Acoustical detection of coronary occlusions using neural networks

Acoustical detection of coronary occlusions using neural networks M. Akay and W. Welkowitz Biomedical Engineering 08855, USA

Department,

Rutgers

University,

PO Box. 909, Piscataway,

NJ

Received August 1992, accepted February 1993

ABSTRACT A nonlinear neural network classtfier was applied to noninvasive acoustic detection of coronary artery disease; the cbrsifier included a feature vector, derived porn diastolic heart souna!s, and a multi-layered network trained by the backpropagation. T?te feature vector is based on the linear prediction coeficients of the autoregressive method afler an adaptive line enhancement method was used as the input pattern to the neural network. Onehundred and twelve recordings (70 abnormal, 42 normal) were studied and the network was trained on a randomly chosen set of six abnormal and six normal patients. It was tested on a database consisting of 700 recordings to which it had not been exposed. 7’he network correctly identified 50 of the 64 patients with coronary artery disease and 32 of the 36patients without any coronary artery occlusions. i?ese results showed that this neural network is capable of distinguishing normalpatientsfrom abnormalpatients. In addition, the diagnostic capability of this approach is much better than any other available noninvasive approach. Keywords:

Coronary occlusion, diastolic heart sounds, neural networks

INTRODUCTION One-third of ail deaths are estimated to be due to coronary artery disease (CAD)‘-‘; for this reason, early detection of coronary artery disease is one of the most important medical research areas. The most reliable way to diagnose CAD is with an invasive procedure known as cardiac catheterization, in which a catheter is inserted into an artery (usually brachial or femoral) and advanced into the heart. Once in the heart, dye can be released to observe the cardiac structures, including the coronary arteries”. Although direct assessment of coronary occlusions is conclusive only through coronary angiography, this method is expensive”. A reliable noninvasive method for early detection of coronary artery disease is clearly desirable. Coronary stenoses have been shown to produce sounds due to the turbulent blood flow in partially occluded arteries4-“; these sounds are rarely heard, but are a sign of coronary stenoses”-I*. During d&stole, coronary artery blood flow is at its maximum and the sounds associated with turbulent blood flow arteries are corona through partially occluded loudest”g. If these signals could be re7 iably detected, they would provide a simple, noninvasive approach to the detection of coronary artery disease. Studies involving turbulent blood flow have been carried out in many components of the cardiovascular system and it has been widely reported that turbuCorrespondence and reprint requests to: Dr Metin Akay

lence produced by stenoses produces sounds due to the vibration of the surrounding structureslo-‘L. These sounds have been detected and analysed, and the results generally show that the high-frequency energy increased when the degree of stenosis was increased’&‘“. However, for severe obstructions, (above 95% occlusion), sounds may not be produced due to the very low blood flow. For small blockages, occlusions as small as 25% narrowing have been detected, thereby substantiating the use of this technique as an assessment of arterial narrowing in vessels of the neck, thorax, and abdomen”~“‘. The auditory corn onent associated with coronary stenosis is simi Par to that found in partially occluded carotid arteries, but is much more attenuated by the intervening tissue”“. It is also masked by comparatively loud valve sounds. These loud valve sounds can be eliminated by isolating diastolic portions of the acoustic signal using a time window synchronized with the cardiac cycle q-y. Figure 1 shows the isolated diastolic segment. Previous angioplasty and normal/abnormal studies by the authors’ group evaluated the diagnostic capability of the acoustic approach in detecting coronary artery disease. In one study, heart sounds of 23 angioplasty patients were recorded before and after angioplastic surgery. In addition, two normal patients were included in the database7--x. The with corrective spectral features most associated surgery were a consistent decrease in the highfrequency ower above 300 Hz7-s, the relatively low frequency P100-200 Hz) peak did not change with the

@ 1993 Butterworth-Heinemann for BE.5 0 I3 I -5425/93/O(i46’)-0.i J. Biomed. Eng. 1993, Vol. 1.5,November

469

Acoustical detection of coronary occlusions using neural networks: M. Akay and W. Wekowi&

