«Detection of coronary artery disease with an electronic stethoscope Schmidt, Samuel Publication date: Document Version Publisher's PDF, also known as ...»
Even though several methods were applied for the analysis, no significant difference was observed between the murmurs and surrogates of the murmurs generated by a linear stochastic process. We concluded that there were no signs of nonlinear characteristics or low dimensional chaos in murmurs from the carotid artery. The carotid artery was chosen since the stenosis is located close to the skin which ensured a good signal to noise ratio and a simpler transfer function between the origin of the murmur and the recording spot. There might be other circumstances related to CAD murmurs which might allow more deterministic murmurs, for example if turbulence isn’t fully developed the velocity fluctuations of the vortexes in post stenotic flow will be more deterministic. However study 3 underlines that careful analysis must be conducted before nonlinear dynamics is assumed. The finding in study 3 was confirmed in study 5 where the performance of spectral entropy, which doesn’t handle non-linearity, exceeded sample entropy which handles nonlinear dynamics. In a supplementary study of diastolic sounds a high correlation was found between the sample entropy and features from an AR-model (see appendix).
7.4. Features for detection of CAD
Different types of feature were evaluated in a cross validation study. Since the existing literature differs in the choice of frequency bands several filter configurations were tested. Only features from lower frequency bands showed a significant difference between non-CAD and CAD subjects. A wide range of features from low frequency bands were significant, but principle component analysis of all features showed only one significant PCA component which indicates that the different features describe the same phenomena. The best feature was the pole magnitude of the 1 st pole in a 6th order AR-model of the 25-250 Hz frequency band. In CAD subjects the pole magnitude was increased, which demonstrated a relative power increased at frequencies at 20-30 Hz.
Similarly, an increased power was observed at lower frequencies (25-80Hz) in CAD patients by 1/3 octave band analysis. That features from the higher frequency bands showed a poor classification performance contradict with findings in prior studies which showed a good classification performance of features from higher frequency bands such as the 180-1200 Hz band [3, 4]. Also the preliminary study conducted in the early phase of the current thesis showed a good classification performance of features from high frequency bands in low noise recordings. However the findings are not surprising since study 4, which used the same dataset as study five, found that noise influenced the performance of high frequency features dramatically. An alternative solution would have been to only analyse recordings with a low level of noise contamination, but this would have excluded approximately 60% of the recordings.
Since the features from the AR model were the best performing features, AR features were chosen for multivariate classification. By combination of the AR features from the 25-250 Hz band and the 250-1000 Hz band a CAD score was constructed. The AUC of the CAD-score was 0.73 (0.685-0.776), sensitivity was 72% and specificity was 65.2%. This was only slightly better than the performance obtained with only one AR-pole from the 25-250 Hz frequency band.
There was a clear gender difference in the CAD score. In the non-CAD subjects the females scored significantly lower compared to males. Further analysis showed a significant gender difference in low frequency features, but no significant gender difference was observed in the AR-feature from the 250-1000 Hz band. Since more males than females suffered from CAD the gender difference might explain a part of the increased CAD-score in CAD subjects, but when the CAD score was tested separately in males and females the CAD score was increased in both males and females. When the two genders was separated the AUC was 0.722 (95% CI: 0.664for males and 0.638 (95% CI: 0.518-0.759) for females, which indicates a performance drop in the females. This is contradictory to unpublished findings referred to by Semmlow et al.  where a CAD detection algorithm based on high frequency features showed a poor performance in males and high performance in females. The genders differences show that further studies must take the gender difference into account.
7.5. The cause of increased low frequency power
Study 4 and 5 identify a new feature for detection of CAD from heart sounds. The studies showed a power increases at lower frequencies in subjects with CAD. Figure 1 below illustrates recordings from a non-CAD patient and a CAD patient. The recordings were filtered with a 20-50 Hz band pass filter. The recordings are from two males with approximately the same BMI. The amplitudes in both the systoles and diastoles are increased in the CAD subject. This indicates that the change also might occur in the systolic periods.
Figure 1. Two band pass filtered (20-50Hz) heart sound recordings from a CAD and a non-CAD patient.
