Proceedings of the 26th Annual International Conference of the IEEE EMBS San Francisco, CA, USA • September 1-5, 2004

Detection of One Lung Intubation by Monitoring Lungs Sounds L. Weizman, J. Tabrikian and A. Cohen Department of Electrical and Computer Engineering, Ben Gurion University of the Negev, Beer-Sheva, Israel

Abstract— Analysis of lungs sounds for monitoring and diagnosis of pulmonary function is well known. One of the applications of this method is detection of One Lung Intubation (OLI) during anesthesia or intensive care. In this paper, an algorithm for detection the One-Lung ventilation situation from the lungs sounds is presented. The algorithm assumes a MIMO (Multiple Input Multiple Output) system, in which a multi-dimensional AR (Auto-Regressive) model relates the input (lungs) and the output (recorded sounds). The unknown AR parameters are estimated, and a detector based on the estimated eigenvalues of the source covariance matrix is developed, in order to detect one lung ventilation situation. Testing the algorithm on real breathing sounds, which were recorded in a surgery room, shows more than 90% accuracy in OLI detection.

of respiratory sound signals, in application to diagnosis and monitoring of pulmonological malfunctioning, is well known in the literature [4],[5],[6]. This technology, when applied to anesthesia and intensive care, has the potential of providing reliable, direct, easy to apply and relatively inexpensive solution to the OLI problem. The concept of this technology is presented in this paper. II. EXPERIMENT STRUCTURE In order to examine the suggested method and to check the possibility of developing a monitoring tool in the future, a database of recorded breathings was established. The database was composed of 24 patients which were recorded in a surgery room in both situations: during correct ventilation, when the tip of the tube is placed above the carina, and during a situation of OLI when the tip of the tube is under the carina and only one lung is ventilated. During the experiments the microphones were attached to the patients’ back, as shown in Fig. 1, recorded the breathing sounds of the patients in both situations. The ventilations were performed manually and not mechanically, in order to achieve higher signal-to-noise ratio in the recorded sounds, and the real position of the tube was validated each time by fiber-optic. The experiments were performed in the main surgery room of medical center Soroka - Israel, during the anesthesia part in the beginning of the surgery.

Keywords— AR, LUNGS, MIMO, OLI

I. INTRODUCTION Ventilation of patients under general anesthesia or in intensive care, is performed by an endotracheal tube, which is placed in the trachea. The location of the tip of tube is critical: it should be placed, and maintained above the bifurcation. A correct position of the tube, in which both lungs are ventilated, is called Tracheal Intubation (TRI). If the tube is misplaced or shifted due to patient movements, cases of One Lung Intubation (OLI) may occur. Prolonged cases of OLI should be avoided since it may cause insufficient oxygenation and may damage the non-ventilated lung. OLI was found to be the most probable cause of desaturation and the forth cause of malfunction during anesthesia [1]. There are currently several means for OLI detection: • The stethoscope is the simplest mean but it requires high attention and its reliability is low. • The Capnograph, which measures the exhaled amount of CO2 was proved to be an unreliable method for detection of OLI [1]. • Oxygen saturation, measured by pulse oximeter is one of the most widely used method today [2],[3] and the most reliable one, but its results are provided with latency of 2 to 5 minutes, which may be too long to prevent damage. Although the above methods are used today to detect OLI, there is currently no reliable method, which provides a realtime OLI detection. The proposed method for OLI detection is based on analyzing the breathing sound signals. An algorithm for detection of the number of ventilated lungs from the recorded breathing sounds is developed and tested. Analysis

0-7803-8439-3/04/$20.00©2004 IEEE

Fig. 1 Experiment structure – The locations of the microphones in the back of the patient.

917

Authorized licensed use limited to: Hebrew University. Downloaded on June 20,2010 at 13:23:33 UTC from IEEE Xplore. Restrictions apply.

