HBNET: A blended ensemble model for the detection of cardiovascular anomalies using phonocardiogram

Abstract

BACKGROUND:

Cardiac diseases are highly detrimental illnesses, responsible for approximately 32% of global mortality [1]. Early diagnosis and prompt treatment can reduce deaths caused by cardiac diseases. In paediatric patients, it is challenging for paediatricians to identify functional murmurs and pathological murmurs from heart sounds.

OBJECTIVE:

The study intends to develop a novel blended ensemble model using hybrid deep learning models and softmax regression to classify adult, and paediatric heart sounds into five distinct classes, distinguishing itself as a groundbreaking work in this domain. Furthermore, the research aims to create a comprehensive 5-class paediatric phonocardiogram (PCG) dataset. The dataset includes two critical pathological classes, namely atrial septal defects and ventricular septal defects, along with functional murmurs, pathological and normal heart sounds.

METHODS:

The work proposes a blended ensemble model (HbNet-Heartbeat Network) comprising two hybrid models, CNN-BiLSTM and CNN-LSTM, as base models and Softmax regression as meta-learner. HbNet leverages the strengths of base models and improves the overall PCG classification accuracy. Mel Frequency Cepstral Coefficients (MFCC) capture the crucial audio signal characteristics relevant to the classification. The amalgamation of these two deep learning structures enhances the precision and reliability of PCG classification, leading to improved diagnostic results.

RESULTS:

The HbNet model exhibited excellent results with an average accuracy of 99.72% and sensitivity of 99.3% on an adult dataset, surpassing all the existing state-of-the-art works. The researchers have validated the reliability of the HbNet model by testing it on a real-time paediatric dataset. The paediatric model’s accuracy is 86.5%. HbNet detected functional murmur with 100% precision.

CONCLUSION:

The results indicate that the HbNet model exhibits a high level of efficacy in the early detection of cardiac disorders. Results also imply that HbNet has the potential to serve as a valuable tool for the development of decision-support systems that aid medical practitioners in confirming their diagnoses. This method makes it easier for medical professionals to diagnose and initiate prompt treatment while performing preliminary auscultation and reduces unnecessary echocardiograms.

Keywords

Blended ensemble mel frequency cepstral coefficient meta-learner phonocardiogram softmax regression

1. Introduction

Valvular heart disease (VHD) is a predominant cause of death and truncated life. Due to the elevated prevalence and mortality rate of VHD, it remains a trending research subject. The number of cardiac patients is increasing yearly, covering almost 41 million globally [2]. One of the ways to limit the increasing number of cardiac cases is through the timely detection and treatment of the disease. Heart auscultation is the most straightforward and non-invasive method for initially diagnosing heart disease. However, the doctor’s experience and expertise determine whether the diagnosis is accurate. The hearing capability is not the same among individuals, and we cannot fully rely on cardiac auscultation for the diagnosis. The gold standard used for disease detection is an echocardiogram. Echocardiography is an expensive, lengthy diagnosis procedure, and there is a waiting period to get an appointment, which delays the diagnosis of the disease. Another valuable alternative is phonocardiography, a non-invasive method for obtaining recordings of cardiovascular sounds.

A phonocardiogram (PCG) is a visual depiction of heart sounds captured using an electronic stethoscope. The intensity, rate, and inter-relationship among the heart cycles can depict the heart conditions. So, analysing the heart sounds is the best way to detect whether a person has a healthy heart. PCG signals and the assistance of artificial intelligence (AI) can objectively decode cardiac sounds to complement the traditional diagnosis method using an echocardiogram.

This work uses deep learning algorithms to categorise heart sounds. These algorithms can learn complex patterns and extract crucial information from the data. Ensemble learning enhances a model’s efficacy by minimising the variance and bias of weak base learners and increasing the accuracy of strong base learners [3]. Soft voting, a method used in bagging, can reduce variance, whereas weighted average, a boosting method, can reduce bias. The blended ensemble model [4] improves efficiency by increasing the accuracy of the base learners. This work proposes a blended ensemble model, HbNet, consisting of Convolutional Neural Network – Long Short Term Memory (CNN-LSTM) and CNN-Bidirectional Long Short Term Memory (BiLSTM) as base learners and softmax regression as a meta-learner. The PCG signals are preprocessed and feature-mapped using MFCC before being applied to base learners. The data sets used for the study are the GitHub dataset [5] for 5-class adult PCG classification and a self-collected 5-class paediatric dataset to analyse the effectiveness and adaptability of the model for a different dataset. The GitHub dataset comprises the heart sounds of Aortic stenosis (AS), Mitral Regurgitation (MR), Mitral Stenosis (MS), Mitral Valve Prolapse (MVP), and Normal (N). The paediatric dataset consists of Atrial Septal Defect (ASD), Functional Murmur (F), Normal, Pathological (P), and Ventricular Septal Defect (VSD). This work also studies the performance of other prominent ensemble models, soft voting and weighted average ensemble models. The significance of the study is summarised as follows:

