Practical guide for biomedical
signals analysis using machine
learning techniques
Prepared by:
Mostafa Khooshebast
9813353496
HSU
Phonocardiography (PCG)
Rangayyan, 2015
• PCG is the record of sonic vibrations of heart and blood circulation.
• The PCG is a sound signal associated with cardiohemic system (the heart
and blood together) activity.
• The PCG signal recording needs a transducer to convert the sound signal
into an electrical signal by utilizing microphones, pressure transducers, or
accelerometers, which can be positioned on the chest surface.
• The normal PCG signal indicates the overall state of the heart related to the
rhythm and contractility.
• CVDs show variations or different sounds and murmurs that can be utilized
in diagnosis.
Oweis, Hamad, & Shammout, 2014
• In the past, different concepts have been proposed to describe the origins of
cardiac cycle sounds once it was understood that the strength of the left
ventricle’s contraction had a substantial effect on the strength of the first heart
sound.
• Heart sounds are formed when a moving column of blood comes to a sudden
stop or decelerates significantly.
• The strength of a heart sound is based on the level of energy that the moving
column of blood has accomplished.
• The sudden deceleration causes a dissipation of energy that results in the
creation of vibrations affecting the contiguous cardiohemic mass.
• The factors included in the creation of the several heart sounds affecting the
acceleration and deceleration of columns of blood are different. These should be
separately considered for each sound, taking into account the pathophysiology
and the physiology of the involved phase of the cardiac cycle.
Chen, Wang, Shen, & Choy, 2012; Cherif, Debbal,
& BereksiReguig, 2010; Choi, Shin, & Park, 2011;
Safara, Doraisamy, Azman, Jantan, & Ranga, 2012
• Precise and early diagnosis of CVDs is crucial, which can be realized by
heart auscultation.
• It is a widely used technique to screen for physical anomalies of the
cardiovascular system. But perceiving pertinent features and creating
a diagnosis based on the heart sounds detected only by a stethoscope
is a skill that can take longer to learned and master.
• The accuracy and effectiveness of a diagnosis based on heart sound
auscultation can be enhanced significantly by employing digital signal
processing and machine learning techniques to investigate PCG
signals.
Carvalho et al., 2011 – Oweis et al., 2014 -
Zhong & Scalzo., 2013
• Several personal health systems have been implemented to effectively distinguish CVDs
and to support clinical decisions. Usually, heart sound auscultation and ECG are
employed for CVD diagnosis.
• But some heart diseases are hard to detect using ECG. Because heart sound signals are
complex and highly nonstationary in nature, it is not easy to analyze them in an
automated way.
• An auscultation technique employed to support medical doctors with an objective and
accurate interpretation of heart sounds can be employed that distinguishes heart sounds
(S1, S2, S3, and S4) during the heart cycle.
• The crucial points for signal analysis are the first heart sound (S1) and the second heart
sound (S2), as well as systolic and diastolic periods.
• Localization (segmentation) of heart sounds must be done before any analysis.
Techniques reported in the literature associated with heart sound segmentation include
the hidden Markov model. Other methods are based on signal homomorphism,
Eigenvalue spectrum, autoregression models, complexity-based segmentation, STFT, and
wavelet decomposition.
Carvalho et al., 2011 – Oweis et al., 2014 -
Zhong & Scalzo., 2013
• a multiband wavelet energy method was proposed for the
segmentation of S1 and S2 heart sounds. S1 and S2 can be utilized to
determine the heart sound type and distinguish abnormal heart
sounds.
• Wavelet decomposition can be successfully used to extract the
information hidden in S1 and S2 spectrums by choosing appropriate
subbands.
• The proposed method precisely localizes S1 and S2 to improve the
characteristic parameter calculation.
Babaei & Geranmayeh, 2009; Huiying, Sakari,
& Iiro, 1997 - Zhong and Scalzo (2013)
• Different methods have been utilized for heart sound segmentation of
normal and abnormal hearts, particularly for the S1 and S2.
• In Huiying et al. (1997), DWT was applied to phonocardiographic
signals, achieving a 93% accuracy.
• A multiscale wavelet decomposition with a threshold method was
implemented to improve the accuracy of the detection rate. The
suggested algorithm achieved an accuracy of around 92% on
abnormal heart sounds. Some approaches have used an unsupervised
training process for the task of robust and automatic detection and
identification of events of interest in the major heart sounds from
unlabeled PCGs.
