Study of Feature Extraction Method to Detect Myocardial Infraction Using a Phonocardiogram
Ashydiki Malik1,2, Satria Mandala1,2,*, Miftah Pramudyo3
1Human Centric (HUMIC) Engineering, School of Computing, Telkom University, Bandung, Indonesia
2School of Computing, Telkom University, Bandung, Indonesia
3Department of Cardiology and Vascular Medicine Padjadjaran University Bandung, Indonesia Email: 1[email protected], 2,*[email protected],
Correspondence Author Email: [email protected]
Abstract−Myocardial Infraction is one of the most dangerous and often fatal cardiovascular diseases. To detect this disease early, non-invasive methods based on Phonocardiogram (PCG) signals have become a significant focus of research. However, to present, research on feature extraction from PCG signals is still limited. In this research, we propose a study of feature extraction algorithms using Discrete Wavelet Transform (DWT), Mel Frequency Cepstral Coefficients (MFCC), and Entropy methods to detect heart attacks. In the pre-processing stage, we applied noisereduce to remove noise in the PCG signal. Further, we perform feature extraction using DWT, MFCC, and Entropy methods on the processed PCG signal. Following that, we used a detuned KNN with hyperparameters as the classification algorithm to classify the features into two categories: heart attack and non-heart attack. The test results show that DWT, MFCC, and Entropy-based feature extraction methods can make a significant contribution in detecting Myocardial Infraction. In comparison with other feature extraction algorithms, the test results show that the Entropy-based feature extraction method provides the best accuracy of 99%, with 99% sensitivity and 99% specificity. This research makes an important contribution to the development of heart attack detection methods using PCG signals. With promising results, the Entropy-based feature extraction method can be an effective and efficient approach in detecting coronary heart disease early, which in turn can improve patient prognosis and treatment.
Keywords:Myocardial Infraction; Phonocardiogram; Feature Extraction
1. INTRODUCTION
Based on data presented by [1] in 2023, one of the highest causes of death in Indonesia is heart disease, which is 95.68 cases per 100,000 Indonesian population.. Coronary heart disease occurs at any age, from young to old. The Indonesian Ministry of Health states that cardiovascular disease is the number one cause of death globally annually.
Cardiovascular disease is a condition caused by a malfunctioning heart and blood vessels that cannot function normally. An unhealthy lifestyle is one of the triggers for cardiovascular disease. Cardiovascular disease is divided into several types, of which the most common type is Coronary Artery Disease (CAD) or Coronary Heart Disease [2]. The Indonesian Ministry of Health (2017) states that by age 30, heart disease will gradually increase.
Coronary artery disease (CAD) is a heart disease caused by the narrowing of the arteries due to plaque deposition along the inner wall of the coronary arteries, which will interfere with the supply of oxygen and other nutrients in the blood [3]. CAD is one of the main causes of heart attack.
A heart attack is a necrosis that occurs in the myocardium due to a lack of oxygen supply from the coronary arteries, causing blockage of blood flow [4]. Blockages that occur are severe enough to cause the heart to stop beating. Heart attacks are the most common cause of disease in people over 65 years of age [5]. As many as 40%
of people who have Myocardial Infarction die within 1 year after having a heart attack. Myocardial infarction is also caused by the lack of oxygen intake delivered by the blood to the heart muscle [6].
One of the most prevalent and dangerous issues known as cardiovascular disease is myocardial infarction [6]. This condition is caused by inadequate oxygen intake to the heart as a result of coronary artery constriction [3]. According to [7], the onset of this illness is caused by the accumulation of plaque on the artery walls, leading to the constriction of the blood vessels. Currently, some many techniques and methods can be used in diagnosing cardiovascular diseases, such as the use of photoplethysmography (PPG), electrocardiograms (ECG), and phonocardiogram (PCG) signals.
According to [8], [9], the PPG signal is a signal that reads the heart's blood volume. Then ECG is a signal that captures the electrical activity of the heart [10]. Meanwhile, PCG is a signal obtained from heart sounds [11].
Some of these signals have limitations and strengths in detecting heart signals. In particular, PCG signals have the strength of being non-invasive, cost-effective, and simple [12].