1.
The work is the first study in its field to classify PCG signals into 5 distinct categories using a novel method that incorporates a blended ensemble model with hybrid deep learning models and softmax regression.
2.
HbNet is a pioneering model that classifies paediatric heart sounds into more than three classes using a deep learning algorithm. Existing studies categorised PCG data into either two (normal/pathological) or three (ASD, VSD, Normal) categories, whereas the proposed method classifies paediatric PCG into 5 classes (ASD, F, N, P, VSD). Accurate detection of functional murmurs aids in minimising unnecessary echocardiograms, while the precise detection of diseases facilitates the timely initiation of treatment.
3.
The proposed model outperformed all seven evaluated metrics compared to the existing state-of-the-art methodologies. This achievement underscores the significance and relevance of the study.
4.
The study collected and organised a new paediatric PCG dataset comprising data in five categories.
5.
The HbNet model predicted the functional murmur, the most misdiagnosed murmur, with 100% precision and specificity, exemplifying the model’s exceptional performance.

2. Literature review

Artificial intelligence (AI) based Cardio Vascular Disease (CVD) detection is carried out using either deep learning or machine learning [6] techniques, as mentioned in the introduction. AI methods involve denoising, segmentation, feature extraction, and classification. The primary denoising techniques are: 1) Adaptive Algorithms to eliminate the lung sounds [7, 8, 9]. 2) Blind Source Separation Techniques [10, 11], which use the resemblance of spectral information to identify the noise. 3) Wavelet Transforms (WT) with various thresholding techniques [12, 13, 14, 15, 16, 17]. 4) The AR-based Kalman method [18], in which AR tries to find the coefficients, and Kalman uses the coefficients for denoising. This method improves the signal-to-noise ratio.

The most common computer-aided segmentation techniques used to recognise the first heart sound (S1) and second heart sound (S2) are 1) Wavelet Transform, which provides excellent results even in a noisy environment. The wavelet-based segmentation techniques are Adaptive Wavelet Decomposition [19], Mel-Scaled Wavelet Transform (MSWT) [20], DWT with the Fano Factor and Normalised Average Shannon Energy [21], a combination of Empirical Wavelet Transform [22, 23] and Shanon Entropy Envelope (SEE) [24, 25, 26], and Continuous Wavelet Transform [27]. 2) Hilbert Transform exhibits good segmentation accuracy for progressive waves. Recent works [28, 29, 30] use the Hilbert Transform to segment heart sounds. 3) The Hidden Markov Model (HMM) based approach [31, 32, 33, 34] explores non-stationary statistical properties for the segmentation. 4) Deep Learning based segmentation methods involve deep Recurrent Neural Networks (RNN) [35], Bidirectional LSTM (BiLSTM) with Attention [36], and deep Convolutional Recurrent Neural Networks (CRNN) [37]. Deep learning methods provide more excellent results than manual segmentation methods for accurate training data.

The feature extraction process involves the efficient representation of the raw data in another form. Localisation of S1, S2, systole, and diastole, spectrograms using various transforms, MFCC [38], and Gammatone Frequency Cepstral Coefficients (GFCC) [31] are the commonly used feature extraction techniques. Classification is the final stage and can be done using machine learning approaches like K-Nearest Neighbour (K-NN) [39], Support Vector Machine (SVM) [22, 40, 41, 42, 43], Random Forest (RF) [44], and Artificial Neural Network (ANN) [45, 46, 47]. In [48], a comprehensive study was conducted which utilised classification methods such as K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Multilayer Perceptron (MLP), and Deep Neural Network (DNN). The results showed that DNN outperformed the other machine learning algorithms significantly. This finding strongly emphasises the advantages of deep learning algorithms over traditional machine learning methods. Classifying data using machine learning algorithms is a meticulous process requiring data segmentation and crafting intricate features manually. Though this process helps improve classification models’ overall performance for small datasets, it is time-consuming and labour-intensive. Deep learning models, in contrast, autonomously extract relevant features from data. The prominent deep learning algorithms used are CNN [49, 50, 51, 52, 53, 54], RNN [51, 55, 56], Gated Recurrent Unit (GRU) [51, 57], a combination of CNN and RNN, generally known as CRNN [58, 59, 60, 61, 62, 63], Transfer Learning [64, 65, 66] and Convolutional Transformers [67]. The hybrid models (CRNN) performed better than stand-alone deep learning models. These models were lightweight and exhibited great results without segmentation. Ensemble learning enhances classification accuracy over the use of a single classifier by combining multiple classifiers [68, 69, 70, 71, 72]. In [26],