Oweis et al. (2014)
• Teager Energy Operator (TEO) with homomorphic filtering was employed
for heart sound segmentation and obtained 98.3% accuracy.
• The proposed segmentation method divides the signal into periods
showing the two dominant heart sounds, S1 and S2, and time durations of
the systole and diastole.
• The proposed method simplifies the presentation of heart sound signals
obtained with a digital stethoscope. The resulting signals are more
meaningful and easier to judge even for an unskilled doctor, thus
eliminating the need for an expert physician.
• The main advantages of this method are employability in a noisy
environment, reduction of the time needed for the examination process,
reliability, and ease of use.
• The proposed method gives a sensitivity level of 99.1%, positive
predictivity of 97.7%, specificity of 97.43%, and accuracy of 98.3%.
Choi et al., 2011
• Medical doctors employ the PCG signals more competently once these signals are
presented visually instead of using a conventional stethoscope.
• A clinician can utilize this significant signal with valuable prognostic and
diagnostic information for the diagnosis of CVDs. Even though PCG signal analysis
by auscultation is an appropriate clinical tool, the heart sound signals are
nonstationary and very complex. Hence the analysis of this signal in a time or
frequency domain is not easy. Generally, there are two main procedures required
before classification: signal processing and feature extraction.
• Several signal processing algorithms are utilized in the literature such as discrete
Fourier transform, short time Fourier transform, Wigner distribution, Hilbert
transform, continuous wavelet transform, DWT, WPT, TQWT, DTCWT, EMD, and
ensemble EMD. Actually, the Fourier and wavelet family of transforms are usually
employed for PCG signal processing. Because the wavelet achieves a practical
decomposition in both the time and frequency domain, it is one of the best
transforms to analyze nonstationary and transient signals, such as PCG.
Safara, Doraisamy, Azman, Jantan, et al., 2012
• Feature extraction is critical to obtain better classification
performance. A different set of features based on wavelet transforms
was defined recently to classify various types of heart sounds and
murmurs.
• For example, Ahlstrom, Hult, et al. (2006) extracted energy and
entropy from DWT of the PCG signals to classify systolic murmurs.
• Cherif et al. (2010) proposed a feature extraction method in the TF
domain for the investigation of heart sounds. They utilized DWT and
WPT in the PCG signal analysis.
Ahlstrom, Hult, et al. (2006)
• Heart murmurs are usually the initial indication of a heart valve
abnormality, and they are generally realized at auscultation during the
primary health care. To distinguish a normal murmur from a
pathological murmur, a smart stethoscope with decision-support
capabilities would be of great value.
• PCG signals acquired from different patients with mitral insufficiency,
aortic valve stenosis, or normal murmurs, and the data were analyzed
were employed to find an appropriate feature subset for automatic
classification of heart murmurs.
• Furthermore, fractal dimensions, wavelets, Shannon energy, and
recurrence quantification analysis were employed for feature
extraction. NN classifier achieved 86% correct classifications rate.
Gupta, Palaniappan, Swaminathan, and
Krishnan (2007)
• a new technique was proposed for the segmentation of PCG signals
into single cardiac cycle (S1 and S2) employing K-means clustering
and homomorphic filtering.
• Wavelet transform was used for feature extraction, and NNs were
employed for classification of normal, diastolic murmur, and systolic
murmur.
• The S3 and S4 heart sounds are two abnormal heart sound
components designated in the heart.
Tseng, Ko, and Jaw (2012)
• an adaptive TF analysis method (Hilbert-Huang Transform) was
proposed to find the existence of S3 and S4 in PCG signals.
• By utilizing the proposed algorithm, 90.3% of heart sound cycles with
S3 were detected, 9.6% were missed, and 9.6% were false-positive;
94% of S4 were detected, 5.5% were missed, and 16% were false-
positive.
Yuenyong, Nishihara, Kongprawechnon, and
Tungpimolrut (2011)
• A new framework for heart sound analysis was proposed. One of the
most crucial steps in PCG signal analysis is segmentation, because of
the interference of murmurs.
• The framework includes envelope detection for the calculation of
cardiac cycle lengths by employing autocorrelation of envelope
signals.
