| JOURNAL HOME | CME HOME | THIS MONTH | PAST ISSUES | ETOC | COLLECTIONS |
| AUTHORS | REVIEWERS | EDITORIAL BOARD | FEEDBACK | RSS | HELP |
| ||||||||||||||
| ||||||||||||||
|
|
|
|
||||||||||||
|
||||||||||||||
From the Divisions of *Pediatrics; Anesthesiology and Critical Care, Soroka Medical Center; and Electrical and Computer Engineering Department, Ben Gurion University of the Negev, Beer-Sheva, Israel.
Address correspondence and reprint requests to Shai Tejman-Yarden, MD, Division of Pediatrics, Soroka Medical Center, Faculty of Health Sciences, Ben Gurion University of the Negev, Beer Sheva 84101, Israel. Address e-mail to teyman{at}bgumail.bgu.ac.il.
 |
Abstract |
|---|
 Top Abstract IntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
METHODS: Lung sounds were collected from 24 adult surgical patients scheduled for routine surgical procedures. Four piezoelectric microphones attached to the patients' backs were used to sample lung sounds during induction of anesthesia and endotracheal tube positioning. To achieve OLI, the endotracheal tube was inserted and advanced down the airway so that diminished or no breath sounds were heard on the left side of the chest. The tube was then withdrawn stepwise until equal breath sounds were heard. Fiberoptic bronchoscopy confirmed the tube's final position. Acoustic analyses were preformed by a new algorithm which assumes a Multiple Input Multiple Output system, in which a multidimensional Auto-Regressive model relates the input (lungs) and the output (recorded sounds) and a classifier, based on a Generalized Likelihood Ratio Test, indicates the number of ventilated lungs without reconstructing the original lung sounds from the recorded samples.
RESULTS: This algorithm achieved an OLI detection probability of 95.2% with a false alarm probability of 4.8%.
CONCLUSION: Higher detection values can be achieved at the price of a higher incidence of false alarms.
|
Introduction |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
Auscultation is the most common method for assessing the position of the ET tube. Though this technique has many advantages, such as ready availability and low cost, it is operator-dependent and has an unacceptably high margin of error that can reach 60% (4–7). Pulse oximetry, by which most cases of OLI are detected, sets off its alarm only after the patient has become symptomatic, and does not specifically indicate the actual cause of hypoxemia. Capnography, which was once suggested as a means to identify OLI, is not reliable for this purpose (1,8–11). New techniques for ventilation monitoring, such as online spirometry and acoustic reflactometry, are being developed and studied (12), and new studies in dogs using routinely monitored airway pressure and flow as inputs, show good results, with both 90% sensitivity and specificity for OLI detection (13).
Analysis of lung sound signals to diagnose and monitor the respiratory system is a well established method of physical diagnosis. Many studies have been conducted trying to analyze lung functions, ventilation status and different pathological states using lung sounds (14–16). On the basis of these works, we conducted an initial study in which acoustic lung sounds were sampled from patients undergoing lung procedures that required the insertion of a double-lumen tube (12). In that study, we sampled each patient for right, left, and bilateral lung ventilation, using two piezoelectric microphones, one on each side of the chest. The sounds samples were band-pass filtered and for every respiration. The left and the right energy envelopes were calculated from each microphone separately, assuming that each microphone samples its ipsilateral lung. The diagnosis of the ventilation status was based on the energy ratios between the left and the right, and OLI was identified when the detected energy ratio decreased below a certain threshold. This analysis algorithm yielded 92% recognition of selective right one-lung ventilation with a sensitivity of only 90%.
The results of that study corresponded to earlier studies on lung auscultation which presumed that gas passing through the Murphy's eye or through a narrow space between the ET tube and the bronchial wall during selective endobronchial intubation might generate contralateral sound signals (6). The use of a double-lumen tube with inflated balloons demonstrated that significant air leak is not the cause of the phenomenon and that each microphone attached to the chest actually samples both lungs, but with different ipsilateral and contralateral contributions. We concluded in that study that each lung does not produce a uniform sound; it produces various different sounds at different times and simultaneously receives transmitted sounds from the contralateral lung. The significance of these findings is that any algorithm based on the assumption that each microphone samples its ipsilateral lung independently and that the energy envelope represents only one source signal does not adequately reflect this acoustic system (12).
Hence, we developed a new algorithm which assumes a MIMO (Multi Input Multi Output) system, in which a multidimensional Auto-Regressive (AR) model relates the input (lungs) and the output (recorded sounds) and a classifier, based on a Generalized Likelihood Ratio Test, indicates the number of ventilated lungs (17,18). In the present study, we implemented this algorithm on sampled lung sounds from surgical patients having routine operations. The patients were intubated with a conventional single-lumen ET tube.
