«Detection of coronary artery disease with an electronic stethoscope Schmidt, Samuel Publication date: Document Version Publisher's PDF, also known as ...»
The weight of the accelerometer will dampen the signal, especially at higher frequencies. Therefore Padmanabhan et al. developed a lightweight accelerometer [57, 59]. Since the principle of the accelerometer was similar to the FYSPac2 accelerometer develop by Vermarien et al. the name FYSPac2 will be used as a reference to the accelerometer developed by Padmanabhan et al. The concept was a balancing beam mounted on a small stand which was attached to the chest wall, see Figure 12. Chest wall vibrations make the balancing beam bend and the deflection is collected with piezoelectric plates mounted on the beam. By adjustment of the two masses in the end of the bean the resonance frequency of the sensor is adjustable. The result was a sensitive accelerometer with a relative flat frequency response from 200-800 Hz. The FYSPac2 accelerometers sensitivity to sound pressure 2 was approximately 100 mV/Pa in the 200-800 Hz frequency band . The resonance frequency was approximately 1050 Hz. A disadvantage of the sensor is that it fragile and thereby not suitable for clinical use .
The FYSPac2 accelerometer was used in several studied by the research group at Rutgers University who published the majority of studies related to acoustic detection of CAD [65, 66, 70, 71, 73, 75, 82, 84, 85, 91].
Figure 12. Illustration of the lightweight FYSPac2 accelerometer .
An electronic stethoscope based on Electromagnetic Diaphragm principle was marketed in 2003 by Thinklabs. The stethoscope’s diaphragm is coated with a conductive surface. Behind the diaphragm a metal plate is located. The diaphragm and the plate work as a capacitor with a variable capacitance depending on the distance between them. The Thinklabs stethoscope was used in some recent studies about acoustic detection of CAD [63, 78, 79]. Other transducer principles include piezoelectric contact sensors  which was used by Chen et al. in studies of CAD murmurs .
1.4.2. Spectral analysis Semmlow et al. publish the first signal processing study of CAD murmurs in 1983 .
The signal was recorded using an air-coupled microphone. By power spectral density (PSD) analysis of the diastolic segments they showed a relative increase in the spectral energy above 90 Hz in 12 CAD patient compared to 12 normal subjects.
The spectral approach was continued by Akay et al. in several studies. The power spectra were estimated using parametric models such as the autoregressive (AR) models and Eigenvector based spectral models in subjects before and after angioplasty [69, 88, 90]. Both methods showed a decrease in power above 200 Hz after removal of the stenosis with angioplasty, see Figure 13. Parametric models were chosen because of 2Usually accelerometer sensitivity is not measured in pressure units. The authors of the study used a special setup to correlate the accelerometer output to sound pressure which is more comparable to microphones.
their noise robustness and their fitness for detection of spectral peaks such as the coronary artery resonance frequency.
Figure 13. Power spectrum of a parametric model of diastolic sound before and after angioplasty.
 Later the parametric models were used to discriminate CAD subject from non-CAD subjects [68, 72, 85, 89]. A CAD related increase in energy above approximately 200Hz was observed in all studies. This was observed with both the FYSPac2 sensor and the air coupled microphone. Since spectral peaks were expected in the diastolic sounds from CAD subjects the diastolic recording segments were filtered with an adaptive line enhancer before modeling . The adaptive line enhancer was an adaptive filter adjusted for linear prediction, thereby enhancing the spectral peaks at the expense of the wide band part of the signal. The filter thus emphasized more distinct spectral peaks .
In a comparative study the classification performance of the Fast Fourier transform (FFT) and three different parametric spectral methods such as AR, Autoregressive moving average model (ARMA) and Eigenvectors were evaluated in recordings from 80 subjects obtained with the FYSPac2 accelerometer . The different methods showed diverging results. The FFT analysis showed an increase in energy above 500 Hz in the CAD subjects compared to non-CAD subject, but the Eigenvector method showed that CAD was related to an increase in energy between 300-500 Hz. The AR and ARMA models showed an increase in the 400-800 Hz band. For all three parametric models the magnitude of the second pole was used as a discriminator in a classification test. The Eigenvector method was the best performing classifier. In 80 cases the sensitivity was 79.2% and specificity was 90.6% . The usefulness of the Eigenvector method was further confirmed in a study including 100 patients were sensitivity was 88.8% and specificity was 78.2% Tateishi et al. analyzed heart sounds above 400 Hz and found that the power ratio calculated as the power in the 400-700 Hz band divided by the power in the 400-1500 Hz band increased in CAD subject . Recordings were made from five positions on the chest from 168 subjects. Only recordings obtained from the forth intercostal space showed a significant difference between CAD and non-CAD subject. The sensitivity was 71% and specificity was 65%.
Recently the Think Lab stethoscope was used in a study by Gauthier et al. . Using the power spectral density they found that the power ratio between the frequency band above and below 130 Hz increased in CAD subjects compared to healthy subjects.
The different studies all showed that the diastolic power at higher frequencies was increased in CAD subjects. However the definition of higher frequencies differs widely from study to study. For example in the first study by Semmlow et al. CAD was related to an increase in power above 90 Hz, where the parametric modeling studies usually related CAD to an increase in power above 200-300 Hz. The variations are probably related to differences in the transducers, differences in the spectral analysis methods and small population sizes in some studies. An example of the influence of the transducer is that the FYSPac2 accelerometer used in several parametric modeling studies is not linear outside the 200-800 Hz frequency band. The high success rate of the parametric modeling might indicate that the spectral peaks are present in the CAD murmurs.