Phonocardiogram is a cardiac acoustic recording technology that is frequently used to assess for heart disease, according to [3]. A Phonocardiogram is a technique for electronically capturing the acoustic vibrations of the heart as signals. According to [13], the system works by capturing the outcomes of blood turbulence reflections that occur when the heart valves seal.
A dynamic microphone array is used in this approach. PCG signals are preferred for identifying cardiac valves and arterial stenosis. The cardiac cycle is made up of four signals, two of which are audible (S1; S2) and two of which are low amplitude and difficult to hear (S3; S4).
At this time, research on Myocardial infarction detection using PCG signals is limited. Some studies focus more on classification studies than feature extraction from PCG signal-based Myocardial infarction, one of which
is as done [6] to detect myocardial infractions using the KNN subspace ensemble model with MCFF feature extraction. This model produces an accuracy of 94.9%. This study only focuses on classification, while the extraction of myocardial infraction features is not discussed in the study.
M. Zarrabi et al. [13] examine a classification system that predicts heart attacks with ECG, PCG, and Clinical Feature signals. The feature extraction used in this study is discrete wavelet transform (DWT). The test results get an accuracy rate of 99% with sensitivity and specificity of 98% and 100% respectively.
Sotaquira et al. experimented with a heart rate classifier that can discriminate between normal and pathological using DNN and weighted probability comparisons the following year [9]. This study employs many feature extraction techniques, including time domain, frequency domain, and time-frequency domain.
This experiment obtained a sensitivity rate of 91.3 percent and a specificity rate of 93.8 percent. Samanta et al. [2] investigated a multi-channel system for detecting CAD using PCG signals. Time domains and frequency domains are two of the feature extractions employed in this investigation. The suggested technique obtains an accuracy of 82.57%, whereas the standard CAD identification system achieves 68.93% utilizing single-channel data.
In 2019, H. Li et al. [14] also conducted research on CAD detection with dual-input neural networks and deep learning features using ECG and PCG signals. The research used ECG signals using the time domain, frequency domain, and time-frequency domain feature extraction, while PCG signals used the time domain, frequency domain, energy domain, entropy domain, and kurtosis domain feature extraction. The results of this study obtained a high accuracy value of 95.62% by combining the two ECG and PCG signals.
Pathak et al. [15]} conducted research on PCG-based CAD detection that can tolerate noise. the resulting system gets maximum value in sensitivity accuracy and specificity for systems that can withstand noisy environments. Mandala et al. [16] investigated CAD identification methods using machine learning and evaluated them using three kernels. With six characteristics, the discrete wavelet transform (DWT) is employed for feature extraction.
The investigation was carried out at four separate places and yielded a reasonably accurate result of 66%.
Ali et al. [17] investigated machine learning techniques for artificial intelligence-based heart disease diagnosis devices. In this study, two feature extractions are used: Framingham risk factor extraction and feature fusion layer.
The accuracy percentage of the testing findings over the training data is roughly 83%. In the same year, Jayasree and Rao [18] reviewed the fundamental concepts of coronary artery disease (CAD), angiography, and literature on angiography approaches used to identify and predicted Coronary Artery Disease using deep learning techniques.
In 2020, researchers H. Li et al. [3] experimented with the fusion framework by combining the application of deep learning and artificial features for CAD detection. The feature extractions used in this study are time, frequency, time-frequency, energy, statistics, entropy, and MFCC. The results of this experiment have a high accuracy rate of 90.43% with sensitivity and specificity of 93.67% and 83.36% respectively. In the following year, P. Li, Y. Hu, and Z. P. Liu [10] conducted an experiment on classification in detecting CAD using MFCC and KNN and using SVM for CAD categorization. The results of the experiment have a high accuracy rate for CAD types, namely 88.0%, 89.2%, 91.1%, and 85.3% for normal, DVCAD, SVCAD, and TVCAD.
Arslan and Karhan [19] used a 5-fold cross-validation framework and 10-fold cross-validation Data Analysis Protocol (DAP) to do a research on the multi-classification of five PCG classes (healthy, aortic stenosis, mitral stenosis, mitral regurgitation, and mitral valve prolapse). MFCC is the feature extraction method employed in this research.