Table 1
Summarised studies on the recent adult and paediatric PCG

Author, year	Methodology	Validation	Results	No. of classes	Remarks
Adult PCG classification
Boulares et al., 2020 [65]	Transfer learning using Matlab Pre-trained CNN models	3 fold	Max ACY: 77%(VGG 19)	3(Normal, Murmur, Extra-systole)	Set a benchmark for transfer learning in PCG classification. Accuracy is not great
Jahmunah et al., 2020 [61]	WaveNet architecture CNN $+$ LSTM $+$ Autoencoders	10 Fold	Train ACY: 97.0%RECALL: 92.5%SPEC: 98.1%	5(AS, MR, MS, MVP, N)	No manual feature extraction. Computationally complex and requires large memory space
Zeng et al., 2021 [46]	Tunable Q-factor wavelet transform (TQWT)Variational mode decomposition (VMD)Phase space reconstruction (PSR)Artificial neural networks	10 fold	ACY: 97.89%	2(Normal, Abnormal)	Complex Architecture and requires more training time
Hazeri et al., 2021 [43]	Springer’s segmentation algorithmTime-frequency domain featuresSVM	10 fold	ACY: 86%SENS: 93.3%SPEC: 78.5%	2(Normal, Abnormal)	Time-frequency feature-based classification. Requires additional ECG signal for segmentation
Sameul Shuvo et al., 2021 [59]	3 CNN Feature extractorsBiLSTM	10 fold	ACY: 99.6% SENS: 98.3% SPEC: 99.58	5(AS, MR, MS, MVP, N)	A lightweight model with high accuracy
Alkhodari et al., 2021 [62]	MODWT $+$ CNN-BiLSTM	10 fold	ACY: 98.7SENS: 98.6SPEC: 99.5	5(AS, MR, MS, MVP, N)	Less training and testing time. No significant improvement in performance
Khan et al., 2022 [52]	Power Spectral DensityCNN	10 fold	ACY: 98.8SENS: 97.3SPEC: 99.2	5(AS, MR, MS, MVP, N)	Lightweight model, but its accuracy is lesser than CardiXnet
Al-Issa et al., 2022 [63]	FFT $+$ CNN $+$ LSTM	10 fold	ACY: 98.8F1-Score- 98.5	5(AS, MR, MS, MVP, N)	No manual feature extraction. Less computational complexity, low accuracy
Pachori et al., 2022 [23]	Fourier-Bessel Series Expansion-basedEmpirical Wavelet TransformPhase Space Reconstruction	10 fold	ACY: 98.8SENS: 97.3SPEC: 99.2	4(AS, MR $+$ MVP, MS, N)	Lower computational complexity and comparable results with DL algorithms
Qaisar et al., 2022 [67]	Convolutional Vision Transformer	10 fold	ACY: 99SENS: 96.7SPEC: 96.3	5(AS, MR, MS, MVP, N)	Performance lower than cardioXnet, considerable computational complexity but lower compared to predominant transfer learning techniques
Shahid et al., 2023 [66]	Chirplet Z transform Transfer Learning	–	ACY: 85–99%	5 (AS, MR, MS, MVP, N)	Computational cost is low
Bhardwaj et al., 202 [53]	Continuous Wavelet Scalograms 2D CNN	5 fold	ACY: 98.32%	5 (AS, MR, MS, MVP, N)	Simple model and accuracy are lesser than existing systems