A is an L × ML matrix defined as:  a 11 T ⋅ ⋅ ⋅ a 1L T    ⋅   ⋅ A= ⋅ ⋅ ,   ⋅   ⋅ a T ⋅ ⋅ ⋅ a T  LL   L1

In order to detect OLI cases from the recorded sounds, processing of the breathing sound signal is required. A simple analysis of comparing the amplitude of the recorded sounds in right and left sides has already been proposed but it was not reliable enough. The reason for this is that each one of the microphones records sounds generated by both lungs, and therefore even during OLI situation all of the microphones record the breathings. Analysis on dog’s breathing sounds has also been performed [6] in which classification by LPC and PARCOR features provided reliability of 83% in OLI cases detection. Because of the fact that the experiments were performed on dogs, no conclusion about human being can be obtained.

where aij is a M × 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]. C is a L × K matrix whose elements, cij, relate the sensor i with the source j. Finally, e[n] is a L × 1 vector that representing the additive white Gaussian noise. We assume that the noise and source signals are independent, zero mean, Gaussian with covariance matrices σ 2 I and I respectively. The last assumption is done without loosing of generality, because the covariance of the sources is controlled by the matrix C, as it can clearly be seen from (3). As a result, we obtain that the conditional distribution of y[n]/u[n] is Gaussian: y[n]/u[n] ~ N ( Au[n], R), where R = CC T + σ 2 I . We can

III. PROBLEM FORMULATION In order to determine the number of lungs that are ventilated from analysis of the recorded sounds, an AR (Auto-Regressive) model that relates the lungs and the microphones was assumed. This model was chosen because it is commonly used in applications of speech and audio processing and its computational complexity is relatively simple. In addition, a high order AR model can be a good approximation to ARMA (Auto-Regressive Moving Average) model [7]. In the first stage, we assume that each lung is a point source, and develop an OLI detector for this case. In the second stage the point source assumption is removed and the algorithm is modified accordingly, as will be discussed in section V. Our goal is to estimate the number of sources from the received signals by the sensors. Let K and L denote the number of sources (lungs) and sensors (microphones), respectively (K
x[ n] = [x1 [ n] x 2 [n] ⋅ ⋅ ⋅ x K [ n]] . T

note that the unknown parameters: A,R, σ 2 ,M and K are need to be estimated from a set of N measurements, y[1],…,y[N], and its past values, u[1],…,u[N]. IV. THE ML ESTIMATOR In order to determine the number of sources, K, we need first to estimate the unknown matrices, A and R, from the N samples of the data: y[1],…,y[N]. For this purpose, the Maximum–Likelihood (ML) estimator is used. The ML estimator of the matrices A and R, is obtained by maximizing the conditional probability density function (pdf) of the output samples given its past values, which is: f (y[1],..., y[ N ] / u[1],..., u[ N ]; R, A) = (7)

(1)

And y[n], the L × 1 vector of measurements is defined by:

y[ n] = [ y1 [ n]

y 2 [n] ⋅ ⋅ ⋅

y L [ n]]

T

(2)

[

]

u[n] = u1T [n] u T2 [n] ⋅ ⋅ ⋅ u TL [n]

T

N

1

(2π )

LN / 2

N/2

∏ exp{−

(y[n] − Au[n]) T R −1 (y[n] − Au[n])}

1 2

R The log-likelihood function can be maximized by equaling its derivations with respect to R and A and solving two matrix equations. This process yields:

The relation between the source signals and the measurements is given by a MIMO (Multiple-Input Multiple-Output) AR model: (3) y[ n] = Au[ n] + Cx[ n] + e[ n], where u[n] is a ML × 1 vector defined as follows:

(6)

and

(4)

 N  N  T T ˆ A ML =  ∑ y[ n]u [ n]  ∑ u[ n]u [ n]   n =1  n =1  ˆ R ML =

and ui[n] is a M × 1 vector which contains the past values of the i-th sensor, yi[n], up to sample M: T u i [n] = [ y i [n − 1] y i [n − 2] ⋅ ⋅ ⋅ y i [n − M ]] . (5)

n =1

N

1 N

∑ y[n]y

T

n =1

ˆ [n] − A ML

N

1 N

−1

∑ u[n]y

(8a)

T

[n]

n =1

(8b)

Substituting the maximum likelihood estimates (8) in the log-likelihood, with some straightforward manipulations, we obtain: log [ f (y[1],..., y [ N ] / u[1],..., u[ N ] )] = (9 ) LN N 1 ˆ − log( 2π ) − log R − NL 2