• DWT utilized features extraction, principal component analysis was
utilized for dimension reduction, and NN bagging predictors were
employed for classification. Maximum classification performance
employing 10-fold cross-validation was 92%.
Choi (2008) and Jiang (2010)
• WPT and SVM classifier were employed to distinguish normal heart sounds from
murmurs; 96% sensitivity and 100% specificity was achieved.
• A novel PCG spectral analysis technique employing the normalized AR power spectral
density curve with the SVM classifier to classify cardiac sound murmurs, normal sound
signals, and abnormal sound signals (split sounds, aortic insufficiency, aortic stenosis,
atrial fibrillation, mitral stenosis, and mitral regurgitation). As a result, the proposed PCG
spectral analysis method achieved 99.5% sensitivity and 99.9% specificity in classifying
normal and abnormal heart sounds.
• A novel PCG signal analysis technique was proposed for mitral and aortic abnormal
murmurs using WPD. Normal heart sounds were acquired from subjects without any
heart problems by employing a wireless electric stethoscope. Abnormal murmurs were
aortic stenosis and mitral stenosis. They utilized the position index and the maximum
peak frequency of the WPD coefficient associated with the maximum peak frequency,
and the ratios of entropy information and the wavelet energy to accomplish higher
accuracy for the heart murmur detection. The proposed abnormal murmur detection
technique achieved a classification performance of 99.78% specificity and 99.43%
sensitivity.
Ahlstrom, Heoglund, et al. (2006) - Babaei
and Geranmayeh (2009)
• NN was employed to differentiate normal murmurs from aortic
stenosis, and a specificity of 88% and a sensitivity of 90% were
achieved.
• A MLPNN was used for the classification of PCG signals to
differentiate murmurs and achieved 94.24% accuracy.
Babaei and Geranmayeh (2009)
• Cardiac auscultatory ability of medical doctors is critical for correct
diagnosis of heart diseases. Several abnormal heart sounds with
indistinguishable main specifications and diverse details, including ambient
noise, are essentially similar, thus it is important to train, evaluate, and
improve the skills of medical students in identifying and differentiating the
main indications of the cardiac diseases.
• A useful wavelet-based multiresolution algorithm was employed to extract
the key characteristics of three well-known heart valve disorders:
pulmonary stenosis, aortic stenosis, and aortic insufficiency sounds.
• An ANN classifier and Daubechies wavelet filter with five decomposition
levels were employed for the most prominent diseases. The proposed ANN
achieved 94.42% classification accuracy.
Herzig, Bickel, Eitan, and Intrator (2015)
• proposed new cardiac monitoring based on PCG signal analysis.
Specifically, they studied two morphological features and their
relations with physical changes.
• The proposed system was verified using data taken from different
patients during laparoscopic surgeries. During laparoscopic surgery,
externally induced cardiac stress allowed analysis of each patient with
regard to his or her own baseline.
• Furthermore, they revealed that the proposed features vary during
cardiac stress, and the variation is more significant for patients with
cardiac complications.
Ahlstrom,Hult, et al. (2006)
• Different features can be extracted based on wavelet transform to distinguish murmurs
from normal heart sounds or to differentiate different types of murmurs.
• A method including entropy, fractal dimension, and energy was proposed to extract
features from PCG signals to classify systolic murmurs.
• Heart murmurs are generally the first sign of pathological variations of the heart valves,
and they can be detected during auscultation in primary health care.
• Separation of a pathological murmur from a physiological murmur is not an easy task,
hence, an “intelligent stethoscope” with decision-support capabilities can be of great
value.
• PCG signals were collected from several patients with mitral insufficiency, aortic valve
stenosis, or physiological murmurs to find an appropriate feature subset for automatic
classification of heart murmurs.
• Wavelets, Shannon energy, recurrence quantification, and fractal dimensions analysis
were employed for feature extraction. The selected multidomain subset achieved the
best results, with 86% correct classifications employing an NN classifier.
Choi (2008)
• WPD and the SVM classifier were utilized for valvular heart disorder
(VHD) detection.
• The WPE was utilized to include distribution information of energy
through the whole frequency range of PCG signals.
• Moreover, the PCG signal parameters (meanWPE and stdWPE) of the
position indices of the terminal nodes were proposed as a feature.
• The experimental results of the proposed VHD detection method
achieved a specificity of 96% and a sensitivity of 100% for both the
training and testing data.