The source of the increase in low frequency power is not known, but an increase at these frequencies is usually not related to cardiovascular murmurs. A more likely source is changes in ventricular movements caused by changes in the compliance of the left ventricle. It is known that CAD can increase ventricular stiffness, even before myocardial infarction (MI) . The diastolic effect of increased ventricular stiffness is a change in filling patterns. Typically, the relation between inflow in early diastole and late diastole is altered . Changes in ventricular compliance are often reflected as S3 and S4 sounds in heart sound recordings. To test if S3 and S4 sounds caused the increase in low frequency power, a small retrospective study was conducted. The S3 and S4 sounds were identified by visual screening the recordings included in study five. S3 sounds were found in 8.6% of the non-CAD patients and in 15.9 % of the CAD patients. Similarly, S4 sounds were found in 8.6% of the non-CAD patients and in 14.3 % of the CAD patients. The mean CAD score 1 was -0.21 (STD 0.44) in non-CAD subjects without S3 and S4 sounds and -0.12 (STD 0.48) in non-CAD subjects with S3 or S4 sounds. In CAD subjects the mean CAD score was 0.23 (STD 0.34) and 0.22 (STD 0.37) in respectively subjects with and without S3 and S4 sounds. Therefore, the S3 and S4 sound cannot explain the increase in low frequency power observed in CAD subjects.
Low frequency chest wall vibrations related to ventricular movements can be measured with the use of seismocardiography . The typical frequency range of seismocardiography is 0.3-50 Hz [8-10], which is overlapping with the frequency range analyses in the current study. Seismocardiography is characterized by rather deterministic wave patterns which reflect cardiac events. Studies have shown that MI alters both the systolic and diastolic Seismocardiographic patterns [8-10].
Figure 2 shows heart sound recordings from a new sensor developed for future research by the current author and colleagues. The frequency response of the new sensor was flat from 0.5 Hz to 1000 Hz. The recording contains a low frequency signal with a In study 5 a CAD score higher than zero was associated with CAD.
distinct morphology which includes landmarks known from seismocardiography. In the CAD subject the amplitude of the low frequency vibration was increased compared to the non CAD subject. When the 20-40 Hz band pass filter was applied the distinct morphology diminishes, but the amplitude difference remains. This indicates that the increase in low frequency power observed in the current study might be further understood and quantified if the frequency range of the sensor is expanded to include lower frequencies. If the power of the low frequency part of the signal is related to ventricular movements and compliance it is likely to be uncorrelated to features from higher frequency bands which is related to CAD murmurs. Therefore, if the noise problems are solved, a combination of the low frequency features and high frequency features might improve the classification performance significantly.
Figure 2.top figures show Heart sounds recorded with a sensor with an extended frequency range.
The bottom figures show the same recordings but after band pass filtering.
To further understand the change in low frequency power, the relation to physiological variables was evaluated briefly. Unfortunately, physiological data such as blood pressure and BMI was accessible in only 66 subjects out of the 140 subjects included in study 4 and study 5. The correlation (r) between the diastolic power in the 20-40 Hz band and Age, BMI, systolic blood pressure and diastolic blood pressure was respectively 0.17, -0.27,-0.09 and 0.17. The most evident, but still weak, correlation was the negative correlation with BMI. This was expected since an increased BMI is likely to increase the distance from the heart to the transducer. In females the breast the increases the distance between the heart and the stethoscope which might explain the weaker low frequency power in females.
Figure 3. Scatter plot between BMI and the low frequency power of the diastolic heart sound.
7.6. Clinical implication of current findings For sensitivities of the current method exceed the ECG stress test, but the specificity of the stethoscope based method was low. Thereby, the clinical benefits of the current method without improvements might be limited. However, despite a relative low diagnostic performance the method might provide valuable information in a broader risk estimation strategy. The Framingham risk score defines the 10 year risk for CAD in three levels: low (10%), intermediate (10-20 %) and a high (20%) risk. If for an illustrative purpose the 10 year risk is converted to prevalence the positive predictive value in patients with intermediate risk will be 18.6%-34% and the negative predictive value 90.3-95.4%. Therefore, nearly all patients can be reclassified to either low risk 10% or high risk 20%, which certainly has clinical value.
7.7. Recommendations for new hardware The purpose of the current thesis was to develop a method for detection of CAD with an electronic stethoscope, but the studies revealed several weaknesses of the electronic stethoscope. Consequently, a recommendation for a new system for detection of CAD
8. Conclusion The purpose of the current work was to develop an algorithm for detection of CAD with an electronic stethoscope. Therefore, a method was developed for automatic segmentation of the heart sounds into systolic and diastolic periods. Next a simple framework was developed for robust extraction of descriptive features. To gain further insight in the characteristics of cardiovascular murmurs, a study was conducted to evaluate whether the murmurs can be described by nonlinear dynamics. The study did not find evidence of nonlinear dynamics in cardiovascular murmurs. Instead, the murmurs might be characterized as a non-stationary linear stochastic process.