The aim of the study was to examine the accuracy of the new system for detecting OLI.
|
METHODS |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
Patients
Twenty-four adult surgical patients (ASA I and II) signed informed consent forms before participation in the study. All were scheduled for a routine surgical abdominal or orthopedic procedure that required the insertion of a single-lumen ET tube.
The Sampling System
The sampling system consisted of four Tommyscope acoustic sensors (KOL Medical Ltd. Innovative Medical Equipment, Ramat-Gan, Israel), an amplifier, an antialiasing filter, and a laptop computer, which was equipped with an A/D converter. The Tommyscope sensors are piezoelectric microphones with internal low-pass filters, which suit the spectral range of the breath sounds signal. The amplifier has a controlled gain for each channel and the laptop computer is equipped with a Keithley A/D converter that has 12 bit resolution (12).
Technique
Before anesthesia, the four Tommyscope piezoelectric acoustic sensors were placed on the patient's back to sample breath sounds. Two microphones were attached on each side of the back in parallel positions. Two microphones in-between the scapulae were placed on each side of the spinal cord and two, a few centimeters below them on each side of the back. After induction of anesthesia with thiopental and succinylcholine, using a method described by Heinonen et al. (19), the ET tube was inserted and advanced down the airway until diminished or no breath sounds were heard by the performing physician on the left side of the chest using a standard stethoscope. The ET tube was then withdrawn stepwise until distinct bilateral breath sounds could be heard, using only the stethoscope. At the end of the ET tube positioning procedure, the correct position of the ET tube was confirmed by fiberoptic bronchoscopy. During induction, tube positioning and bronchoscopic conformation of the tube's position, lung sound samples were collected. The samples were recorded in real-time for analysis at a later time in the signal-processing laboratory.
Signal Preprocessing
Processing of the sampled signal was performed using the MATLAB software (Mathworks Ltd.). The recorded sounds were sampled at 4 kHz to attenuate some of the irregular background noise outside of the spectral range for breathing signals. The recorded signals were band-pass filtered by a Butterworth filter with a bandwidth of 100 Hz–600 Hz. Because of the cut-off frequency of 600 Hz, down-sampling operation with a factor of 0.3 was performed (17).
Since the signal amplitude from each microphone depends on the particular location of the microphone on the patient's body, the anatomy of the particular patient, and the gain of the sampling system, each channel's output had a different signal amplitude, and the noise variance, differed among microphones. To solve this problem, each channel was normalized according to its noise level.
The preprocessing techniques were designed to reduce ambient noise associated with the operating room and the algorithm itself was robust enough to determine OLI without interference from the ambient noises present in the operating room (17).
Signal Processing and Breath Classification
The recorded sounds were analyzed by an algorithm which assumes a MIMO system, in which a multidimensional AR model relates the input (lungs) and the output (recorded sounds) (Fig. 1). In this model, each ventilated lung represents a sound source and the goal of the algorithm is to calculate the number of sources from the received signals, two sources for bilateral lung ventilation (TrI) and one source for one-lung ventilation (OLI). The number of the sources in each respiration is estimated by calculating a four-by-four matrix of values (derived from the four microphones), the 
matrix. The four eigenvalues (self-values) of this matrix represent the sound sources of the system (right lung, left lung, echoes, and ambient noise sounds). These values are calculated out of the recorded samples, without reconstructing each source separately. A brief description of the breath classification algorithm, originally introduced by Weizman et al. (18), is described in Appendix A.
|
OLI Determination
The four eigenvalues of the matrix represent the sound sources in the following manner: the highest eigenvalue represents the right lung, the second highest eigenvalue represents the left lung, and the third and the fourth eigenvalues (which are much smaller values) represent the echoes and the ambient noise. For OLI detection, a certain threshold of the second eigenvalue of the matrix or the ventilation discrimination index needed to be determined. Any value below this threshold would indicate OLI, and any value above it would indicate bilateral lung ventilation (17,18,20).
To set this threshold, a distribution curve of bilateral ventilation values and OLI values was calculated from the sound signals recorded from all 24 patients. Since the available database was small, in order for us to build a normalized histogram of the ventilation discrimination index from which this certain threshold could be determined, we used the leave-four-out method. Using this method, a database from 20 patients served as a threshold template to which data from four patients were referred. This scheme was repeated six times until a full histogram was built.