Cherif et al. (2010)
• Medical doctors can utilize the PCG signal once it is presented visually,
rather than by a conventional stethoscope. This signal delivers
significant prognostic diagnostic information to the clinician. Although
PCG signal analysis by auscultation is suitable as a clinical tool, it is
not easy to analyze in the time or frequency domain as they are
nonstationary and more complex.
• TF feature extraction was studied by Cherif for the recognition of
heart sounds. They highlighted the importance of the choice of
wavelet in PCG signal analysis using the DWT and the WPT.
• The performance of the DWT and the WPT in PCG signal analysis was
assessed and discussed.
Homaeinezhad, Sabetian, Feizollahi, Ghaffari,
and Rahmani (2012)
• A mathematical modeling of the cardiac system with multiple
measurement signals was presented: PCG, ECG, and arterial blood pressure
(ABP).
• Moreover, the PCG framework can produce the S4-S1-S2-S3 cycles in
normal and cardiac illness conditions such as insufficiency, gallop,
regurgitation, and stenosis.
• The amplitude and frequency content (5–700Hz) of individual sound and
variation patterns can be defined in the PCG model.
• The three models were realized to produce artificial signals with different
cardiac abnormalities for quantitative recognition and performance
evaluation of numerous PCG, ECG, and ABP signals based on theDWT,
Hilbert transform, geometric features, and the principal components
analyzed by geometric index.
Naseri and Homaeinezhad (2012)
• A PCG signal measurement quality assessment technique was developed that consists of three
main steps: preprocessing, quality assessment, and advanced measurement algorithms.
• The preprocessing step consists of normalization, wavelet-based threshold denoising, and
baseline wander removal.
• The quality assessment routine contains the energy and noise level of the PCG signal.
• The advanced quality measurement step is principally employed by the S1 and S2 sound
intervals.
• The developed algorithm achieved 95.25% classification accuracy. Furthermore, they proposed a
new framework to classify PCG sounds.
• First, the PCG signal undergoes preprocessing, then two windows were moved into the
preprocessed data, and two frequency and amplitude-based features were extracted from these
segments in each analysis window. After feature extraction, combining these features created a
synthetic decision-making basis. To classify the defined PCG sounds, first S1 and S2 were
distinguished. Then, a new decision statistic was utilized to distinguish infrequent S3 and S4
sounds.
• The proposed framework achieved a positive predictive value of 98.60% and average sensitivity
of 99.00%.
Safara, Doraisamy, Azman, and Jantan (2012)
• Spectral, temporal, and geometric features were combined for PCG signal classification.
• The feature set is composed of zero-crossing rates as the temporal feature, and spectral flux,
spectral energy entropy, spectral roll-off, and spectral centroid as the spectral features.
• Moreover, curve length, area under curve, summation of the first order derivatives, summation of
the second order derivatives, and centralized mean square values were utilized in the feature set
as well.
• They introduced a new entropy to investigate the heart sounds in classification of five types of
heart sounds and murmurs which are normal, mitral stenosis, mitral regurgitation, aortic stenosis,
and aortic regurgitation.
• WPT was employed for extracting features from the heart sound, and the entropy was calculated
to derive feature vectors.
• The proposed scheme achieved 96.94% accuracy with BayesNet.
• A multilevel basis selection (MLBS) was proposed to preserve the most informative basis of a WPT
tree by eliminating the less informative basis by using three exclusion criteria: noise frequency,
energy threshold, and frequency range. MLBS achieved an accuracy of 97.56% for classifying
normal heart sounds, aortic regurgitation, mitral regurgitation, and aortic stenosis (Safara et al.,
2013).
Ramovic, Bandic, Kevric, Germovic, and
Subasi (2017)
• The PCG taken from normal subjects generally includes two separate tones,
S1 and S2. Moreover, an auscultation technique utilized to supply medical
doctors with objective and precise interpretation of heart sounds can be
employed to distinguish four sounds, namely, S1, S2, S3, and S4, during the
heart cycle.
• A framework was proposed to distinguish four heartbeats efficiently
utilizing the combination of multiscale wavelet transform and TEO to
improve the accuracy of the recognition process.
• Wavelet transform was added to the module in the form of creating several
combinations for a signal on which TEO is applied.