2

2

918 Authorized licensed use limited to: Hebrew University. Downloaded on June 20,2010 at 13:23:33 UTC from IEEE Xplore. Restrictions apply.

samples each, with 80% of overlapping. An arbitrary AR order of 15 was set, and the unknown matrices A and R were estimated for each window. Fig. 3 shows the second highest eigenvalues of Rˆ as a function of time. As it can clearly be seen from Fig. 3, we can easily decide between OLI and TRI cases, by the value of the second highest eigenvalue in every breathing. The results of the proposed algorithm were consistent over 24 experiments made on patients as explained in section II.

In order to be avoided from over estimating of K and M, a penalty function that depends on the free parameters of the model should be added to (9). This function can be computed by model order selection criteria, such as MDL [8],[9] and AIC [10]. Simulation results, assuming the proposed model and implementing these criteria have shown good results in estimating both the AR order model, M, and the number of sources, K. V. MODIFICATION OF THE SOLUTION FOR REALISTIC MODEL It can be noted that the only expression that depends on K and M in (9) is | Rˆ |, which is actually the product of the eigenvalues of Rˆ . According to [9], some of these eigenvalues represent the energy of the source signals (marked as li) and some of them represent the noise level (marked as σ 2 ). According to the selected model and [9], under each hypothesis (OLI or TRI) the eigenvalues of R are expected to be: TRI : l1 , l 2 , σ 2 , σ 2 ,..., σ 2 , (10a)

4

x 10

2

$!!#!!"

.

(10b)

L −1

Two Lungs

-1

5

10

15

4

x 10

One Lung

2

20

Time [Sec.]

25

30

35

40

35

40

35

40

35

40

Lower left microphone

Two Lungs

0

-2

5

10

4

x 10

One Lung

5

15

20

Time [Sec.]

25

30

Upper right microphone

Two Lungs

0 5

As it can clearly be seen from (10), the second highest eigenvalue of R can be used as an indicator for OLI situation. Because of the expected mismatches between the selected model and the real one, even during cases of OLI the second highest eigenvalue of R is higher than the noise level, σ 2 . The reason for this, is that the lungs are not point sources as they were treated in the selected model. In fact, each lung is composed of several distributed and independent point sources whose energies are distributed over all of the eigenvalues of R . Therefore, even during OLI situation the second highest eigenvalue is above the noise level. During OLI situation only one lung is ventilated and therefore, the point sources are less scattered than in case of TRI where both lungs are ventilated. This fact causes the energy of these point sources to be less scattered over the eigenvalues of R . Therefore, in OLI case the second highest eigenvalue of R value is smaller than in TRI case. This is the reason why choosing the second highest eigenvalue as a detector provides a reliable detection of OLI situations, as will be shown in the next section.

5

10

4

x 10 4

One Lung

2

15

20

Time [Sec.]

25

30

Lower right microphone

Two Lungs

0

-2

5

10

15

20

Time [Sec.]

25

30

Fig. 2 Some recorded breathing cycles by the four microphones, OLI and TRI cases are included. 6

x 10 5

Second highest eigenvalue of R

L−2 2

OLI: l1 , σ , σ , σ ,..., σ 2

Upper left microphone

0

$! !#!! "

2

One Lung

1

VI. RESULTS

4

3

2

One Lung

1

0

Fig. 2 shows a few breathing cycles, of both OLI and TRI situations, recorded by the four microphones. As it can be seen from this figure, determination between OLI and TRI cases by only the amplitude of the recorded sounds is not a simple task, as explained in section II. The breathing signals were limited to cut-off frequency of 4kHz and the data were divided into windows of 2000

Two Lungs

5

10

15

20

Time [Sec.]

25

30

35

ˆ as a function of time. OLI and TRI cases Fig. 3 Second eigenvalue of R can clearly be determined.

The curves that represent the distribution of the second highest eigenvalue under each hypothesis are shown in

919 Authorized licensed use limited to: Hebrew University. Downloaded on June 20,2010 at 13:23:33 UTC from IEEE Xplore. Restrictions apply.