Figure 2

Ills

Figure 1

The isolated diastolic heart sound waveform

surgery and the amplitude of this peak was almost the same before and after angioplastic surgery. Though the exact cause is unknown, this peak may be due to low frequencies produced by aortic pumping or ventricle filling. Angioplasty generally produced a consistent shift away from the unit circle in the second and third ole pairs obtained with the autoregressive (AR) me tR od. A parallel normal/abnormal study showed that the percentage of spectral energy above 300 Hz differed between normal and diseased patients with the energy over 300 Hz being greater in diseased atients 5y7. A g ain, in almost all subjects, it was foun 1 that the second and usually the third poles of the normal patients were further from the unit circle than those of coronary artery diseased patient5-7. Although the results of our previous studies which applied the AR, autoregressive moving average (ARMA) methods to recordings showed considerable promise after adaptive filtering, further improvements may be necessary for a mass screening procedure for detection of coronary artery disease. Recently, we focused on the application of the neural networks to improve the diagnostic capability of the acoustical approach; in that study, we utilized an adaptive filter to reduce background noise in the diastolic heart sounds. The AR method was applied to the filtered diastolic heart sound recordings obtained from 100 patients to provide a lo-point analog feature vector. Then, a neural network was a plied to the feature vector. In order to capture fully a P1 relevant information related to the disease states of the patients, the nonlinear and multi-layered architecture of the neural network was utilized. A neural network was chosen since it has the intrinsic ability to learn from the feature vector and make less restrictive assumptions about the feature vector than other traditional statistical methods. Finally, a neural network is nonparametric and capable of forming highQni;nlinear decision boundaries in the feature space .

METHODS Feature vector selection for neural network When

470

using

neural

networks

to perform

J. Biomed. Eng. 1993, Vol. 15, November

pattern

Adaptive Line Enhancer

representative features should be classification, applied in order that the neural network may learn from those features. Determining the meaningful and representative feature vector is a very critical and important step since the learning process of the neural network depends on the feature vector. However, in practice, the biomedical signals are corrupted by background noise. For this reason, these signals are preprocessed so that the intrinsic features of the sources are extracted and randomness in the discrimination process is minimized before the actual classification process proceeds. Here we utilized the adaptive line enhancer (ALE) method to minimize the background noise contaminating the desired diastolic heart sounds associated with corona occlusions. The ALE method was chosen since it 7 oes not require a reference (noise) signal. The primary (input) signal, x(n), provides its own reference signal which is a delayed replica of itself, the only difference between the primary and the reference signal being the delay, d. If the delay is properly chosen, the noise in the reference signal, r(n), is uncorrelated with the primary noise signal, u(n). The output of the adaptive line enhancer estimates the desired signal y(n) as shown in Figure 2. The details of the ALE method have been described elsewhereg~‘g-2’. In this study, we utilized the least mean square (LMS) implementation of the line enhancement method. We selected the delay d = 5 for ALE (see Figure 3, R&2) since this value is large enough so that the noise, v(n), in the primary signal will be uncorrelated with u(n - d) in the reference signal. The ALE filter order was taken as m = 10 based on our previous findings (see Figure l(a), &f 2). After enhancing the diastolic heart sounds and eliminating the background noise using the ALE method, average autocorrelation function lags were calculated for each filtered recording (over 10 periods). Then, the linear prediction coefficients, a(m), of the filtered signal were calculated for each atient recording using the modified Yule-Walker MYW) AR method. These parameters were used to P provide a lo-point analog (for each recording) feature vector as the input pattern to the neural network. The MYW AR method has been shown to give accurate and efficient calculations of the prediction coefficients of the AR method 22P23. The overdetermined type MYW AR method was chosen because it shows better performance than other block processing AR methods when the poles of the filter function, A(z), are sufficiently close to the unit circle22,23. The extraction of the feature vector (the prediction coefficients of the MYW AR methods), a(m), can be carried out in two steps. The first step was to calculate the estimated autocorrelation functions (ACF), Z$y(k)2’. The second step involved the calculation of the MYW AR prediction coefficients from the follow-

Acoustical detection of coronary occhuionr using neural networks: M. Akay and W. WeUrow&

ing equation:

Ryy (0)

Ryy(l). ** Ryyb- 1) 4)

R;y(1)

... = ...

...

R,,(k- 1)

...

R,,b-m)

44

Ryy(4

(1) where k represents the highest order of the ACF and m represents the highest order of the AR process. This equation was solved by using the least squares algorithmZ2T2”. The details of the MYW AR method have been described elsewhere22*23. For the initial estimation, a(0) = 1 was chosen. A study using the MYW AR method has shown that for filter order, m, greater than 10, the minimal meansquare error as a function of filter order was relatively stable. Based on previous data (F&we 7a, b, I&$ 2) and our initial empirical findings, filter orders between 10 and 15 were judged sufficient to represent the signal records. Further, our results were insensitive to the filter orders within this range. For these reasons, a filter order of m = 10 and an autocorrelation order of k = 15 were considered adequate for representing the original signal.