The DNN model showed the best classification performance, with 98.9% precision, 98.7% recall, 98.8%
F1 score, and 98.9% accuracy using 5-fold cross validation, and a correlation value of 0.981 using the DAP technique. Chang et al. [20] also undertook research on artificial intelligence-based cardiac disease diagnosis devices utilizing machine learning methods. A random forest classifier with an accuracy of 83% on training data is used in this study.
Based on previous research related to detecting the presence of myocardial infarction, research that focuses on feature extraction is still hard to find. Therefore, this research will focus on the development of feature extraction in improving performance in heart attack detection using PCG signals. This research aims to conduct a study of feature extraction methods for detecting myocardial infractions based on PCG signals and develop a classification system in the process of comparing three feature extraction methods.
2. RESEARCH METHODOLOGY
2.1 Research Stages
As shown in Figure 1, this research system's design is divided into numerous phases. A classification system that will be judged on its effectiveness comes after various stages of dataset retrieval.
Figure 1. Flowchart System
The system as shown in Figure 1, that was developed in this study begins with the input of PCG data collected from Hasan Sadikin Hospital, Bandung. The data used in this study is data in the form of phonocardiogram signals for 30 seconds from 140 patients. PCG data will be managed at the preprocessing stage to remove noise on the PCG signal with the "noisereduce" method. Feature extraction to extract features on the PCG signal to improve the accuracy of MI (Myocardial infarction) detection and get the results of MI detection accuracy in classification with the KNN model. The last step is to examine the performance results of several feature extraction methods to determine the best performance.
2.2 Material Data
This study used data from Hasan Sadikin Hospital, Bandung. The data was in the form of phonocardiogram signals for 30 seconds from 140 patients. each patient has recorded four recordings at different heart locations namely Apex, Left Upper Sternal Border (LUB), Left Lower Sternal Border (LLSB), and Right Upper Sternal Border (RUSB).
1) Normal Signal: There were 70 healthy subjects, with a total of 280 records at each different point for all subjects. As presented in the following Figure 2, illustrates the raw data of the normal signal at each different collection point.
Figure 2. PCG Signal of Normal Subject
2) Myocardial Infraction Signal: There were 70 subjects with Myocardial infarction, with a total of 280 records taken at each different point similar to the normal signal. The signals of myocardial infarction subjects can be seen as shown at each different point as shown in Figure 3.
Figure 3. PCG signal of Subject MI
2.3 Test Metrics
The test metrics used in testing the algorithm are metrics that are also used in previous studies [6], [13], [16].
Includes accuracy and specifications.
1) Accuracy
𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝑇𝑁
𝑇𝑃+𝐹𝑃+𝐹𝑁+𝑇𝑁 (1)
2) Spesificity 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑇𝑁
𝑇𝑁+𝐹𝑃 (2)
3) Sensitivity 𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑃
𝑇𝑃+𝐹𝑁 (3)
There are several formulas used in the heart attack classification stage. Formula 1 is used to calculate accuracy, Formula 2 is used to calculate specificity, and Formula 3 is used to calculate sensitivity. In these formulas, there are variables such as TP (True-Positive) which represents the number of correctly classified heart attack signals, TN (True-Negative) which represents the number of correctly classified normal signals, FP (False- Positive) which represents the number of incorrectly classified heart attack signals, and FN (False-Negative) which represents the number of incorrectly classified normal signals [2].
2.4 Preprocessing
In this study, the first step performed is preprocessing which includes the process of removing noise on all signals.
The method used to reduce noise is noisereduce. Noisereduce was developed by [21] which is a Python library useful for denoising. Several previous studies in the field of audio signals have used Noisereduce. For example, [22] used noisereduce to clean noise in spontaneous speech audio recordings by achieving a high accuracy of 92.72%.
2.5 Feature Extraction
Feature extraction used in this study for PCG signals that have been removed noises using 3 methods, which are:
DWT, MFCC, and Shannon Entropy.
1) Discreate Wavelet Transform
According to [13] Discreate Wavelet transforms (WT) is a technique for presenting effective information on signals. Discreate Wavelet Transform (DWT) is generally the process of combining low-pass and high-pass which will produce approximation coefficients and detail coefficients from the signal. The DWT has the ability to describe high-frequency energy changes in a diastolic murmur [16]. The formula used by [23] approximation coefficient and detail coefficient as follows:
𝐴 = 𝑐𝐴𝑛∑𝑛𝑖=0𝑐𝐷𝑛 (4)
In Formula 4, Symbol A depicts the wavelet coefficient values while symbol n indicates the number of approximate levels present.
