Heart sound classification with signal instant energy and stacked autoencoder network

https://doi.org/10.1016/j.bspc.2020.102211Get rights and content

Highlights

  • Denoising, segmentation, and classification are the basic technologies used in the analysis of heart sound signals.

  • These processes are both time consuming and have a lot of workload.

  • It has been shown that instant energies of heart sounds can be used for direct classification in this study.

  • The instant energies of the heart sounds were classified with Autoencoder network.

  • The proposed method can be easily used for the automatic diagnosis of heart disease from the heart sounds very effectively.

Abstract

Recently, different signal processing and classification methods have been tried to increase the success of classification for a heart sound analysis. For this purpose, in many studies, S1 and S2 segments of heart sounds were obtained by using methods such as Shannon energy, discrete time wavelet transform, Hilbert transform, and then classified. In this study, the use of signal energy, which is generally used to segment S1-S2 sounds in heart sounds, in direct classification was investigated. The instant energies of the heart sounds obtained by the resampled energy method were used directly for classification. The classification was done with a stacked autoencoder network. Experiments were carried out with the PASCAL B-training data set to test the performance of the proposed method. The results were compared with the data from previous studies for the same data set. As a result of the research, it is seen that the classification performance criterias obtained with the proposed method are as similar as the segmented classification. Thus, it was concluded that the instant energy of the heart sounds, and a stacked autoencoder networks can be very easily used for the diagnosis of heart diseases from heart sounds and a more efficient, and effective classification performance can be obtained.

Introduction

The most common method used by doctors to diagnose heart diseases is auscultation. Auscultation means listening to heart sounds with a stethoscope. Regular or irregular rhythms of heart sounds provide important information about heart diseases such as pathological cardiac status, valve diseases, and heart failure. Therefore, the heart sounds are of a great importance in the early diagnosis of heart diseases [[1], [2], [3]].

In fact, the heart sounds have four main components: S1, S2, S3, and S4 that repeat at regular intervals. However, the two main components of the heart sound are S1-Systole and S2-Diastole, and the sounds except these are mostly inaudible during auscultation. Fig. 1 gives an example of normal, murmur and extrasystole heart sounds in the dataset used in this study and S1-S2 sounds. Also abnormalities are shown on the figure. In studies, it has often been emphasized that these components must be well defined in order to easily and accurately diagnose heart diseases. Therefore, in many studies, heart sound signals were preprocessed and segmented into S1-S2 sounds by different methods, before classification. In such studies, the first step for automatic diagnosis of heart diseases from heart sounds is usually segmentation of heart sounds. Correct separation of systolic and diastolic regions is very important in the correct detection of cardiac sound components. Thus, pathological conditions in a heart sound can be obviously separated [4].

The aim of this study is to show that direct classification can be made easily with instant signal energy data, without segmentation by using more complex methods or needing feature selection with hybrid methods. It is also to discuss the obtained results by comparing with other examples. In this context, an exemplary study was performed in which the instant energies of the heart sounds calculated using the resampled energy method proposed by Deperlioglu [1] were classified with stacked autoencoder network. The rest of the article was organized as follows. In the second part, related works that reveal the purpose of the article were given. In the third section, a brief information about the data set and used methods were given. In the fourth section, results and comparative evaluations were made. In the last section, the contributions of the proposed method are presented.

Section snippets

Related works

As mentioned before, the majority of research on heart sounds has been conducted to improve classification success. Segmentation of S1 and S2 sounds in heart sounds, determination of peak values of S1 and S2 sounds, the use of different filtering methods and the use of different classification techniques are the most preferred ways to increase the classification success. Some case studies are given in Table 1.

In his study, Deperlioglu segmented S1-S2 sounds using the resampled energy method and

Materials and method

As can be seen from the examples above, signal energy is often used as a support method for segmentation. The aim of this study is to discuss the results obtained by demonstrating that direct classification can be done easily with signal energy data without the need of segmentation or more complex methods and algorithms. For this, the resampled energy method proposed by Deperlioglu was used [1]. For evaluation of proposed method, the classification was made with PASCAL B-training heart sounds

Results and discussion

The most appropriate SAEN configuration was obtained by making use of previous experiences for classification application. In the classification application of the signal energy data, 80% samples were used as a training data and 20% samples were used as a test data from PASCAL B-training data set. The maximum number of a learning epochs was 1000. Each classification study was repeated 20 times for different the training and test dataset.

In 20 trials, the lowest overall classification accuracy

Conclusion

In this study, it was found that the classification accuracy rate was almost same with the classification of the segmented data. From the obtained results, it is seen that the instant energy values could be classified directly as the segmented data or features. In addition, considering the energy values of heart sound data in the same dataset classified directly by ANN in the reference 4, it is seen that the SAEN classification is more successful. It is understood that the energies of the heart

CRediT authorship contribution statement

Omer Deperlioglu: Methodology.

Declaration of Competing Interest

We certify that we have NO affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers’ bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or non-financial interest (such as personal or professional relationships, affiliations, knowledge or beliefs) in the subject matter or materials discussed in this

References (31)

  • H. Ali et al.

    Heart sound signal modeling and segmentation based on improved Shannon energy envelogram using adaptive windows

    Asialink International Conference on Biomedical Engineering & Technology

    (2007)
  • J. Pedrosa et al.

    Automatic heart sound segmentation and murmur detection in pediatric phonocardiograms

  • A. Moukadem et al.

    High order statistics and time-frequency domain to classify heart sounds for subjects under cardiac stress test

    Comput. Math. Methods Med.

    (2015)
  • A. Raza et al.

    Heartbeat sound signal classification using deep learning

    Sensors

    (2019)
  • F. Demir et al.

    Towards the classification of heart sounds based on convolutional deep neural network

    Health Inf. Sci. Syst.

    (2019)
  • Cited by (27)

    • Research of heart sound classification using two-dimensional features

      2023, Biomedical Signal Processing and Control
      Citation Excerpt :

      Currently, convolutional neural networks [17], deep neural networks [18], and deep recurrent neural networks, among others, are primarily used by scholars in China and abroad to process heart sounds. Deperlioglu Omer et al. [19] proposed a method that uses instant energy and a stacked self-superimposed network to classify, and the average ac-curacy is 99.61%. Khan Faiq Ahmad et al. [20] presented the differences between the various operation methods on segmented and unsegmented signals, based on the time and frequency features extracted from the original signal.

    • PCG classification through spectrogram using transfer learning

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      The signal is divided into cycles for feature extraction from pre-trained AlexNet and a voting scheme is employed for final classification resulting in high classification rates. An analysis of methods listed in Tables 9 and 10 also reveals that deep neural networks have been employed for binary [19,35,42,45] as well as multi-class classification [32–34,36–41]. It is important to mention that from the viewpoint of application, binary classification can be used to recommend a subject to the cardiologist if the corresponding PCG signal is identified as abnormal in computerized analysis.

    View all citing articles on Scopus
    View full text