Classification of Heart Sounds Using Grey Level Co-occurrence Matrix and Logistic Regression

Every human being's heart is their central organ. The purpose of the heart is to circulate blood throughout the body. The heart is situated between the fourth and sixth ribs on the left side of the chest. Because the heart beats instinctively and autonomously, it is an organ whose function is not controlled by the brain [1]. The diagnosis results are highly influenced by the subjectivity of the physician, as they need expertise and sensitivity to diagnose normal or abnormal heart sounds, often known as pathological murmurs. An essential diagnostic technique for cardiac disease is auscultation, a low-cost screening technique. Issues with valvular function and heart disease diagnosis are related to heart valves. The range of frequencies for normal heart sounds is 20 Hz to 500 Hz, whereas the range for aberrant heart sounds is up to 1000 Hz. A murmur is the result of an improper valve opening or stenosis, which makes blood flow through a restricted aperture and causes blood to backflow [2]. Digital signal processing techniques have been used in numerous researches to diagnose cardiac problems because heart sound assessment is still highly subjective [3].

Early detection of abnormal heart sounds is needed and can also be used to support healthcare professionals [4]. Several studies have been carried out to develop heart sound classification applications. Hamidi et al. classified cardiac sound signals using fractal dimension and curve fitting [5]. Using the closest neighbor as a classifier, the approach was evaluated on three datasets and yielded the maximum accuracy of 92%. The fractal approach was also employed by Juniati et al. [6] and Komalasari et al. [7] with a promising accuracy of up to 100% for regular and murmurs. Fauzi et al., meantime, categorizes cardiac sounds using statistical traits and empirical mode decomposition (EMD). When using KNN as a classifier [8], the highest accuracy is 98.2%. Other researchers have also processed cardiac sounds using the time-frequency domain. Wang et al. extracted features using wavelet-time entropy [2]. Excellent classification findings have been obtained from some of these experiments. However, another viewpoint is required to visualize and analyze heart sound texture while conducting research. Similar to the subsequent study, a heart sound spectrogram acquired from STFT was processed using the Grey Level Difference Matrix (GLDM) [9]. 73% accuracy was the best accuracy achieved when using SVM as a classifier. Nonetheless, the findings of this study are still regarded as inadequate and have room for improvement.

GLDM seeks to extract the heart sound spectrogram's texture [9]. Other techniques, such GLCM, can be applied in addition to GLDM. This is the reason that the GLCM approach was evaluated in this work for feature extraction in the classification of heart sounds. Furthermore, visualizing heart sounds using GLCM can significantly enhance the diagnostic heart sound analysis and categorization process. GLCM is used for feature extraction in image processing, which is then used for classification with machine learning. The results of previous research are good, especially for biomedical image processing, with an accuracy level above 80% [10-13]. Previous research was also carried out for sound classification with GLCM, with accuracy results of up to 90% [14]. This could be the basis for a similar development of heart sounds to classify abnormal classification. GLCM uses textural features from heart sound spectrograms to discover patterns linked to various heart conditions [15]. To convert the signal from one dimension to two dimensions, heart sounds are converted into the time-frequency domain. After that, normalization is done to make the SFTT value equal to the image's pixel value. The transformed image from the heart sound spectrogram is subjected to GLCM, and its properties are computed. It is anticipated that the accuracy that results would be better than the GLDM approach in earlier studies. This study makes use of a number of models created to identify which GLCM feature produces the most accurate model for categorizing heart sounds. All GLCM features such as Dissimilarity, Correlation, Homogeneity, Contrast, ASM and Energy with directions 0°, 45°, 90° and 135°.

2a.png

(a)

2b.png

(b)

Figure 2. (a) Normal heart sound; (b) Spectrogram of normal heart sound

3a.png

(a)

3b.png

(b)

Figure 3. (a) Murmur heart sound; (b) Spectrogram of murmur heart sound

Figures 2 and 3 present visual representations of both normal cardiac sound signals and murmurs, together with the corresponding spectrums that come from these acoustic phenomena. A spectrogram can be visualized as a two-dimensional image that exhibits a texture indicative of the energy distribution of a signal over both the time and frequency domains. Prior to performing feature extraction using GLCM, a conversion process is undertaken to transform the values obtained from the spectrogram into integer values ranging from 0 to 255, as has been previously demonstrated in prior research [9].

4.png

Figure 4. Distribution of several GLCM in the 0° direction (a) Dissimilarity; (b) Correlation; (c) Homogeneity; (d) Contrast; (e) ASM; (f) Energy

As stated earlier, this study utilized six features that were computed at four different angles: 0°, 45°, 90°, and 135°. In Figure 4, a boxplot illustrates the computed values distribution of data for each class for each feature in the 0° orientation. There is a tendency for all distribution data per each class trait to overlap with one another, as shown in each graph for each feature. This pattern of distribution data per class suggests that this could lead to a comparatively reduced level of accuracy.