To identify features for classification between CAD and non-CAD patients a wide range of features was examined in 430 recordings from 140 patients. New efficient features were identified from lower frequency bands. The new features were related to the changes in the power distribution of the low frequency part of the signal. An advantage of these features was a high degree of robustness against noise. This was in contrast to features from higher frequency bands which were very sensitive to noise.
The mechanism behind the observed change in the low frequency part of the signal is not known, but the change might be related to changes in ventricular compliance.
A cross validation test of a multivariate classification algorithm resulted in an area under the receiver operating characteristic of 0.73, the sensitivity was 72% and the specificity was 65.2%. The relative low performance reflects several weak points of the current electronic stethoscope when used in clinical settings. The limitations include friction noise, sensitivity to ambient noise and short recording time. If these limitations are handled the performance of features from the high frequency bands might improve significantly. Thereby, a system based on a combination of features related to low frequency vibrations and features suited for quantification of the high frequency CAD murmurs is likely to be a successful non-invasive test for CAD.
3M Littmann, "3M™ Littmann® Electronic Stethoscope Model 3100 user interface," vol. 2011,.
  V. Padmanabhan and J. L. Semmlow, "Dynamical Analysis of Diastolic Heart Sounds Associated with Coronary-Artery Disease," Ann. Biomed. Eng., vol. 22, pp. 264-271, 05, 1994.
 Akay, Yasemin M., Akay, Metin, Welkowitz, Walter, Semmlow, John L., Kostis,John B., "A comparative study of advanced signal processing techniques for detection of coronary artery disease," in IEEE Proceedings of the Annual Conference on Engineering in Medicine and Biology, 1991, pp.
 M. Akay, "Harmonic decomposition of diastolic heart sounds associated with coronary artery disease," Signal Process, vol. 41, pp. 79-90, 1995.
 J. Semmlow and K. Rahalkar, "Acoustic detection of coronary artery disease," Annual Review of Biomedical Engineering, vol. 9, pp. 449-469, 2007.
 W. J. Paulus, D. L. Brutsaert, T. C. Gillebert, F. E. Rademakers, S. U. Sys, A. F. Leite-Moreira, O. M.
Hess, Z. Jiang, P. Kaufmann, L. Mandinov, C. Matter, P. Marino, D. G. Gibson, M. Y. Henein, J.
Manolas, O. A. Smiseth, M. Stugaard, L. K. Hatle, P. Spirito, S. Betocchi, B. Villari, O. Goetzsche and A. M. Shah, "How to diagnose diastolic heart failure," Eur. Heart J., vol. 19, pp. 990-1003, 1998.
 V. Fuster, R. A. O. Rourke, R. A. Walsh and p. Poole-Wilson, Hurst's the Heart. McGraw-Hill, 2008.
 D. M. Salerno, J. M. Zanetti, L. A. Green, M. R. Mooney, J. D. Madison and R. A. Van Tassel, "Seismocardiographic changes associated with obstruction of coronary blood flow during balloon angioplasty," Am. J. Cardiol., vol. 68, pp. 201-207, 1991.
 D. M. Salerno, J. M. Zanetti, L. C. Poliac, R. S. Crow, P. J. Hannan, K. Wang, I. F. Goldenberg and R.
A. Van Tassel, "Exercise seismocardiography for detection of coronary artery disease," American Journal of Noninvasive Cardiology, vol. 6, pp. 321-330, 1992.
 D. M. Salerno and J. Zanetti, "Seismocardiography for monitoring changes in left ventricular function during ischemia," Chest, vol. 100, pp. 991-993, 1991.
Appendix The appendix contains two publications by the current author, which are cited in the Thesis.
An abstract S. Schmidt, C. Holst-Hansen, E. Toft and J.J. Struijk, "Detection of coronary artery disease with an electronic stethoscope: Is it possible?" in Summer Meeting, Danish Cardiovascular Research Academy, Sønderborg, Denmark, 2007, pp. No. 26.
A conference paper:
Samuel Schmidt, John Hansen, C. H. Hansen, Egon Toft and Johannes J. Struijk, "Comparison of sample entropy and AR-models for heart sound-based detection of coronary artery disease," Comput. Cardiol., vol. 37, pp. 385-388, 2010.