«Detection of coronary artery disease with an electronic stethoscope Schmidt, Samuel Publication date: Document Version Publisher's PDF, also known as ...»
 Akay, Yasemin M., Akay, Metin, Welkowitz, Walter, Semmlow, John L., Kostis,John B., "Noninvasive detection of severity of occlusions associated with coronary artery disease," in IEEE Proceedings of the Annual Conference on Engineering in Medicine and Biology, 1991, pp. 2101-2102.
 M. Akay, "Harmonic decomposition of diastolic heart sounds associated with coronary artery disease," Signal Process, vol. 41, pp. 79-90, 1995.
 M. Akay, "Noninvasive diagnosis of coronary artery disease using a neural network algorithm," Biol.
Cybern., vol. 67, pp. 361-367, 1992.
 Akay, M., Bauer, M., Semmlow, J.L., Welkowitz, W., Kostis,J., "Analysis of diastolic heart sounds before and after angioplasty," in IEEE Proceedings of the Annual Conference on Engineering in Medicine and Biology, 1988, pp. 257-259.
 Akay, M., Bauer, M., Semmlow, J.L., Welkowitz, W., Kostis,J., "Autoregressive modeling of diastolic heart sounds," in IEEE Proceedings of the Annual Conference on Engineering in Medicine and Biology, 1988, pp. 172-174.
 Akay, M., Semmlow, J.L., Welkowitz, W., Kostis,J., "Parametric modeling of diastolic heart sounds before and after angioplasty," in IEEE Proceedings of the Annual Conference on Engineering in Medicine and Biology, 1989, pp. 51-52.
 J. Semmlow and K. Rahalkar, "Acoustic detection of coronary artery disease," Annual Review of Biomedical Engineering, vol. 9, pp. 449-469, 2007.
 Zhao, Z., Ma,C., "An intelligent system for -oninvasive diagnosis of coronary artery disease with EMD-TEO and BP neural network," in 2008 International Workshop on Education Technology and Training and 2008 International Workshop on Geoscience and Remote Sensing, ETT and GRS 2008, 2009, pp. 631-635.
 Watakabe, M., Mita, K., Akataki, K., Itoh,Y., "Mechanical behaviour of condenser microphone in mechanomyography," Medical and Biological Engineering and Computing, vol. 39, pp. 195-201, 2001.
 Vermarien, H., Van Vollenhoven,E., "The recording of heart vibrations: A problem of vibration measurement on soft tissue," Medical and Biological Engineering and Computing, vol. 22, pp. 168Pasterkamp, H., Kraman, S.S., DeFrain, P.D., Wodicka,G.R., "Measurement of respiratory acoustical signals: Comparison of sensors," Chest, vol. 104, pp. 1518-1525, 1993.
 Watakabe, M., Itoh, Y., Mita, K., Akataki,K., "Technical aspects of mechnomyography recording with piezoelectric contact sensor," Medical and Biological Engineering and Computing, vol. 36, pp. 557Chen, T.-H., Xing, S.-X., Guo, P.-Y., Tang, H.-T., Yu,Z., "The design of digital collecting system of heart sound signals based on DSP," in 2010,.
 S. Schmidt, C. Holst-Hansen, E. Toft and 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.
 S. E. Schmidt, C. Holst-Hansen, C. Graff, E. Toft and J. J. Struijk, "Segmentation of heart sound recordings by a duration-dependent hidden Markov model," Physiol. Meas., vol. 31, pp. 513-529, 2010.
 S. E. Schmidt, M. Graebe, E. Toft and J. J. Struijk, "No evidence of nonlinear or chaotic behavior of cardiovascular murmurs," Biomedical Signal Processing and Control, 2010.
 S. E. Schmidt, E. Toft, C. Holst-Hansen and J. J. Struijk, "Noise and the detection of coronary artery disease with an electronic stethoscope." in The 5th Cairo International Conference on Biomedical Engineering 2010, Cairo, 2010,.
Publish in 5th Cairo International Biomedical Engineering Conference proceedings Study 5 Acoustic features for the identification of coronary artery disease Submitted to IEEE Transactions on Biomedical Engineering 2011
7. Discussion Five studies were conducted to develop a method for detection of CAD with an electronic stethoscope. In study one a method was developed for segmentation of the heart sounds into systolic and diastolic periods. In study two a framework was developed for robust extraction of features. Study three examined the fundamental characteristics of murmurs. The influence of noise was estimated in study 4 before the classification performance of different types of features was tested in study five. The current chapter discusses the findings in these five studies.
Since the CAD related murmurs are expected to peak in the diastolic periods, identification of the diastolic periods is essential. A duration dependent Markov model (DHMM) was applied for segmentation of heart sounds without the need for an additional reference signal. The concept of the DHMM fits the problem of heart sound segmentation well. The states of the repeating heart cycle can be modeled as a Markov process. As in a hidden Markov model the actual state of the heart at a given time is unknown, but the heart sounds are observable and related to the state of the heart. Since the duration of the states in the heart cycle is relatively stable over a recording period of a few heart beats, the probability of transition from the current state to the next state is related to the time spent in the current state. This aspect was modeled by the duration dependent Markov model.