• The aim of combining wavelet and TEO was to explore how different details
taken from wavelet transform affect the success of TEO in detecting S1, S2,
S3, and S4 heart sounds.

Pcg

  • 1.
    Practical guide forbiomedical signals analysis using machine learning techniques Prepared by: Mostafa Khooshebast 9813353496 HSU Phonocardiography (PCG)
  • 2.
    Rangayyan, 2015 • PCGis the record of sonic vibrations of heart and blood circulation. • The PCG is a sound signal associated with cardiohemic system (the heart and blood together) activity. • The PCG signal recording needs a transducer to convert the sound signal into an electrical signal by utilizing microphones, pressure transducers, or accelerometers, which can be positioned on the chest surface. • The normal PCG signal indicates the overall state of the heart related to the rhythm and contractility. • CVDs show variations or different sounds and murmurs that can be utilized in diagnosis.
  • 3.
    Oweis, Hamad, &Shammout, 2014 • In the past, different concepts have been proposed to describe the origins of cardiac cycle sounds once it was understood that the strength of the left ventricle’s contraction had a substantial effect on the strength of the first heart sound. • Heart sounds are formed when a moving column of blood comes to a sudden stop or decelerates significantly. • The strength of a heart sound is based on the level of energy that the moving column of blood has accomplished. • The sudden deceleration causes a dissipation of energy that results in the creation of vibrations affecting the contiguous cardiohemic mass. • The factors included in the creation of the several heart sounds affecting the acceleration and deceleration of columns of blood are different. These should be separately considered for each sound, taking into account the pathophysiology and the physiology of the involved phase of the cardiac cycle.
  • 4.
    Chen, Wang, Shen,& Choy, 2012; Cherif, Debbal, & BereksiReguig, 2010; Choi, Shin, & Park, 2011; Safara, Doraisamy, Azman, Jantan, & Ranga, 2012 • Precise and early diagnosis of CVDs is crucial, which can be realized by heart auscultation. • It is a widely used technique to screen for physical anomalies of the cardiovascular system. But perceiving pertinent features and creating a diagnosis based on the heart sounds detected only by a stethoscope is a skill that can take longer to learned and master. • The accuracy and effectiveness of a diagnosis based on heart sound auscultation can be enhanced significantly by employing digital signal processing and machine learning techniques to investigate PCG signals.
  • 5.
    Carvalho et al.,2011 – Oweis et al., 2014 - Zhong & Scalzo., 2013 • Several personal health systems have been implemented to effectively distinguish CVDs and to support clinical decisions. Usually, heart sound auscultation and ECG are employed for CVD diagnosis. • But some heart diseases are hard to detect using ECG. Because heart sound signals are complex and highly nonstationary in nature, it is not easy to analyze them in an automated way. • An auscultation technique employed to support medical doctors with an objective and accurate interpretation of heart sounds can be employed that distinguishes heart sounds (S1, S2, S3, and S4) during the heart cycle. • The crucial points for signal analysis are the first heart sound (S1) and the second heart sound (S2), as well as systolic and diastolic periods. • Localization (segmentation) of heart sounds must be done before any analysis. Techniques reported in the literature associated with heart sound segmentation include the hidden Markov model. Other methods are based on signal homomorphism, Eigenvalue spectrum, autoregression models, complexity-based segmentation, STFT, and wavelet decomposition.
  • 6.
    Carvalho et al.,2011 – Oweis et al., 2014 - Zhong & Scalzo., 2013 • a multiband wavelet energy method was proposed for the segmentation of S1 and S2 heart sounds. S1 and S2 can be utilized to determine the heart sound type and distinguish abnormal heart sounds. • Wavelet decomposition can be successfully used to extract the information hidden in S1 and S2 spectrums by choosing appropriate subbands. • The proposed method precisely localizes S1 and S2 to improve the characteristic parameter calculation.
  • 7.
    Babaei & Geranmayeh,2009; Huiying, Sakari, & Iiro, 1997 - Zhong and Scalzo (2013) • Different methods have been utilized for heart sound segmentation of normal and abnormal hearts, particularly for the S1 and S2. • In Huiying et al. (1997), DWT was applied to phonocardiographic signals, achieving a 93% accuracy. • A multiscale wavelet decomposition with a threshold method was implemented to improve the accuracy of the detection rate. The suggested algorithm achieved an accuracy of around 92% on abnormal heart sounds. Some approaches have used an unsupervised training process for the task of robust and automatic detection and identification of events of interest in the major heart sounds from unlabeled PCGs.