2) MFCC
Based on [24] MFCC (Mel Frequency Cepstral Coefficients) is one of the feature extraction methods commonly used for speech recognition in recent years the use of MFCC is also extensive in signal processing.
According to [3] MFCC has the ability to present time-frequency information that is widely used in speech recognition, which involves three extraction steps. The MFCC technique involves the operations of Discrete Fourier Transform (DFT), logarithm, scale mel, and Discrete Cosine Transform (DCT) [19]. The following steps in MFCC as conducted [24]:
a) Windowing
At this stage, the signal is divided into frames with a duration of about 25 ms. After that, the hamming window method is used to reduce the effect of discontinuity. The calculation uses the formula stated in Formula 5 below:
𝑀𝑛= 0,54 − 0,46 (2𝜋(𝑛−1)
𝑁−1 ) (5)
In Formula 5, Mn represents the value of the number of samples.
b) Filterbank
In this step, a conversion is performed to obtain the desired non-linear frequency. The filterbank calculation can be done with the following formula:
𝑓𝑚𝑒𝑙= 𝑙𝑜𝑔10(2595) (1 + 𝑓
700) (6)
In Formula 6, f indicates the frequency value of the signal.
c) Discrete consine transform
At this stage, signal compression will be carried out using the DCT algorithm, this stage is optional.
3) Shannon Entropy
According to [25], Shannon Entropy is the calculation of the average amount of information in a signal.
Formula 7 used in the calculation of entropy on the signal is as follows:
𝑆 = − ∑𝑚𝑖=1|𝑋|2𝑙𝑜𝑔|𝑋|2 (7)
2.6 Scenario of Experiments
In this research, testing is needed to determine the success of the feature extraction algorithm that was built and test what features have a major contribution to the heart attack detection stage. The following are the feature extraction methods used:
Table 1. Feature Extractions No Feature Extraction
1 Discreate Wavelet Transform
2 MFCC
3 Shannon Entropy
Using the three feature extractions mentioned above, this research will test the MI detection method by comparing three alternative feature extraction algorithms, namely DWT, MFCC, and SE, before performing the classification process.
3. RESULT AND DISCUSSION
3.1 Preprocessing Result
In this study, the preprocessing stage will use the noisereduce method to reduce noise. In our illustration, there are Figure 4 and Figure 5 which are examples of signals that have undergone the denoising process, both in normal conditions and PCG signals. In this study, the preprocessing stage will use the noisereduce method to reduce noise.
In our illustration, there are Figure 4 and Figure 5 which are examples of signals that have undergone the denoising process, both in normal conditions and PCG signals.
Figure 4. Normal signal that has cleared
Figure 5. MI signal that has cleared
3.2 Feature Extraction Results
Phonocardiogram (PCG) signals that have gone through the preprocessing stage using noisereduce will be carried out at the feature extraction stage. The feature extraction algorithms used in this study are 3 different algorithms, namely DWT, MFCC, and Shannon entropy. The proposed algorithm is intended to be able to identify the characteristics of Myocardial Infarction and be able to detect Myocardial infarction. The following are the results of each feature extraction algorithm
1) Discreate Wavelet Transform
In this algorithm, several features of the PCG signal are obtained, namely maximum value, minimum value, median, standard deviation, skewness, variance, kurtosis, quartile 1, quartile 3, average, and inter quartile range (IQR) as in Table 2.
Table 2. Wavelet Feature Extraction Results
Feature Value
Normal MI
Mean wavelet 1.502731e-06 1.435873e-09 Std wavelet 0.000743 0.000511 Max wavelet 0.010726 0.010863 Min wavelet -0.007889 -0.008997 Med wavelet 2.350813e-09 5.162848e-10
Var wavelet 5.527448e-07 2.614612e-07 Skew wavelet 2.231036 0.093306
Q1 wavelet -1.797313e-06 -2.598059e-06 Q3 wavelet 1.679290e-06 2.369187e-06 IQR wavelet 3.476603e-06 4.967246e-06 MinMax wavelet 0.018615 0.019861
Kurt wavelet 56.758168 58.236027 2) MFCC
In this algorithm, several characteristics of the PCG signal are obtained, namely the maximum value, median, quartile 1, quartile 3, mean, variance, skewness, standard deviation, inter quertile range (IQR), the range between the minimum and maximum values, and kurtotis as in Table 3.