The DHMM’s capability to identify S1 and S2 sounds was tested in recordings from 73 patients. The sensitivity was 98.8% and positive predictivity was 98.6%, which shows that the method is robust and accurate and that the DHMM model is suited for modeling of the heart sounds. Heart valve murmurs such as murmurs from aorta stenosis did not reduce the performance significantly. The strength of the model, the duration dependency of state transitions, is also the limitation of the model since it limited the performance in highly arrhythmic patients. A potential solution to this problem is to further customize the probability distributions of the systolic and diastolic durations to the individual patient. In the current implementation probability distributions of the systolic and diastolic durations were determined by a normal distribution were only the means were determined individually from each subject. The mean durations were estimated from the autocorrelation of the signal envelope in the current implementation, but the degree of arrhythmia might also be estimated from the autocorrelation, and therefore, the duration distributions might be fitted to the degree of the arrhythmia. However no models are perfect and the perfect segmentation of heart sounds is not achievable, therefor an automatic post validation method would be needed to reduce the risk of erroneous classifications caused by incorrect segmentation.
The high degree of robustness was further confirmed in study five were the diastole locations in 435 recordings were corrected if they were incorrectly placed. This happened in 3% of the diastoles.
7.2. Noise and noise reduction
In study 2 the focus was to develop a framework for extraction of diastolic heart sound features. The goal was to reduce the effect of friction spikes and other types of noise. A simple framework was developed by subdivision of the diastolic periods into subsegments of short duration. The noise level of the sub-segmentation was estimated by the variance and the level of stationarity. Sub-segments with high variance and a high level of non-stationarity were removed before the magnitude of the 1 st pole in an ARmodel was extracted as a feature from the sub-segments. The final feature value was then calculated as the median of the feature values from the sub-segments. According to the study at a dataset consisting of 50 recordings the sub-segmentation improved the separation capability of the AR-pole considerably. The framework was further used for feature extraction in study 4 and 5.
Study 4 was conducted to evaluate the influence of noise on features for detection of CAD. Four types of noise were analyzed: ambient noise, recording noise, respiration noise and abdominal noise. The influence of friction spikes (friction noise of short duration) was not included in the study since these were present in nearly all recordings, but recording noise included friction noise of longer duration. 633 recordings from 140 patients were analyzed by listening and visual inspection. The degree of contamination from the four noise types was quantified according to noise intensity and duration. The magnitude of AR-poles was used as features and calculated from a low frequency band (25-250) and a high frequency band (250-1000 Hz). The classification potential of the AR-poles was evaluated by the area under the receiver operating characteristic (AUC). In 75.7% of the recordings noise contamination was identified. The AUC was first calculated in the clean recordings (recordings without observed noise) before noisy recordings were added gradually as more and more noise was tolerated. The trend in the AUC from the high frequency band was that the AUC dropped as noisier recordings were included in the analysis. Even weak noise seams to influence the performance of the feature from high frequency band. In contrast, the AUC from the low frequency band was influenced only by very extensive noise. The study clearly indicates that noise is a significant problem for analyses of the high frequency part the signal. A limitation of the study was that the estimates of the influence at specific noise levels and durations were imprecise. This was due to the fact that every time the noise tolerance was increased with one step only a few new recordings were added. For example if the tolerance for ambient noise was increased from moderate noise of maximum one second to moderate noise of maximum two seconds only few new recordings were added to the analysis.
The result of study 4 clearly indicates that, even when study 2 showed that the subsegmentation method improved the classification performance, the noise issue isn’t solved. Ambient noise and recording noise were the most common noise sources in study 4 and they both had significant effect at classification performance of the ARpoles from the high frequency bands. The effect of ambient noise might be reduced by either active or passive noise reduction. Passive noise reduction may include better shielding of the microphone or uses of other transducer types such as accelerometers, which are more robust to ambient noise. In an active noise reduction setup a reference signal from an external microphone might be used for adaptive filtering. The problem of recording noise was typical due to scratching between the stethoscope diaphragm and the skin of the chest. In some recent versions of electronic stethoscopes, such as 3M Littmann Model 3100, the material of diaphragm was chosen to reduce the friction noise , but the safe solution might be to attach the transducer to the chest wall. The third most common noise source was respiration noise. Respiration noise can be limited by asking the patients to hold their breath. Therefore, the influence of three most common noise sources ambient noise, recording noise and respiration might be reduced by changes in the recording equipment and the examination protocol.
7.3. The potential of nonlinear signal processing techniques
Since the murmurs are described as broad-banded in the frequency domain, the most obvious signal model is a linear stochastic process, but as proposed by Padmanabhan et al. the murmurs might be dominated by non-linear dynamics such as chaos . Clearly the murmurs cannot be described by a simple deterministic model, but maybe the murmurs might be described by more complicated nonlinear dynamics such as low dimensional chaos. Therefore, the null hypothesis that cardiovascular murmurs were from a linear stochastic process was tested using recordings of carotid artery murmurs.