Neural networks Recent developments in the field of artificial neural networks have made them a powerful tool for analysing signals. The application of artificial neural networks (ANNs) has opened a new area for solving problems not resolvable by other signal processing techniques. A number of ANN algorithms and their applications have been widely reported’5-‘8. Among the many neural network models multi-layer perceptrons which are considered the most useful learning modelsi5-is have been widely used in the biomedical field24,2”. In this study, a three layer network trained by the backpropagation algorithm will be used. This approach has also been found to perform well insolving a number of problems related to speech and visual pattern and recognition, synthesis recognitionis. Using a recursive updating algorithm, all connection weights starting at the output layer (nodes) and working back to the first hidden layer were updated as follows’“-‘s: A w,(t)

= n6jXj

w;li(t+ l)=

H$~((t)+~6jXj

(2) (3)

where B$( t) corresponds to the weight from node i of first layer or from an input to node j of second layer at the time t, xj corresponds the output of the unit of the secondary layer; v output of the j* corresponds to learning rate control constant; 6j corresponds to the error. In order to increase the speed of convergence, a momentum term will be added to the updating

equation. Details of the neural networks have been described elsewhere’5-18. In this study, a three layer network trained by the backpropagation algorithm was used. The target for the output was coded as 0 for the input pattern obtained from normal patient recording and 1 for the input pattern obtained from abnormal patient recording. The winner-take-all decision rule used in this study meant that the input pattern vector was set to the class belonging to the output node whose value is the biggest. Detection rate (sensitivity) was defined as the number of the correctly diagnosed abnormal patients divided by the total number of the abnormal patients in the test set. Specificity was defined as the number of the correctly diagnosed normal patients divided by the total number of the normal patients in the test set. The learning coefficient, 7, and the momentum factor, (Y, of the neural network were set at 0.5 and 0.1, respectively. Initial weights of the neural network were random. Details of the neural network have been described elsewhere’5-‘8. Patient

analysis

Patients were selected from those undergoing catheterization and/or angioplasty at the Cardiodynamics Laboratory of Robert Wood Johnson University Hospital. Diastolic heart sounds were recorded horn the 4th intercostal space on the chest of patients usin a specially designed high-sensitivity accelerometer in conjunction with a portable digital pulse code modulation (PCM) data recorder (TEAC, RD-1 10T). These sounds were recorded while the patients held their breath and were supine”“. The recorded data were loaded from the digital tape to a computer (HP a standard data acquisition system 9000) usin (3852% HP‘i . All the patients were proven by cardiac catheterization as normal or abnormal. For each patient, 10 cardiac cycles were digitized (sampling frequency = 4 kHz). Each cycle consisted of 1024 samples. As detailed elsewherekg, the diastolic heart sounds were passed through an anti-aliasing analog filter set for a cut-off frequency of 1200Hz and a high-pass digital filter set for a cut-off frequency of 180Hz. For each patient, since there is some slight change in parameter values from period to period, the average linear prediction coefficients were calculated by averaging over 10 cardiac cycles. Before the analysis, the direct current (DC) component from each record is eliminated (period by period). The ALE method was used to eliminate the background noise from the diastolic heart sound periods.

RESULTS The network was trained using a randomly chosen subset of the feature vector obtained from 12 atients; of these, six were normal and six abnorma B. It was tested on a database consisting of 100 (36 normal and 64 abnormal) recordings to which it had not been exposed. This neural network was trained until the averaged MS error per pattern at the output was less than 0.0 1. Figure 3 shows the average RMS error per