Table 3. MFCC Feature Extraction Result
Feature Value
Normal MI
Mean mfcc -40.691204 -42.735283 Std mfcc 218.82877 224.17154 Max mfcc 171.80734 185.06921 Min mfcc -1019.07355 -1040.87920 Med mfcc 2.255058 1.231523
Var mfcc 47886.027 50252.880 Skew mfcc -4.087075 -4.088933 Q1 mfcc 0.000000 -0.938724 Q3 mfcc 13.019967 8.764449 IQR mfcc 13.019967 9.703173 MinMax mfcc 1190.8809 1225.9484
Kurt mfcc 14.839055 14.851157 3) Shannon Entropy
In this algorithm, get the entropy value on normal and MI data as in Table 4.
Table 4. Shannon Entropy Extraction Result
Feature Value
Normal MI
Entropy 13.399985 13.400227 3.3 Experiments Results
The above-mentioned feature extraction algorithms will be put through performance testing by comparing the three various feature extraction algorithms, DWT, MFCC, and SE, before executing the classification process with the tuned KNN.
Table 5. Experiment Results
Scenario Accuracy Sensitivity Specivity
DWT 97% 95% 98%
Scenario Accuracy Sensitivity Specivity
MFCC 98% 98% 98%
SE 99% 99% 99%
In Table 5, it is found that Shannon entropy has the highest performance compared to WT and MFCC.
3.4 Discussion
Based on the analysis results in the previous scenario, this study analyzes the performance of various algorithms separately to detect myocardial infarction using Phonocardiogram (PCG) signals. The test results show that Shannon Entropy (SE) characteristics have a significant influence on the accuracy of myocardial infarction detection.
In three situations studied using classical machine learning and K-Nearest Neighbors (KNN) classifier models, SE features excelled in terms of accuracy, sensitivity, and specificity. The SE feature achieved the best performance in the test, with accuracy reaching 99%, sensitivity of 99%, and specificity of 99%. This shows that the SE feature has an excellent ability to detect myocardial infarction based on PCG signals.
Furthermore, the Mel Frequency Cepstral Coefficients (MFCC) feature also performs well with 98%
accuracy, 98% sensitivity, and 98% specificity in detecting myocardial infarction. Meanwhile, the Discrete Wavelet Transform (DWT) feature has 97% accuracy, 95% sensitivity, and 98% specificity, which although slightly lower than the SE and MFCC features, still gives good results.
Table 6. Comparative research results
Work Signal Feature Extraction Accuracy
Khan et al. [6] PCG MFCC 94,9%
M. Zarrabi et al [12] ECG + PCG DWT 99 %
Our Work PCG
DWT 97%
MFCC 98%
SE 99%
The results shown in Table 6 support the use of Shannon Entropy in the feature extraction process in myocardial infarction detection using PCG signals. The SE feature has the best test accuracy of the three algorithms studied. This can be explained by the fact that the SE feature is based on the distance between consecutive signal peaks in the PCG signal. In addition, the entropy in the SE characteristic was also found to be an important factor in the classification process.
Although Mel Frequency Cepstral Coefficients (MFCC) and Discrete Wavelet Transform (DWT) also contain some characteristics that can provide information about the PCG signal, the test results showed that they did not have a significant influence on the high level of accuracy found in this study.
Overall, this work suggests that combining Shannon Entropy (SE) extraction features with machine learning approaches such as KNN has the potential to be an effective strategy for diagnosing myocardial infarction using PCG data. The findings show that myocardial infarction diagnosis systems based on SE characteristics have a high degree of accuracy.
It should be mentioned, however, that this study has several limitations. To begin, this study just employs one classifier model (KNN), however the results may be expanded by employing different models. Furthermore, to guarantee the validity and reliability of the suggested method, this research may be expanded through using a bigger dataset and a broader spectrum of patient situations.
More work on developing and upgrading these approaches is intended to make significant progress in the early diagnosis of myocardial infarction, which would enhance the prognosis and treatment of heart attack patients.