  • 8.
    Oweis et al.(2014) • Teager Energy Operator (TEO) with homomorphic filtering was employed for heart sound segmentation and obtained 98.3% accuracy. • The proposed segmentation method divides the signal into periods showing the two dominant heart sounds, S1 and S2, and time durations of the systole and diastole. • The proposed method simplifies the presentation of heart sound signals obtained with a digital stethoscope. The resulting signals are more meaningful and easier to judge even for an unskilled doctor, thus eliminating the need for an expert physician. • The main advantages of this method are employability in a noisy environment, reduction of the time needed for the examination process, reliability, and ease of use. • The proposed method gives a sensitivity level of 99.1%, positive predictivity of 97.7%, specificity of 97.43%, and accuracy of 98.3%.
  • 9.
    Choi et al.,2011 • Medical doctors employ the PCG signals more competently once these signals are presented visually instead of using a conventional stethoscope. • A clinician can utilize this significant signal with valuable prognostic and diagnostic information for the diagnosis of CVDs. Even though PCG signal analysis by auscultation is an appropriate clinical tool, the heart sound signals are nonstationary and very complex. Hence the analysis of this signal in a time or frequency domain is not easy. Generally, there are two main procedures required before classification: signal processing and feature extraction. • Several signal processing algorithms are utilized in the literature such as discrete Fourier transform, short time Fourier transform, Wigner distribution, Hilbert transform, continuous wavelet transform, DWT, WPT, TQWT, DTCWT, EMD, and ensemble EMD. Actually, the Fourier and wavelet family of transforms are usually employed for PCG signal processing. Because the wavelet achieves a practical decomposition in both the time and frequency domain, it is one of the best transforms to analyze nonstationary and transient signals, such as PCG.
  • 10.
    Safara, Doraisamy, Azman,Jantan, et al., 2012 • Feature extraction is critical to obtain better classification performance. A different set of features based on wavelet transforms was defined recently to classify various types of heart sounds and murmurs. • For example, Ahlstrom, Hult, et al. (2006) extracted energy and entropy from DWT of the PCG signals to classify systolic murmurs. • Cherif et al. (2010) proposed a feature extraction method in the TF domain for the investigation of heart sounds. They utilized DWT and WPT in the PCG signal analysis.
  • 11.
    Ahlstrom, Hult, etal. (2006) • Heart murmurs are usually the initial indication of a heart valve abnormality, and they are generally realized at auscultation during the primary health care. To distinguish a normal murmur from a pathological murmur, a smart stethoscope with decision-support capabilities would be of great value. • PCG signals acquired from different patients with mitral insufficiency, aortic valve stenosis, or normal murmurs, and the data were analyzed were employed to find an appropriate feature subset for automatic classification of heart murmurs. • Furthermore, fractal dimensions, wavelets, Shannon energy, and recurrence quantification analysis were employed for feature extraction. NN classifier achieved 86% correct classifications rate.
  • 12.
    Gupta, Palaniappan, Swaminathan,and Krishnan (2007) • a new technique was proposed for the segmentation of PCG signals into single cardiac cycle (S1 and S2) employing K-means clustering and homomorphic filtering. • Wavelet transform was used for feature extraction, and NNs were employed for classification of normal, diastolic murmur, and systolic murmur. • The S3 and S4 heart sounds are two abnormal heart sound components designated in the heart.
  • 13.
    Tseng, Ko, andJaw (2012) • an adaptive TF analysis method (Hilbert-Huang Transform) was proposed to find the existence of S3 and S4 in PCG signals. • By utilizing the proposed algorithm, 90.3% of heart sound cycles with S3 were detected, 9.6% were missed, and 9.6% were false-positive; 94% of S4 were detected, 5.5% were missed, and 16% were false- positive.
  • 14.
    Yuenyong, Nishihara, Kongprawechnon,and Tungpimolrut (2011) • A new framework for heart sound analysis was proposed. One of the most crucial steps in PCG signal analysis is segmentation, because of the interference of murmurs. • The framework includes envelope detection for the calculation of cardiac cycle lengths by employing autocorrelation of envelope signals. • DWT utilized features extraction, principal component analysis was utilized for dimension reduction, and NN bagging predictors were employed for classification. Maximum classification performance employing 10-fold cross-validation was 92%.