J. Biomed. Eng. 1993, Vol. 15, November

471

Acoustical dehtion

of coronq

occlusions using neural networks: M. Akay and W. Welkowitz

400

600

600

1000

Epoch Count Figure 3

Average

RMS error versus epoch count for training set

0

200

101

400

600

600

1000

Epoch Count Figure set

4

Incorrectly

diagnosed

patients verms epoch count for test

pattern for training set used in this study. Figure 4 shows the diagnostic performance of the neural network versus the number of epoch. As shown in of the test set qualitative errors Figure 4, the trajecto decreases very rapid y first, then increases after 400 epochs. At 400 epochs, 18 patients were incorrectly diagnosed. Several network configurations were tested. It was found that a network with 10 input nodes, 10 hidden layer nodes, and 2 output nodes showed the best performance. Using more than 10 hidden layer nodes did not improve the network performance. All training was done in a supervised fashion, which means that the inputs and desired outputs were known during the training process. The network correctly identified 50 of the 64 patients with coronary artery disease and 32 of the 36 patients with no apparent occlusions on the test set. The network performed with a detection rate (sensitivity) of 78.13% and a false alarm rate (1 .O-specificity) of 11.11%.

7

DISCUSSION

AND CONCLUSION

In this study, a neural network was applied to diastolic heart sounds recorded in a relatively noisy

472

J. Biomed. Eng. 1993, Vol. 15, November

room to detect coronary artery disease, noninvasively. The network had a detection rate 78.13% and a false alarm rate 11.11 O/o, whereas the previous study using the power spectral characteristics of the AR method showed a detection rate 75% and a false alarm rate of 13.89%‘. Although these results showed that neural networks are potentially capable of separating normal from abnormal patients, further studies should be done on a very large database to validate the effectiveness of this approach for mass screening procedure for detection of coronary artery disease. Considering the incorrectly diagnosed diseased patients by neural network, two patients had lOOO/o LAD occlusions. Since these blockages would permit no blood flow, the network detected these diseased There is no clear trend among patients as ‘normaPg. the other mis-diagnosed patients. The diagnostic ability of the neural networks can be improved further by using some physical examination variables such as sex, age, body weight, smoking condition, diastolic, and systolic pressure which are all routinely available to physicians. In addition to the prediction coefficients of the AR method, the physical examination parameters can be included in the feature vector. This feature vector, consisting of the prediction coefficients of the AR applied to the diastolic heart sound associated with coronary occlusions, can be used as the input pattern to the neural network. None of the patients studied had aortic regurgitation, mitral stenosis or other audible diastolic murmurs. Further work will involve the development of additional signal processing techniques to permit evaluating such patients. These results compare quite favourably with other noninvrisive methods for detecting CAD. For example, the sensitivity of the cardiointegram (CIG) technique developed by Teichholz et al., was found to be 73% with a specificity of 780/0~~.This approach was considered to be a moderately useful noninvasive method to detect CAD. However, valid comparisons between our approach and other diagnostic methods should be done on the same database. The im ortance of an effective procedure for nonivasive Py detecting coronary occlusions is selfevident. The advantages of this approach over any other invasive approaches are: noninvasive, assive (no radiation), fast, and inexpensive. In a 1 dition, this approach as described above is based on the measurement associated with turbulence, which is related to stenosis, and not on symptoms as with other nonivasive approaches. This is a very important fact in early detection and continuous monitoring of the coronary occlusions. Although these results are very encouraging for mass screening procedure to detect coronary artery disease, this acoustic approach failed to localize the coronary stenosis. Further work will involve extracting the feature vector(s) from the multi-channel recordings to localize the coronary occlusions. In order to further explore the extraction of the useful information regarding the complex diastolic heart sounds produced by single and multi-lesions, the analysis of the diastolic heart sounds will be approached using neural networks. For this purpose,

Acousticaldetection of cororq

another feature vector based on the minimum-norm (Eigenvector) parameters and the physical examination parameters will also be introduced.

ACKNOWLEDGEMENT The authors wish to thank Drs J.L. Semmlow, J. Kostis for providing data and technical assistance, and Y.M. Akay, J Redling, D. Shen, A. Smith, and V. Padmanabhan for collecting and preprocessing the data used in this study. This work was supported by a grant from Colin Medical Instrument Corporation.