1.4.3. Time-frequency studies Murmurs are non-stationary signals. Therefore, Akay et al. applied adaptive filters for tracking spectral changes over the diastolic periods [65, 66]. They monitored the magnitude of the second pole in the adaptive filter throughout the diastolic period. In one study 10 patients were monitored before and after angioplasty . Using a blind protocol the authors was able to identify whether the recording was obtained before or after angioplasty in 9 out of the 10 cases. Figure 14 shows the magnitude trajectories of the second pole throughout the diastole before and after angioplasty. The pole magnitude was increased in the 200-300 ms interval before angioplasty, which was the case in 9 out of 10 cases. In a second study including 35 subjects (non-CAD and CAD subjects) the findings of increased magnitude of the second poles in the diastolic interval from 200-300 ms was confirmed . Furthermore, the variance of the pole magnitude was increased in CAD patients.
Figure 14. Magnitude trajectories of the second pole throughout the diastole.
The Solid line was before angioplasty and the dotted line after angioplasty .
Zhao et al. calculated the instantaneous frequency throughout the diastolic period using the Hilbert Huang Transform . In a case study of a CAD patient undergoing coronary angioplasty the mean weighted instantaneous frequency was 155 Hz before angioplasty and 98.3 Hz after. Similarly, the variance of the instantaneous frequency decreased from 42 Hz to 16.7 Hz after removal of the stenosis. The finding indicated that CAD increases the non-stationarity of the diastolic heart sound.
1.4.4. Nonlinear dynamics Since murmurs originate from turbulent flow it has been argued that the murmurs reflect the non-linear and chaotic characteristic of turbulence. Padmanabhan et al.
applied the Grassberger method for estimation of the correlation dimension of an underlying attractor . The dimension is related to the degrees of freedom of the underlying system. The hypothesis was that if the signal is governed by a finite dimension attractor the Grassberger correlation integral will saturate even when the embedding dimension of the phase space is increased. Or explained in another way if the attractor can be described by a certain number of variables the complexity of an ideal attractor model will not increase even if a higher number of variables are available for modeling of the attractor. Opposite, if the signal is a completely stochastic process the dimension is infinite and the correlation integral will not saturate.
Padmanabhan et al. found at that the Grassberger correlation integral saturated in 10 diseased subjects and that the Grassberger correlation integral didn’t saturate in 5 normal subjects. This indicates that the diastolic sound recordings from normal subjects were dominated by random noise and that diastolic sound from CAD subjects were influence by a dynamical system.
Akay et al. analyzed the complexity of diastolic periods using Approximate Entropy . They found that Approximate Entropy was increased in 30 CAD subjects compared to 10 normal subjects, indicating that CAD increases the complexity of the diastolic sounds.
1.4.5. Multivariate classifiers The parameter values from the parametric models were used as input to neural networks in a study including 100 subjects . The tests showed a sensitivity of 78% and specificity of 89%. A second set of recordings from 112 subjects were analyzed with wavelets and classified with a neural network . From the third wavelet band extremas of the wavelet coefficients were identified and the statistical moments Mean, Variance, Skewness and Kurtosis were calculated from the extremas. These features were combined with physiological variables such as sex, age, body weight, smoking condition, and systolic and diastolic pressure in a neural network. 82 of the recordings were reserved for the test. The result was a sensitivity of 78% and specificity of 89%.
The multivariate method did therefore not perform better compared to the Eigenvector based parametric model.
Zhao et al. did several studies using multivariate classification. Based on the Hilbert Huang Transform, features such as Mean, Variance, Skewness and Kurtosis of the average instantaneous frequency were extracted. The same feature set was used in two different classifiers: a support vector machine  and a neural network . Tested in 37 subjects the sensitivity of the support vector machine was 85% and the specificity was 100%. The neural network was tested in 40 subjects, giving a sensitivity of 95% and a specificity of 85%. In a study by Chen et al. wavelets were used for denoising of the recordings before features were extracted with parametric models . The study showed a sensitivity of 87.5% and 100% specificity in 28 subjects. The weakness of those studies was a low number of subjects and that no blind protocols were used.
1.4.6. Detection of CAD using an electronic stethoscope Two of the described studies used the electronic stethoscope for data collection [63,79].
These studies were published parallel to the current work and seem to confirm the suitability of the electronic stethoscope for detection of CAD.
1.4.7. Summary of prior art CAD is associated with an increased energy at higher frequencies, but the specific high frequency bands which were affected by CAD diverge from study to study. Several studies were successful in application of parametric modeling for spectrum analysis and feature extraction. Other studies showed that there are indications of non-linear dynamics in diastolic heart sound. Some studies confirm that CAD increases the nonstationarity of the diastolic heart sound. Multivariate classification methods have been applied for classification. The largest multivariate classification study was tested in 82 subjects. Even though physiological variables such as sex, age, body weight, smoking condition, and systolic and diastolic pressure were combined in a neural network with four heart sound based features, the classification performance did not exceed the performance of the pole magnitudes from parametric models.