4. CONCLUSION
This research aims to detect heart attack or myocardial infarction using DWT (Discrete Wavelet Transform), MFCC (Mel Frequency Cepstral Coefficients), and Entropy feature extraction methods on Phonocardiogram (PCG) signals. The results show that the use of a combination of DWT, MFCC, and Entropy feature extraction methods can effectively detect heart attacks. The DWT method is used to obtain information about amplitude and frequency changes in PCG signals. DWT can reveal patterns and structures in the signal related to heart attacks.
Furthermore, the MFCC method is used to analyze the spectral characteristics of the PCG signal. These features can help identify typical patterns associated with myocardial infarction. The Entropy method, on the other hand, is used to measure the complexity of the PCG signal and distinguish between normal patterns and abnormal patterns that may indicate a heart attack. Test results show that the model using a combination of DWT, MFCC, and Entropy feature extraction provides high accuracy in detecting myocardial infarction. This accuracy is crucial in the diagnosis and appropriate treatment measures in the early stages of a heart attack. In addition, this approach also provides accurate information to doctors for effective treatment planning. This study makes an important contribution to the development of myocardial infarction detection methods using PCG signals. The promising
results show that DWT, MFCC, and Entropy feature extraction methods have the potential to be an effective and efficient approach to detecting heart attacks early. This could have a positive impact on patient prognosis and enable faster and more timely medical intervention. Nonetheless, it should be noted that this study still requires further testing and validation using a larger data set. In addition, future studies could consider using other extraction features and analyze their effect on the overall myocardial infarction detection performance. With further efforts in developing and improving this method, it is hoped that greater progress can be made in the early detection of heart attacks, which will ultimately improve the quality of care and patient safety.
ACKNOWLEDGMENT
We thank Hasan Sadikin Hospital for granting permission and giving the PCG data that were used in this search.
Also, we appreciate the financial support from our educational institutions as well as the moral and financial support from our parents. Their support was essential to the effective conclusion of this investigation. We also value the ideas and contributions made by the relevant experts, researchers, and reviewers, as well as their insightful criticism. Last but not least, we want to give our gratitude to our family and friends for their encouragement and moral support. Even though not all parties can be specifically acknowledged, we want to express our gratitude and appreciation to everyone who helped make this study a success. Thank you once again for your support, and we hope that this study provides significant benefits in the field of myocardial infarction detection and treatment.
REFERENCES
[1] E. F. Santika, “10 Penyakit Penyebab Kematian Tertinggi di Indonesia,” Databoks, 2023.
[2] P. Samanta, A. Pathak, K. Mandana, and G. Saha, “Classification of coronary artery diseased and normal subjects using multi-channel phonocardiogram signal,” Biocybern Biomed Eng, vol. 39, no. 2, pp. 426–443, Apr. 2019, doi:
10.1016/j.bbe.2019.02.003.
[3] H. Li et al., “A fusion framework based on multi-domain features and deep learning features of phonocardiogram for coronary artery disease detection,” Comput Biol Med, vol. 120, May 2020, doi: 10.1016/j.compbiomed.2020.103733.
[4] C. Gopalan and E. Kirk, Biology of Cardiovascular and Metabolic Diseases. Academic Press, 2022.
[5] J. L. Rodgers et al., “Cardiovascular Risks Associated with Gender and Aging,” J Cardiovasc Dev Dis., vol. 6, no. 2, 2019.
[6] U. M. Khan, Z. Mushtaq, M. Shakeel, S. Aziz, and S. Z. H. Naqvi, “Classification of Myocardial Infarction using MFCC and Ensemble Subspace KNN,” in Proc. of the 2nd International Conference on Electrical, Communication and Computer Engineering (ICECCE), Istanbul, 2020.
[7] S. Parmet, T. J. Glass, and R. M. Glass, “Coronary artery disease,” JAMA, vol. 292, no. 20, p. 2540, 2004.
[8] M. F. Ihsan, S. Mandala, and M. Pramudyo, “Study of Feature Extraction Algorithms on Photoplethysmography (PPG) Signals to Detect Coronary Heart Disease,” in 2022 International Conference on Data Science and Its Applications (ICoDSA), 2022.
