placed in series with the subject. To avoid interfering with the subject s breathing, ZpN is usually small compared to ZRs. It follows that A P, as measured with a differential pressure transducer, is but a small fraction of the common pressure applied to both inputs of this transducer. An accurate estimate of ZRs therefore requires a high dynamic symmetry of the transducer, as defined by its common-mode rejection ratio (CMRR): CMRR is defined as CMRR =. log ( P/A P ) where AP is the spurious differential pressure recorded by the transducer when the same pressure P is simultaneously applied to both inputs of the transducer. It has recently been shown that, with common pneumotachographs, substantial errors are made in respiratory parameters when CMRR is less than db at Hz, par- ticularly when ZRs is high [5]. So far, transducer asymmetry, which increases with frequency, has been viewed as an insuperable circumstance [2], [5], and no correction scheme has been devised to eliminate the corresponding error. Indeed, transducers have always been modeled as one input-one output systems, with a single frequency response [2],[7], while dynamic asymmetry can only be accounted for by a two input-one output model. As a consequence, accurate impedance data can only be obtained on a limited frequency range using the most symmetrical transducers presently available. In this paper, we present a general theoretical model of the most usual impedance measuring system taking into account the asymmetry of differential pressure transducers. From this model, we propose a simple dynamic calibration procedure which allows correction of impedance data for transducer asymmetry. THEORY Fig. l(a) shows a diagram of the setup generally used to measure respiratory input impedance. All magnitudes are in the frequency domain and their dependence on frequency is omitted. The excitation pressure generated by the loudspeaker is recorded at the mouth (PE) with a differential pressure transducer (PTI). The excitation flow ( V, ) entering the respiratory system is measured as the pressure drop (A P = PI - P2) across a pneumotachograph (PN) by means of another differential pressure transducer (PT2). Thus, from the recorded signals Sp and Sv, a measured impedance value ZM ZM = sp/sv (1) is calculated for the true respiratory impedance ZRS defined as The mouthpiece is, in fact, included in the respiratory system. Its small influence is eliminated by correcting ZRs for the known impedance of the mouthpiece. The main feature of the following theoretical approach is to consider a differential transducer as a linear system with two inputs and one output [Fig. l(b)]. Thus, a transducer is characterized by two associated transfer functions [I] corresponding to the positive and negative pressure ports: Hp and HP for PTI, H; and Hi for PT2. Each of these transfer functions accounts for the relationship between the signal provided by the transducer and the corresponding pressure applied to one of its ports while the other is maintained at the atmospheric ground level. Therefore, from the block diagram in Fig. l(b), the recorded signals Sy and Sp are Sy = PIH; i- P2H; Sp = PEHp. (3) Sp does not depend on HP since the negative port of PT1 is open to the atmosphere. This does not mean that the response of PT1 is unaffected by the dimensions of the tubing attached to the negative port. In fact, as occurs with a compliant-membrane differential pressure transducer with connecting tubes [SI, the transfer function 1989 IEEE