40

VII. DISCUSSION AND CONCLUSIONS

Fig. 4. As a result from these curves we can draw the Detection Error Tradeoff (DET) curve, which is the probability of miss, Pmiss, as a function of probability of false-alarm, PFA. The DET curve provides information about the receiver performance, where each point on the curve shows the PFA and Pmiss for a given threshold. In Fig. 5 we can see the DET curve of the proposed decision system. The Equal Error Rate (EER) point which is defined as the point on the DET curve where Pmiss= PFA is 2.7%.

This paper endeavored to take the OLI detection research one step forward by showing the concept of a new automatic system for detection of OLI. An algorithm for detection of OLI by monitoring lungs sounds was developed and tested. It has been shown that assuming a MIMO AR model and selecting the second highest eigenvalue of the residual covariance matrix as a feature proves itself as a reliable method for detection of OLI. Implementing the method into a real-time machine in order to detect OLI while the patient is mechanically ventilated, is supposed to give better results. The reason for this is that while the patient’s lungs are mechanically ventilated, the variance of the amplitude of the breathing signals is much smaller than it is when using manually ventilation (as performed in our experiments). This fact should cause to better performances of our method.

Distribution of second highest eigenvalue under each hypothesis 0.35

Probability Distribution

0.3

0.25

0.2

ACKNOWLEDGMENT

0.15

The authors wish to thank Prof. G. Gurman, Dr. S. Teiman and Dr. A. Zlotnik from the medical center Soroka – Israel, for their major contribution with collecting the real data and supporting the medical part of the paper.

0.1

0.05

0

0

REFERENCES

2 4 6 8 10 12 Normalized values of second highest eigenvalue of R

[1] R. K. Webb, J. H. van der Walt, W. B. Runciman, J. A. Williamson et. al. , “Which Monitor? An Analysis of 2000 Incident Reports,” Anesthesia and Intensive Care, Vol. 21, pp. 529-542, 1993. [2] W. B. Runciman, R. K. Webb, L. Barker and M. Currie, “The Pulse Oximeter: Applications and Limitations – An Analysis of 2000 Incident Reports,” Anesthesia and Intensive Care, Vol. 21 pp.543-550, 1993. [3] T. A. Webster, “Now that we have Pulse Oximeters and Capnographs, we don’t need Precordial and Esophageal Stethoscope,” J. of Clinical Monitoring, Vol. 3, pp. 191-192, 1987. [4] A. Cohen and D. Landsberg, “Analysis and Automatic Classification of Breathing Sounds,” IEEE Trans. Vol. BME-31, pp. 585-590, 1984. [5] A. Cohen and A. Berstein, “Acoustic Transmission of the Respiratory System using Speech Stimulation,” IEEE Trans., Vol. BME-38, pp. 126-132, 1991. [6] G. Sod-Moriah, A. Cohen and G. Gurman, “Detection of One lung Intubation Incidents in General Anesthesia and Intensive Care,” Proc. of the 13th Int. Conf. BIOSIGNAL 96, pp. 282-284, Brno, Czech Republic, 1996. [7] B. Porat, Digital Processing of Random Signals, Theory and Methods, Prentice Hall, Englewood Cliffs, NJ, 1994. [8] J. Rissanen, “Modeling by shortest data description,” Automatica, Vol. 14, pp. 465-471, 1978. [9] M. Wax and T. Kailath, “Detection of signals by information theoretic criteria,” IEEE Trans. Acoust. Speech, Signal Processing, Vol. ASSP-33, pp.387-392, 1985. [10] H. Akiake, “A new look at the statistical model identification,” IEEE Trans. Automat. Contr. Vol. AC-19, pp. 716-723, 1974.

Fig. 4 The distribution of the second eigenvalue. The left curve represents the distribution of the normalized values when OLI situation occurs, and the right curve represents the values on TRI situation.

DET Curve

0

Pmiss

10

-1

10

-2

10

-2

10

-1

PFA 10

0

10

Fig. 5 The DET of a classifier based on the second highest eigenvalue of estimated R.

920 Authorized licensed use limited to: Hebrew University. Downloaded on June 20,2010 at 13:23:33 UTC from IEEE Xplore. Restrictions apply.