Heart Signal Analysis Using Multistage Classification Denoising Model

,


Introduction
Cardiac arrest has become a common disease in today's world.According to the World Health Organization (WHO), cardiovascular diseases (CVDs) are responsible for 17.9 million deaths per year, making them the leading cause of mortality.Heart attacks and strokes contribute to over 80% of fatalities caused by CVDs, with one-third of these deaths occurring prematurely in individuals under 70 years.It is crucial to identify and treat cardiovascular diseases promptly to improve patients' quality of life and well-being.Tis can also reduce the frequency of occurrences and prevent the onset of various complications [1].A phonocardiogram (PCG) signal is a graphical representation of the sounds made by the opening and closing of heart valves.Te signal shows how the heart sounds change over time.By monitoring these signals, doctors can detect heart abnormalities early on, which can help to reduce mortality rates [1].
Researchers have used time-frequency domain characteristics to train conventional machine-learning models for heart signal categorization.In [2], a similar technique is proposed that utilizes spectrogram phase and magnitude features to identify heart valve issues.Te authors in [3] presents a TQWT-based two-stage abnormality detection approach, where SVM is used in the frst stage for binary classifcation and the KNN is used in the second stage for further classifcation.In addition, Shannon energy [4], cochleagram features [5], and mel-frequency cepstral coefcients (MFCC) [6] are also considered as other characteristics to detect abnormalities in signals.To detect abnormalities in signals, Karhade et al. [7] present the time-frequency domain deep learning (TFDDL) framework techniques, which combine CNN architecture with timefrequency domain polynomial chirplet transform.A novel approach combining the wavelet scattering transform with the twin support vector machine (TSVM) was employed in the classifcation process [8].
Te Yaseen Khan dataset [6], Physionet Challenge 2016 [9], fetal heart sound [10], and PASCAL heart sound datasets are available for the detection of abnormality in the PCG heart sound signals.
1.1.Contribution of the Paper.Te primary contribution of the paper lies in its introduction of a multistage model that combines CNN and LSTM networks for heart sound signal classifcation.Tis innovative combination enables the extraction of intricate quasi-cyclic features from the heart sound signal, ultimately leading to more efective signal classifcation.
1.2.Organization of the Paper.Tis paper explains in detail the proposed model and its evaluation process.In Section 2, the step-by-step procedure of the proposed model is described, which includes signal preprocessing, decomposition, and classifcation using a novel hybrid deep-learning approach of the CNN and LSTM.Section 3 presents the outcomes of our experiments, demonstrating the performance of individual components such as the CNN, LSTM, and the hybrid model in classifying heart sound signals.Section 4 includes a comparative analysis with recently available methods, taking into account the model's performance under noisy conditions.Finally, Section 5 concludes the study comprehensively.

Literature Review
Various models of machine learning and deep learning (DL) have been proposed for detecting heart sound abnormalities.In recent times, the deep learning neural network (DNN) has become a powerful tool for identifying abnormal heart sounds due to its strong feature representation capability.Tomae and Dominik [11] created an end-to-end deep neural network that focuses on temporal or frequency features to extract hidden features in the temporal domain.Ryu et al. [12] developed a CNN model specifcally designed for segmented PCG classifcation.Recently, Humayun et al. [13] created a time-convolutional (tConv) unit to discover hidden features from temporal properties.Chakraborty et al. [14] used the cardiac cycle spectrum for training on a 2D-CNN.Figure 1 shows the evolution of abnormality detection methods.
Mel-frequency cepstral coefcients (MFCC) and discrete wavelet transform (DWT) characteristics were introduced by Yaseen et al. [6] and combined with support vector machine (SVM) for automated identifcation of heart valve diseases and other conditions.In the meanwhile, Zabhi et al. [15] classifed HVDs using an ensemble of 20 feedforward neural networks (FFNNs) with time, frequency, and timefrequency characteristics.Nevertheless, approaches that depend on morphological characteristics need the identifcation of S1 and S2 sound peaks [15,16].Te phonocardiogram (PCG) signal contains pathological fuctuations and noise interference, which make it difcult to detect these peaks [17,18].An integrated method is presented with k-NNACO algorithm.According to Rajathi and Radhamani [19], an analysis of accuracy and error rate is used to determine its efectiveness.An empirical Sstudy on initial proposed algorithms is proposed by Meena et al. [20] on the classifcation of heart disease dataset-its prediction and mining.All available methods need to be further improved in several areas.To cut computational costs, the detection algorithm's efciency must be improved.Also, false alarms should be reduced.Tis is especially important because the user will be interacting with the system; therefore, a low-cost solution is required.Table 1 shows details about similar work done by other researchers on diferent datasets.

Proposed Model
Te proposed model classifes signals into the initial stage at the user end and a later stage in the clinical end as shown in Figure 2.
Te suggested method uses TQWT transforms to help a convolutional neural network (CNN) categorize heart sounds into fve types.Te suggested approach is divided into three stages, preprocessing, denoising, and classifcation.Preprocessing is done on the signals in the frst stage to make sure they are of the same length and normalized.After that, these signals are broken down into six levels: one approximation level coefcient and fve detailed levels.Once this decomposition is complete, the output is combined into a vector that is fed into a one-dimensional CNN model.Lastly, a training session and subsequent dataset validation are performed on the hybrid model.

Preprocessing.
Te dataset for this study encompasses fve classes of distinct heart sound signals, each sampled at a frequency of 8 kHz.Te following preprocessing procedures are applied to each signal.

Resampling.
To address the frequency range constraint of the fast Fourier transform (FFT), which cannot capture pathological sounds below 500 Hz, the signal sampling frequency is downsampled from 8 kHz to 1 kHz [36].

