Limits...
A framework for automatic heart sound analysis without segmentation.

Yuenyong S, Nishihara A, Kongprawechnon W, Tungpimolrut K - Biomed Eng Online (2011)

Bottom Line: Geometric mean was used as performance index.Average classification performance using ten-fold cross-validation was 0.92 for noise free case, 0.90 under white noise with 10 dB signal-to-noise ratio (SNR), and 0.90 under impulse noise up to 0.3 s duration.Further work include building a new training set recorded from actual patients, then further evaluate the method based on this new training set.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of Communication and Integrated Systems, Tokyo Institute of Technology, Japan 2-12-1-W9-108 Ookayama, Meguro-ku, Tokyo, Japan. toey123@gmail.com

ABSTRACT

Background: A new framework for heart sound analysis is proposed. One of the most difficult processes in heart sound analysis is segmentation, due to interference form murmurs.

Method: Equal number of cardiac cycles were extracted from heart sounds with different heart rates using information from envelopes of autocorrelation functions without the need to label individual fundamental heart sounds (FHS). The complete method consists of envelope detection, calculation of cardiac cycle lengths using auto-correlation of envelope signals, features extraction using discrete wavelet transform, principal component analysis, and classification using neural network bagging predictors.

Result: The proposed method was tested on a set of heart sounds obtained from several on-line databases and recorded with an electronic stethoscope. Geometric mean was used as performance index. Average classification performance using ten-fold cross-validation was 0.92 for noise free case, 0.90 under white noise with 10 dB signal-to-noise ratio (SNR), and 0.90 under impulse noise up to 0.3 s duration.

Conclusion: The proposed method showed promising results and high noise robustness to a wide range of heart sounds. However, more tests are needed to address any bias that may have been introduced by different sources of heart sounds in the current training set, and to concretely validate the method. Further work include building a new training set recorded from actual patients, then further evaluate the method based on this new training set.

Show MeSH
Feature extraction from heart sound segment.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3045988&req=5

Figure 10: Feature extraction from heart sound segment.

Mentions: The feature set used in this work consists of two parts: three features come from the envelope segments and another 32 come from heart sound segments. Features were obtained from heart sound segments using the same method as in [12] where d2 DWT coefficients decomposed for six levels using Daubechies-2 wavelet were partitioned into 32 non-overlapping windows and the signal energy of each window was used as a feature. This procedure is illustrated in Figure 10. It yields a 32-element feature vector for each heart sound segment. Another three features were obtained from envelope segments. Figure 11 shows an envelope segment. Each peak is marked with a circle and a peak interval and is shown by the arrows. The features derived from it were: the number of peaks, the average distance (in samples) between consecutive peaks, and the signal energy of the whole segment. The peak detection algorithm used to obtain the first two of these features is an improvement over simple thresholding; its pseudo code is shown below.


A framework for automatic heart sound analysis without segmentation.

Yuenyong S, Nishihara A, Kongprawechnon W, Tungpimolrut K - Biomed Eng Online (2011)

Feature extraction from heart sound segment.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3045988&req=5

Figure 10: Feature extraction from heart sound segment.
Mentions: The feature set used in this work consists of two parts: three features come from the envelope segments and another 32 come from heart sound segments. Features were obtained from heart sound segments using the same method as in [12] where d2 DWT coefficients decomposed for six levels using Daubechies-2 wavelet were partitioned into 32 non-overlapping windows and the signal energy of each window was used as a feature. This procedure is illustrated in Figure 10. It yields a 32-element feature vector for each heart sound segment. Another three features were obtained from envelope segments. Figure 11 shows an envelope segment. Each peak is marked with a circle and a peak interval and is shown by the arrows. The features derived from it were: the number of peaks, the average distance (in samples) between consecutive peaks, and the signal energy of the whole segment. The peak detection algorithm used to obtain the first two of these features is an improvement over simple thresholding; its pseudo code is shown below.

Bottom Line: Geometric mean was used as performance index.Average classification performance using ten-fold cross-validation was 0.92 for noise free case, 0.90 under white noise with 10 dB signal-to-noise ratio (SNR), and 0.90 under impulse noise up to 0.3 s duration.Further work include building a new training set recorded from actual patients, then further evaluate the method based on this new training set.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of Communication and Integrated Systems, Tokyo Institute of Technology, Japan 2-12-1-W9-108 Ookayama, Meguro-ku, Tokyo, Japan. toey123@gmail.com

ABSTRACT

Background: A new framework for heart sound analysis is proposed. One of the most difficult processes in heart sound analysis is segmentation, due to interference form murmurs.

Method: Equal number of cardiac cycles were extracted from heart sounds with different heart rates using information from envelopes of autocorrelation functions without the need to label individual fundamental heart sounds (FHS). The complete method consists of envelope detection, calculation of cardiac cycle lengths using auto-correlation of envelope signals, features extraction using discrete wavelet transform, principal component analysis, and classification using neural network bagging predictors.

Result: The proposed method was tested on a set of heart sounds obtained from several on-line databases and recorded with an electronic stethoscope. Geometric mean was used as performance index. Average classification performance using ten-fold cross-validation was 0.92 for noise free case, 0.90 under white noise with 10 dB signal-to-noise ratio (SNR), and 0.90 under impulse noise up to 0.3 s duration.

Conclusion: The proposed method showed promising results and high noise robustness to a wide range of heart sounds. However, more tests are needed to address any bias that may have been introduced by different sources of heart sounds in the current training set, and to concretely validate the method. Further work include building a new training set recorded from actual patients, then further evaluate the method based on this new training set.

Show MeSH