System Performance Assessment
The statistical analysis of this method is based on its performance, which is the probability to detect OLI against the probability to false alarm. This calculation is a function of the overlap between the values in the distribution curve and is based on the Detection Error Tradeoff (DET) performance curve of this system. To evaluate this system, for every threshold taken from the values, which overlap between the OLI values and the bilateral values, the ventilation discrimination index of each respiration was compared to the normalized distribution curved which was set accordingly. Two types of errors in OLI detection were calculated: Pmiss, (false-negative), the probability of a true OLI to be wrongly detected as bilateral ventilation, and Pfa (false-positive), the probability of bilateral ventilation to be wrongly detected as OLI. For every threshold, these probabilities were calculated and a DET curve, a common method to display these errors, was built. The DET curve indicates the device's performance, where each point on the curve shows the Pfa and Pmiss for a given threshold. The threshold of a real monitoring system should be calculated according to the requested sensitivity of the system, taking into consideration the allowed Pmiss of the system and the resulting Pfa.
|
RESULTS |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
was estimated and its eigenvalues were calculated. In Figure 2, the original lung sound samples before and after the band-pass filtering are presented. In Figures 3 and 4, the filtered lung sounds and the analyses of two trials are demonstrated. Box A shows the filtered sampled lung sounds from each microphone on the patient's back. Box B shows a diagram of the four eigenvalues of the matrix, representing the source signals (the right lung, the left lung, the echoes and the ambient noise). Box C shows a diagram of the second highest eigenvalue of the matrix, the ventilation discrimination index, representing the left lung.
|
|
|
matrix can serve as an indicator for the ventilation status. Although most of the values are distinct, some were found to overlap between the two curves. As a result, this is not an absolute system and there are potential false-positive and false-negative results, depending on the system's sensitivity setting.
|
Figure 6 shows the calculated DET curve for this system. The Equal Error Rate point, the point on the DET curve where Pmiss = Pfa, in this system is 4.8%. The significance of this value is that by using the proposed method an OLI correct detection probability of 95.2% (100% – Pmiss of 4.8%) and Pfalse of 4.8% can be achieved. Moreover, by using this curve, it is possible to set the system to a Pmiss of 2% (detection probability of 98%), but with a Pfa of 9%.
|
|
DISCUSSION |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
In an initial study, acoustic lung sounds were sampled from patients undergoing lung procedures that required the insertion of a double-lumen tube. Lung sound samples were collected by two microphones, one on each side of the chest. In that study, the analysis of the lung sounds, done by an algorithm, which assumed that each microphone samples its ipsilateral lung, yielded 100% recognition of left OLI with a sensitivity of 100% and 92% recognition of right OLI with a sensitivity of 90%. Careful analysis of the sound samples collected in that study showed that each chest microphone actually sampled sounds from both lungs, but with differential ipsilateral and contralateral contributions. Moreover, we found that each lung does not produce a uniform sound but, rather, a series of various sounds at different times during each respiration (12).
In light of these findings, we developed the algorithm presented in this article. This algorithm analyzes each respiration as a MIMO system in which a multidimensional AR model relates the input (lungs) and the output (recorded sounds) without reconstructing the original lung sounds separately. To assess this algorithm, we conducted a study done on 24 patients undergoing routine surgical procedures and intubated with a standard endobronchial tube. Using a technique described by Heinonen et al. (19), we managed to sample each patient during right OLI and bilateral ventilation. Analysis of the lung sounds by this algorithm shows that OLI can be detected with a probability of 95.2% and a probability of 4.8% of false alarm. Moreover, we found out that this system is adaptable; the alarms can be set to the desired sensitivity for each surgical patient, taking into consideration that if the alarm is set to a high sensitivity rate there will be a correspondingly higher rate of false alarms.
Clinically, the system presented here is a prototype for a new device that can monitor ventilation during endobronchial intubation and detect OLI. It is reliable and noninvasive. It yields immediate results and can be used to identify OLI during surgery. Since the assessed variable is not the ET tube's position but rather the patient's ventilation status, it may be possible to use this algorithm to detect other surgical life-threatening events that lead to unequal breath sounds, such as esophageal intubation, massive unilateral atelectasis, or pneumothorax. The most important feature of this system is its ability to detect any ventilation incident as it occurs, while the patient is still well oxygenated and stable. Moreover, since the system is attached to the patients' back, in parallel position to the electrocardiograph monitoring system, it could be designed in the future, to combine these two monitors into one, with both modalities sampled, by the same sensors.