Scaling.
Normalization is performed to overcome the efect of interclass variation on the amplitude of signals that suppress in amplitude variation in the signal.Normalization of signals is performed as follows: . (1) 3.1.3.Resizing.For a given dataset, the duration of the cardiac signal ranges anywhere from 1.15 seconds to 3.99 seconds.Each sample is comprised of three cardiac 2 Journal of Electrical and Computer Engineering cycles, and the duration of these signals might vary owing to diferences in the rate at which the heart beats.After identifying the beginning and end points of the transmission, the signals are shrunk down to the same size, which is 2.8 kilobits.Matlab's "imresize" function, which makes use of bicubic interpolation, is called upon when the sample's size has to be changed.Te change in proportions, as seen in Figure 3, are seen at the original signal.

TQWT Denoising.
In the Mallat method of sub-band coding algorithm, the signal is convolved with two frequency bands, high pass (G) and low pass frequencies (H) [37].In this technique, the signal is divided into a detailed level (D) and approximation (A) coefcients, respectively.At the frst level, the signal x(n) itself will be convoluted with a highpass and low-pass flter.Downsampling approximation applied on the low-pass flter and get next level coefcients, as shown in Figure 4. Te detailed and approximation features at a particular level (j) is obtained as given in equations ( 2) and (3): and With the proposed technique, the heart sound signal is decomposed up to 18 levels, and a rough level coefcient vector is obtained.A tunable Q-factor qavelet transform (TQWT-) based adaptive thresholding technique [38] is used to suppress the in-band noise.Te cardiac sound signal is decomposed by TQWT using eighteen TQWT decompositions.It is described as a time-condensed shortwave signal that transports energy and information.Both low-frequency and high-frequency sounds are categorized as being present.High-frequency transmissions have quality, while low-frequency signals have information.A learning rate of 0.01% is applied in each iteration, totaling 450 iterations.Gradient descent optimization is utilized to determine the optimal settings for the output of the LSTM layer including the dropout layer, which is useful in noise-flled environments and plays a critical role in preventing overftting.Table 2 illustrates the architecture of the proposed model.

Results and Discussion
For the evaluation of the proposed approach, a publically available dataset [6]

Journal of Electrical and Computer Engineering
It is also possible to see how many times each input class was correctly matched with the output class by computing the confusion matrix.Te confusion matrix produced by the DWT denoising method is shown in Figure 6.

Proposed Method Results and Comparison to the Most
Cutting-Edge Research Techniques.Using the same dataset, the suggested approach is evaluated in comparison to other methods that have just recently been presented for the detection of heart sound signals.Te procedures are outlined in the following.Yaseen et al. [6] have been successful in extracting features based on MFCC and DWT.Create and analyze the performance using the SVM, KNN, and deep neural network methodologies.Using the WSST approach and random forest for the diferent categories, Ghosh et al.'s [2] method acquired the time-frequency domain signal, and they calculated the magnitude and phase characteristics.Tese characteristics come from the transmission.Te approach, which is based on deep learning and employs the deep wavelet method, was proposed by Oh et al. [39] and colleagues.Wavelets are a kind of generative model that has been investigated for its potential to produce raw audio signals.Te experiments are performed with the proposed hybrid model, as well as for the CNN model and LSTM model individually.Te confusion matrix obtained using these models are provided in Figure 7. Te CNN model achieved 97.8% accuracy, the LSTM model achieved 43.9% accuracy, and the hybrid model achieved 98.9% accuracy.It shows that the hybridization of CNN and LSTM networks produces better results than the individual models.It is expected because the hybrid of CNN and LSTM networks helps extract relevant patterns and exploit their time dependency.Te CNN model alone produces satisfactory results.However, the LSTM model's performance degrades drastically.Tese results indicate that the hybridization of the diferent models.
Table 3 depicts the performance parameters obtained for the CNN, LSTM, and hybrid models.Te results obtained using the hybrid models are superior for all fve categories compared to the CNN and LSTM models.Te hybrid model achieved 100% F1-score for the classes N and MVP, while 98.04% for AS, 98.88% for MS, and 99.19% for MR.Tese results demonstrate the efcacy of the proposed hybrid model classifying all fve heart sound signal categories, specifcally normal vs pathological cases.Such a system will be helpful for automatically analyzing the heart sound signal.

Comparison with Existing Methods.
Te efectiveness of the proposed model is evaluated by comparing its performance with several recently introduced methods documented in the literature for the same dataset, as illustrated in Table 4. Te proposed hybrid model with 98.9% accuracy is superior to all the compared methods.While some existing methods also show prominent results, the superior accuracy of the proposed model highlights its efectiveness in classifying heart sound signals.Its accuracy may reduce misdiagnoses and improve patient care, benefting individuals with heart valve diseases and other cardiac disorders.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon reasonable request.
Model.Te TQWT denoising output generates approximation level coefcients using TQWT, organized in a 1-D array of length 2942, serving as input for training the CNN model.Te model architecture consists of 5 layers: 1 input layer, 2 CP (convolution and pooling) layers, 1 fully connected layer, and 1 output softmax layer.Each CP layer employs padding to maintain output size matching the input.Following TQWT decomposition, 2942 coefcients are generated for each of the neurons in the layers.Each training epoch comprises 50 iterations, and within each epoch, nine iterations are executed.

Figure 1 :
Figure 1: Time series of diferent techniques.

Table 1 :
Literature review for the classifcation of PCG signal.

Table 2 :
Architecture of each layer used in the proposed multistage model.

Table 4 :
Comparison with existing methods.