[9] M. Sotaquirá, D. Alvear, and M. Mondragón, “Phonocardiogram classification using deep neural networks and weighted probability comparisons,” J Med Eng Technol, vol. 42, no. 7, pp. 510–517, Oct. 2018, doi:
10.1080/03091902.2019.1576789.
[10] P. Li, Y. Hu, and Z. P. Liu, “Prediction of cardiovascular diseases by integrating multi-modal features with machine learning methods,” Biomed Signal Process Control, vol. 66, Apr. 2021, doi: 10.1016/j.bspc.2021.102474.
[11] W. R. Putra, S. Mandala, and M. Pramudyo, “Study of Feature Extraction Methods to Detect Valvular Heart Disease (VHD) Using a Phonocardiogram,” in 2021 International Conference on Intelligent Cybernetics Technology &
Applications (ICICyTA), 2022. [Online]. Available: http://orcid.org/0000-0001-6997-5875
[12] C. Ahlström, “Nonlinear phonocardiographic Signal Processing,” Institute of Technology Linkoping University, Linkoping, 2008.
[13] M. Zarrabi et al., “A system for accurately predicting the risk of myocardial infarction using PCG, ECG and clinical features,” Biomed Eng (Singapore), vol. 29, no. 3, Jun. 2017, doi: 10.4015/S1016237217500235.
[14] H. Li et al., “Dual-input neural network integrating feature extraction and deep learning for coronary artery disease detection using electrocardiogram and phonocardiogram,” IEEE Access, vol. 7, pp. 146457–146469, 2019, doi:
10.1109/ACCESS.2019.2943197.
[15] A. Pathak, P. Samanta, K. Mandana, and G. Saha, “An improved method to detect coronary artery disease using phonocardiogram signals in noisy environment,” Applied Acoustics, vol. 164, Jul. 2020, doi:
10.1016/j.apacoust.2020.107242.
[16] S. Mandala, M. Pramudyo, A. Rizal, and M. Fikry, “Study of Machine Learning Algorithm on Phonocardiogram Signals for Detecting of Coronary Artery Disease,” Ind. Journal on Computing, vol. 5, no. 3, 2020, doi:
10.34818/indojc.2021.5.3.536.
[17] F. Ali et al., “A smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and feature fusion,” Information Fusion, vol. 63, pp. 208–222, Nov. 2020, doi: 10.1016/j.inffus.2020.06.008.
[18] M. JayaSree and L. K. Rao, “Survey on-identification of coronary artery disease using deep learning,” Mater Today Proc, 2020.
[19] Ö. Arslan and M. Karhan, “Effect of Hilbert-Huang transform on classification of PCG signals using machine learning,”
Journal of King Saud University - Computer and Information Sciences, vol. 34, no. 10, pp. 9915–9925, Nov. 2022, doi:
10.1016/j.jksuci.2021.12.019.
[20] V. Chang, V. R. Bhavani, A. Q. Xu, and M. A. Hossain, “An artificial intelligence model for heart disease detection using machine learning algorithms,” Healthcare Analytics, vol. 2, Nov. 2022, doi: 10.1016/j.health.2022.100016.
[21] T. Sainburg, M. Thielk, and T. Q. Gentner, “Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires,” PLoS Comput Bio, vol. 16, no. 10, 2020.
[22] L. Ziming, L. Proctor, P. Collier, D. Casenhiser, E. J. Paek, and S. O. Yoon, “Machine Learning of Transcripts and Audio Recordings of Spontaneous Speech for Diagnosis of Alzheimer’s Disease,” in Alzheimer’s Association International Conference, 2021.
[23] S. K. Ghosh, R. K. Tripathy, and R. N. Ponnalagu, “Evaluation of performance metrics and denoising of PCG signal using Wavelet Based Decomposition,” in 2020 IEEE 17th India Council International Conference (INDICON), 2020, pp. 1–6.
[24] O. El Badlaoui and A. Hammouch, “Phonocardiogram classification based on MFCC extraction,” in 2017 IEEE International Conference on Computational Intelligence and Virtual Environments for Measurement Systems and Applications (CIVEMSA), 2017, pp. 217–221.
[25] F. Khan, M. A. Jan, and M. Alam, Applications of Intelligent Technologies in Healthcare. Springer, 2019.