This study is a preliminary, small scale study that was designed to confirm the ability of this algorithm to detect OLI. Since each sample taken from the recorded data was used both for testing the results and as a baseline, each sample separately had a large impact on the final results. We assume that this is the reason for the significant fraction of overlap between the bilateral ventilation and the OLI curves in the distribution histogram presented in Figure 5. For further assessment of this monitoring system, we have begun a new study that will be conducted on a large number of patients with diverse acoustic properties (influenced by age, body mass indexes, and different pathologies), who need a double-lumen tube ventilation. In this study we will sample the patients during right OLI, left OLI and bilateral lung ventilation, while ventilated with the respirator at the same setting of inadvertent OLI, before and during surgery. Since the baseline will be set according to a large database, we assume that the portion of overlap between the two curves in the distribution histogram will be smaller, (depending on the age and the body mass index), thus the probability to detect OLI will be higher, with a smaller portion of false-positive. Additionally, we intend in this trial to sample patients with different pulmonary pathologies to assess the ability of this algorithm to diagnose respiratory status in spite of the interfering lung sounds, such as crackles or ronchi or altered lung sounds due to the underlying illness.
A limitation to the results in the present study is the fact that the lung sounds were collected during induction, a quiet phase, while no surgical, noisy procedures took place. For accurate assessment of the system's performance, lung sound samples in the new study are to be sampled before and during surgery, as mentioned above. The aim is to assess the overlap of noise and the lung sound in this acoustic system and to learn what measures are needed to cancel noise and accurately analyze this sound system.
In conclusion, we have evaluated a new method for monitoring ventilation using analysis of breath sounds. The data obtained under controlled conditions of ET tube placement during induction of general anesthesia suggest that this method can be used to discriminate right mainstem intubation and resulting one-lung ventilation from bilateral ventilation with a high degree of reliability. The system requires additional validation to detect right mainstem intubation in other patient populations and clinical conditions which result in unequal ventilation.
|
APPENDIX A |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
[n] y2(M)[n] ... yL(M)[n]] and A is an L x ML matrix where aij is an M x 1 vector, which relates the n-th sample of the i-th sensor, yi[n], with the past values of the j-th sensor, yj[n– 1], ..., yj[n – M] when K and L denote the number of sources (lungs) and sensors (microphones), respectively (K < L). C is an L x K matrix whose ij-th element, cij, relates the sensor i with source j. Finally, e[n] is an L x 1 vector that represents the additive white Gaussian noise. The algorithm assumes that the noise and source signals are independent, zero mean, Gaussian with covariance matrices 
2I and I, respectively. The last assumption is employed with no loss of generality, because the covariance of the sources is a function of the matrix C. As a result, the conditional distribution of y[n]|y(M)[n] is Gaussian: y[n]|y(M)[n]
N(Ay(M)·[n],R), where R is defined as: R = CC + 2I. It is assumed that the initial conditions are zero (e[n], x[n] = 0 for n < 0) and that the input and noise signals are stationary. Successful estimation of K, the number of sources (lungs) is the key for OLI detection. For this purpose the maximum–likelihood estimator is used. The maximum–likelihood estimator of the matrices A and R is obtained by maximizing the conditional probability density function of the output samples given its past values. The log-likelihood function can be maximized by equating its derivations with respect to R and A, solving two matrix equations. This process yields:
|
|
and
|
|
The classification between OLI and TrI from A and R is done by a Generalized Likelihood Ratio Test (GLRT) which uses the second highest eigenvalue of the matrix as the detector. It can be noted that the only expression that depends on K and M is , which is actually the product of the eigenvalues of the matrix (20,21). According to Wax and Kailath (22), some of these eigenvalues represent the energy of the source signals (marked as li) and some of them represent the noise level (marked as 2). According to the selected model and Wax and Kailath, under each hypothesis (OLI or TrI) the eigenvalues of the matrix are expected to be:
|
|
and
|
|
As it can be seen from these formulations, the second highest eigenvalue of the matrix can be used as an indicator for OLI situation (the ventilation discrimination index). Since the lungs are not point sources as they were treated in the selected model, even during cases of OLI the second highest eigenvalue of the matrix is higher than the noise level 2. In fact, each lung is composed of several distributed and independent point sources whose energies are distributed over all of the eigenvalues of the matrix. During OLI situation, only one lung is ventilated and therefore, the point sources are less scattered than in case of bilateral ventilation where both lungs are ventilated. This fact causes the energy of these point sources to be less scattered over the second highest eigenvalues of the matrix. Therefore, in OLI case the second highest eigenvalue of the matrix value is smaller than in bilateral ventilation case. This is the reason why choosing the second highest eigenvalue as a detector provides a reliable detection of OLI situations.
|
Footnotes |
|---|
|
REFERENCES |
|---|
|
TopAbstractIntroductionMETHODSRESULTSDISCUSSIONAPPENDIX AREFERENCES |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