Table 1 displays the results of classification tests for each feature in each direction. The contrast in the 90° direction produced the highest accuracy when utilizing a single characteristic, at 75%, while the 90° direction dissimilarity gave the second-highest accuracy, at 74%. This outcome is better than earlier studies that extracted features using GLDM with only two classes of data, yielding an accuracy of 73% for features in a single direction [9]. The pretty good accuracy obtained with only one feature demonstrates the significant potential for increased accuracy by including other parameters.

Table 1. Training results for all features

Next, testing was done with all features (24 features) and all degree of each type of feature extraction (every 4 features), with training data comprising 90% and test data comprising 10% of the 455 total features. It is evident that even with an increase in the number of attributes, the rise is not statistically significant. In addition to using all the features, other classification methods, such as SVM, KNN, and Logistic regression, are also compared. Determining the optimal parameters for all classifier is the next step towards improving accuracy. This is why GridSearchCV, a method for selecting the best parameters from a range of parameters applied to a subset of data, is used with third cross validation [24]. It was found that logistic regression obtained greater accuracy than SVM and KNN. The Dissimilarity and Contrast features have the best accuracy from the previous model test in Table 1, and this combined test also show in Table 2, that the highest accuracy is in the same feature. Dissimilarity had the highest score, 82.64%, among the other categories.

Table 2. All degree GLCM feature extraction accuracy machine learning

Dissimilarity is the best feature that can be extracted from GLCM. Therefore, we reserve this feature's use for testing many other performance metrics. The performance results from the all machine learning model using the dissimilarity feature are shown in Figure 5 and Table 3, which in this example include the confusion matrix, precision, recall, and F1 score.

Because each class's data size is not balanced, measurements using these parameters are required. These findings show that although the accuracy of the model with feature dissimilarity is generally not outstanding, the precision, recall, and F1 Score values for the normal class are still excellent when compared to other classes, is shown in Figure 5. However, only the logistic regression method could classify the murmur class well, as seen in Figure 5(a) and Table 3, with a recall value reaching 100%. Meanwhile, for SVM, the model cannot classify this class well with a recall value of even 0%, as can be seen in Figure 5(b) and Table 3. The artifact class is also not classified well for each model, even though the boxplot in Figure 4 clearly illustrates this class's pattern.

5a.png

(a)

5b.png

(b)

5c.png

(c)

Figure 5. Confusion matrix dissimilarity feature from machine learning method: (a) SVM; (b) Logistic regression; (c) KNN

In Figure 6, other measurement results are carried out using the ROC curve from one class to another (one versus rest classifier) from logistic regression as best machine learning model. From the curve, the area under the ROC curve (AUC) value for the normal and artifact classes is considered very good because it is close to 1, but had lower sensitivity for the murmur class, with an AUC value of 0.7. The average value of the AUC for each class is still 0.93. These results show that the resulting model is still able to classify each class well, even though the accuracy obtained is only 82%.

Table 3. Performance Softmax regression with dissimilarity feature training results for all classes

6.png

Figure 6. ROC for one versus rest for each class

The results of the research above are better than those of research on lung sound classification using GLDM, which had the highest accuracy of 73% [9]. This method uses the distribution of signal energy in the time-frequency domain, which is analyzed using the texture analysis method. The spectrogram analysis technique differs from that of Shanthakumari and Priya [25]. Shanthakumari uses morphological features in spectrograms, such as energy, entropy, variance, and waveform length. The weakness of texture analysis in heart sound analysis is the low image contrast because the heart sound energy is relatively low. Apart from that, the frequency of heart sounds is also low. The image produced from a spectrogram is also greatly influenced by spectrogram parameters, such as N-FFT, window width, and sampling frequency [24]. The resulting combination of parameters and characteristics can increase accuracy in subsequent research. Apart from that, the feature subset selection method can be carried out before exploring more advanced classification methods such as deep learning.

Based on all of the tests mentioned above, the GLCM dissimilarity feature generates a machine-learning model that is reasonably good. The current features can effectively identify all classes, including normal and artifacts, but they have poor sensitivity for class murmurs. The similarity in feature data distribution patterns between the normal and murmur classes accounts for this performance. Nevertheless, because there aren't enough datasets, the murmur class isn't very sensitive. This needs to be enhanced since high sensitivity to abnormal classes is necessary for real-world heart sound classification systems to detect them accurately [26]. The research is limited by the balanced dataset for each class, which prevents the classification model's accuracy from being deemed excessively high even when other metrics like precision, recall, AUC, and F1 are deemed to be rather powerful. Further datasets will be required in the future, particularly for the murmur class. A strategy can be carried out by collaborating with healthcare centers to collect datasets, especially for murmur classes. Another way is to use Augmentation Techniques to add murmur class datasets artificially using GAN or other methods [27].