REFERENCES 1. Guyton AC. Textbook of Medical PhysioloD. Pittsburg: WB Saunders Company, 1987. 2. Ganong WF. Review of Medical Physiology. Los Altos, CA: Lange Medical Publications, 1981, 485-8. 3. Krupp MA. Current Medical Diagnosti and Treatment. Los Altos, CA: Lange Medical Publications, 1982, 193210. 4. Semmlow JL, Welkowitz W, Kostis J, Mackenzie JW. Coronary Artery Disease - correlates between diastolic auditory characteristic and coronary artery stenoses. I%E Trans Biomed Eng 1983; BME-30: 136-9. 5. Akay M, Semmlow JL, Welkowitz W, Bauer M, Kostis J. Detection of coronary occlusions using AR modelling of diastolic heart sounds. IEEE Trans Biomed Eng 1990; BME-37: 366-73. 6. Akay M, Semmlow JL, Welkowitz W, Bauer M, Kostis J. Noninvasive detection of coronary occlusions using Eigenvector methods before and after angioplasty. IEEE Trans Biomed Eng 1990; BME-37: 1095-l 104. 7. Akay M. Noninvasive detection of coronary artery disease using advanced signal processing methods (PhD Thesis]. Piscataway, NJ: Rutgers University, 1990. 8. Welkowitz W, Akay M, Wang J-Z, Semmlow J, Kostis J. A model for disturbed coronary artery flow with phonocardiographic verification. In: Sideman S, Beyer K, eds, Activation, Circulation and Transport In the Cardiac Mu&. Massachusetts: Kluwer Academic Publishers, 1990. 9. Akay YM, Akay M, Welkowitz W, Semmlow J, Kostis J. A comparative study of advanced signal processing techniques for detection of coronary artery disease. Proc Im. Frontiers in Medicine 199 1; 39-2 140.

ot&uians ucing neural networks: h4. Akay and W. Welkowitz

IO. Cheng TO. Diagnostic murmur caused by coronary artery stenosis. Ann Intern Med 1970; 72: 543-6. 11. Dock W, Zoneraich S. A diastolic murmur arising in a stenosed coronary artery. Am J Med 1967; 42: 617. 12. Duncan GW, Gruber TO, Dewey CF, Myers GS, Lees RS. Evaluation of carotid stenosis by phonoangiography. New England JMed 1975; 27: 1124-8. 13. Lees RS, Myers GS. Noninvasive diagnosis of arterial disease, Adv Intern Med 1982; 23: 475-509. 14. Kartchner MM, McRae LP, Morrison FD. Noninvasive detection and evaluation of carotid occlusive disease. Arch Surg 1973; 106: 528-35. 1.5. Lipmann R. Introduction to computing with neural nets. IEEE ASSPMaga&e 1988; 4 (2): 4-22. 16. Huang W, Lippmann R. Comparison between neural network and conventional classifiers. A-oceedings 1987 IEEE Int Conf Neural Networks, 1987, 485-94. 17 McClelland JL, James L, Rumelhart D. Explorations in Parallel Distn.buted Processifig: A Handbook of Models, Programs and Exercises. Cambridge, MA: MIT Press, 1988. 18. Rumelhart D, Hunton, GE, Williams RJ. Learning Internal Representations by Error Propagation, Parallel Distributed Processing; Explorations in the Microstructure of Cognition, Vol. 1. Cambridge, MA: MIT Press, 1986. 19. Chen J, Vandewalle J, Sansen W, Vantrappen G, Janssens J. Adaptive method for cancellation of respiratory artifact in electrogastric measurements. Med Biol Eng Camp 1989; 27: 57-63. 20. Ferrara ER, Widrow B. Adaptive filtering for signal enhancement. IEEE Tram Assp 1981; ASSP-29, No 3. 21. Haykin S. Adaptive F&r Z%eory.Englewood Cliffs, NJ: Prentice Hall, 1986. 22. Makhoul J. Spectral linear properties and applications. Tram ASSP 198 1; ASSP-29: 282-96. 23. Kay SM, Marple SL. Spectral analysis: a modem perspective. Proc IEEE 1981; 69: 1380-1419. 24. Kelly MF, Parker PA, Scott RN. The application of neural networks to myoelectric analysis: a preliminary study. IEEE Trans Biomed Eng 1990; 37: 22 l-30. 25. Marmarou RM, White MW, Marrnarou A. Classification of somatosensory of evoked potentials recorded from patients with severe head injuries. IEEE BME Magazine, 1990; 43-9 26. Padmanabhan V, Fisher R, SemmlowJL, Welkowitz W, Kostis J. High sensitivity PCG transducer for extended frequency applications, Proc IEEE Frontiers in Medicine 1989; 57-9. 27. Teichholtz LE. Cardiointegram detection CAD using in normal resting electrocardiograms. J Am Co11 Cardio; 3: 598.

J. Biomed. Eng. 1993, Vol. 15, November

473