  • 15.
    Choi (2008) andJiang (2010) • WPT and SVM classifier were employed to distinguish normal heart sounds from murmurs; 96% sensitivity and 100% specificity was achieved. • A novel PCG spectral analysis technique employing the normalized AR power spectral density curve with the SVM classifier to classify cardiac sound murmurs, normal sound signals, and abnormal sound signals (split sounds, aortic insufficiency, aortic stenosis, atrial fibrillation, mitral stenosis, and mitral regurgitation). As a result, the proposed PCG spectral analysis method achieved 99.5% sensitivity and 99.9% specificity in classifying normal and abnormal heart sounds. • A novel PCG signal analysis technique was proposed for mitral and aortic abnormal murmurs using WPD. Normal heart sounds were acquired from subjects without any heart problems by employing a wireless electric stethoscope. Abnormal murmurs were aortic stenosis and mitral stenosis. They utilized the position index and the maximum peak frequency of the WPD coefficient associated with the maximum peak frequency, and the ratios of entropy information and the wavelet energy to accomplish higher accuracy for the heart murmur detection. The proposed abnormal murmur detection technique achieved a classification performance of 99.78% specificity and 99.43% sensitivity.
  • 16.
    Ahlstrom, Heoglund, etal. (2006) - Babaei and Geranmayeh (2009) • NN was employed to differentiate normal murmurs from aortic stenosis, and a specificity of 88% and a sensitivity of 90% were achieved. • A MLPNN was used for the classification of PCG signals to differentiate murmurs and achieved 94.24% accuracy.
  • 17.
    Babaei and Geranmayeh(2009) • Cardiac auscultatory ability of medical doctors is critical for correct diagnosis of heart diseases. Several abnormal heart sounds with indistinguishable main specifications and diverse details, including ambient noise, are essentially similar, thus it is important to train, evaluate, and improve the skills of medical students in identifying and differentiating the main indications of the cardiac diseases. • A useful wavelet-based multiresolution algorithm was employed to extract the key characteristics of three well-known heart valve disorders: pulmonary stenosis, aortic stenosis, and aortic insufficiency sounds. • An ANN classifier and Daubechies wavelet filter with five decomposition levels were employed for the most prominent diseases. The proposed ANN achieved 94.42% classification accuracy.
  • 18.
    Herzig, Bickel, Eitan,and Intrator (2015) • proposed new cardiac monitoring based on PCG signal analysis. Specifically, they studied two morphological features and their relations with physical changes. • The proposed system was verified using data taken from different patients during laparoscopic surgeries. During laparoscopic surgery, externally induced cardiac stress allowed analysis of each patient with regard to his or her own baseline. • Furthermore, they revealed that the proposed features vary during cardiac stress, and the variation is more significant for patients with cardiac complications.
  • 19.
    Ahlstrom,Hult, et al.(2006) • Different features can be extracted based on wavelet transform to distinguish murmurs from normal heart sounds or to differentiate different types of murmurs. • A method including entropy, fractal dimension, and energy was proposed to extract features from PCG signals to classify systolic murmurs. • Heart murmurs are generally the first sign of pathological variations of the heart valves, and they can be detected during auscultation in primary health care. • Separation of a pathological murmur from a physiological murmur is not an easy task, hence, an “intelligent stethoscope” with decision-support capabilities can be of great value. • PCG signals were collected from several patients with mitral insufficiency, aortic valve stenosis, or physiological murmurs to find an appropriate feature subset for automatic classification of heart murmurs. • Wavelets, Shannon energy, recurrence quantification, and fractal dimensions analysis were employed for feature extraction. The selected multidomain subset achieved the best results, with 86% correct classifications employing an NN classifier.
  • 20.
    Choi (2008) • WPDand the SVM classifier were utilized for valvular heart disorder (VHD) detection. • The WPE was utilized to include distribution information of energy through the whole frequency range of PCG signals. • Moreover, the PCG signal parameters (meanWPE and stdWPE) of the position indices of the terminal nodes were proposed as a feature. • The experimental results of the proposed VHD detection method achieved a specificity of 96% and a sensitivity of 100% for both the training and testing data.
  • 21.
    Cherif et al.(2010) • Medical doctors can utilize the PCG signal once it is presented visually, rather than by a conventional stethoscope. This signal delivers significant prognostic diagnostic information to the clinician. Although PCG signal analysis by auscultation is suitable as a clinical tool, it is not easy to analyze in the time or frequency domain as they are nonstationary and more complex. • TF feature extraction was studied by Cherif for the recognition of heart sounds. They highlighted the importance of the choice of wavelet in PCG signal analysis using the DWT and the WPT. • The performance of the DWT and the WPT in PCG signal analysis was assessed and discussed.
  • 22.
    Homaeinezhad, Sabetian, Feizollahi,Ghaffari, and Rahmani (2012) • A mathematical modeling of the cardiac system with multiple measurement signals was presented: PCG, ECG, and arterial blood pressure (ABP). • Moreover, the PCG framework can produce the S4-S1-S2-S3 cycles in normal and cardiac illness conditions such as insufficiency, gallop, regurgitation, and stenosis. • The amplitude and frequency content (5–700Hz) of individual sound and variation patterns can be defined in the PCG model. • The three models were realized to produce artificial signals with different cardiac abnormalities for quantitative recognition and performance evaluation of numerous PCG, ECG, and ABP signals based on theDWT, Hilbert transform, geometric features, and the principal components analyzed by geometric index.
  • 23.
    Naseri and Homaeinezhad(2012) • A PCG signal measurement quality assessment technique was developed that consists of three main steps: preprocessing, quality assessment, and advanced measurement algorithms. • The preprocessing step consists of normalization, wavelet-based threshold denoising, and baseline wander removal. • The quality assessment routine contains the energy and noise level of the PCG signal. • The advanced quality measurement step is principally employed by the S1 and S2 sound intervals. • The developed algorithm achieved 95.25% classification accuracy. Furthermore, they proposed a new framework to classify PCG sounds. • First, the PCG signal undergoes preprocessing, then two windows were moved into the preprocessed data, and two frequency and amplitude-based features were extracted from these segments in each analysis window. After feature extraction, combining these features created a synthetic decision-making basis. To classify the defined PCG sounds, first S1 and S2 were distinguished. Then, a new decision statistic was utilized to distinguish infrequent S3 and S4 sounds. • The proposed framework achieved a positive predictive value of 98.60% and average sensitivity of 99.00%.
  • 24.
    Safara, Doraisamy, Azman,and Jantan (2012) • Spectral, temporal, and geometric features were combined for PCG signal classification. • The feature set is composed of zero-crossing rates as the temporal feature, and spectral flux, spectral energy entropy, spectral roll-off, and spectral centroid as the spectral features. • Moreover, curve length, area under curve, summation of the first order derivatives, summation of the second order derivatives, and centralized mean square values were utilized in the feature set as well. • They introduced a new entropy to investigate the heart sounds in classification of five types of heart sounds and murmurs which are normal, mitral stenosis, mitral regurgitation, aortic stenosis, and aortic regurgitation. • WPT was employed for extracting features from the heart sound, and the entropy was calculated to derive feature vectors. • The proposed scheme achieved 96.94% accuracy with BayesNet. • A multilevel basis selection (MLBS) was proposed to preserve the most informative basis of a WPT tree by eliminating the less informative basis by using three exclusion criteria: noise frequency, energy threshold, and frequency range. MLBS achieved an accuracy of 97.56% for classifying normal heart sounds, aortic regurgitation, mitral regurgitation, and aortic stenosis (Safara et al., 2013).
  • 25.
    Ramovic, Bandic, Kevric,Germovic, and Subasi (2017) • The PCG taken from normal subjects generally includes two separate tones, S1 and S2. Moreover, an auscultation technique utilized to supply medical doctors with objective and precise interpretation of heart sounds can be employed to distinguish four sounds, namely, S1, S2, S3, and S4, during the heart cycle. • A framework was proposed to distinguish four heartbeats efficiently utilizing the combination of multiscale wavelet transform and TEO to improve the accuracy of the recognition process. • Wavelet transform was added to the module in the form of creating several combinations for a signal on which TEO is applied. • The aim of combining wavelet and TEO was to explore how different details taken from wavelet transform affect the success of TEO in detecting S1, S2, S3, and S4 heart sounds.