Table 1, continued
Author, year	Methodology	Validation	Results	No. of classes	Remarks
Karhan et al., 2022 [48]	Hilbert-Huang transform Mel-frequency cepstral coefficient KNN, SVM, MLP, DNN	5 fold	ACY: 98.9%	5 (AS, MR, MS, MVP, N)	Complex feature extraction methods and accuracy are lower than lightweight deep learning models
Roy et al., 2023 [73]	CNN Squeeze Net Random Forest	5 fold	ACY: 99.3%	5 (AS, MR, MS, MVP, N)	Noise can degrade the performance as noise cannot be filtered completely
Paediatric PCG classification
Aziz et al., 2020 [22]	Empirical mode decomposition1D Local ternary patternMFCCSVM	10 fold	Mean ACY: 95.25%	3(N, VSD, ASD)	Three handcrafted features are required. Only 56 subjects were considered in the study
Gharehbaghiet al., 2020 [47]	Time Growing Neural Network	5 fold	ACY: 88.5%	2(Atrioventricular defects, others)	The work fails to distinguish between normal and other diseases
Wang et al., 2020 [25]	PCG decomposition recombination methodShanon Energy Entropy (SEE)Artificial Neural Network	Nil	ACY: 93%SENS: 93.5%SPEC: 91.7%	2(Normal, Abnormal)	Noises due to crying and movement cannot be removed. Limited dataset (86 samples)
Gómez et al., 2021 [72]	Manual Segmentation67 Time domain featuresXGBoost	5 fold	CHD/N ACC: 78PDA/N ACC: 77	2CHD/NormalPDA/Normal	Lower Classification accuracy compared to existing methodologies
Edin et al., 2021 [84]	SVM	Nil	ACC: 97%	2 Innocent/Organic	Many data sources are required, such as ECG, blood results, and clinical parameters, in addition to auscultation, which is a limitation
Chen et al., 2022 [58]	CNN-LSTM	Nil	ACC: 85 SENS: 87 SPEC: 82	2 Normal/Abnormal	No segmentation and feature extraction. Lightweight model
Weize et al., 2022 [26]	SEEEnsemble majority voting Random Forest and Adaboost	5 fold	ACC: 95SENS: 94.6SPEC: 96.1	2Normal/Abnormal	Largest paediatric dataset, Complex architecture with large computational complexity and training time
Shekhar et al., 2022 [54]	Segmentation Spectrogram CNN	5 Fold	M/NM SENS: 90 S/NM SENS: 88.4	Murmur (M)/No Murmur (NM) Still (S)/No Murmur	Not mentioned how segmentation or feature extraction done
Zafar et al., 2023 [75]	LPF (0–650) Chunking Padding Pitch shifting 1D CNN	–	ACC: 98.56%	Normal/Abnormal	Results not validated using cross-validation
Sabouri et al., 2023 [76]	Morphology of the heart sound signals’ frequency domains statistical features in the frequency ranges	–	ACC: 96.6	Normal/Abnormal	Simple methodology without using AI algorithms

majority voting utilising Random Forest (RF) and AdaBoost is employed to categorise the paediatric cardiac sounds into two categories, yielding impressive outcomes. Roy et al. [73] proposed a CNN-RF ensemble model and achieved 99.37% accuracy for adult PCG classification. These works not only validate the effectiveness of the ensemble approach but also highlight the potential of utilising ensemble models to achieve superior performance. So, this study combines the merits of hybrid deep learning models and ensemble models to propose HbNet, a novel blended ensemble model that integrates two hybrid models, CNN-BiLSTM and CNN-LSTM, using a softmax regression as a meta-learner.

Recent works on adult and paediatric PCG are showcased in Table 1. Unlike the previous paediatric papers that classified the PCG into normal and pathological [25, 26, 54, 58, 74, 75, 76], this study classifies it into five different classes such as ASD, Functional murmur, Normal, Pathological and VSD, making the work novel. HbNet facilitates the early identification of VHD from heart sounds.

3. Materials and methods

The work proposes a blended ensemble model, HbNet, incorporating CNN-BiLSTM and CNN-LSTM as base learners and softmax regression as a meta-learner for detecting cardiovascular anomalies. Before finalising the base models, we conducted experiments using various deep learning architectures, including CNN, LSTM, GRU, BiLSTM, and hybrid models like CNN-LSTM, CNN-GRU, and CNN-BiLSTM. Notably, the CNN-BiLSTM and CNN-LSTM configurations demonstrated exceptional performance, which led to their selection as the base models. Since the study uses a not-so-large dataset of one-dimensional heart sounds, it would be ideal to introduce feature mapping along with deep learning feature extraction. The acoustic feature mapping employed is MFCC. The mapped features are given to HbNet for classification. The HbNet model is depicted in Fig. 1. In addition to the blended ensemble model, the possibilities of the soft voting method [77] and the weighted average ensemble model [78] have also been explored. Both of these models are implemented using the same two base models.

Figure 1.

HbNet model.

3.1 Dataset

3.1.1 GitHub dataset

The data utilised in this study are acquired from a publicly available database [5]. The PCG signals available in the dataset are collected from various medical websites, such as OpenMichigan, the University of Washington Department of Medicine, and auscultation tutorial CDs available in medical textbooks. The dataset is structured into five categories: N, AS, MR, MS, and MVP. Each category comprises 200 PCG samples, amounting to a cumulative count of 1000 samples. The duration of heart sounds varies from 1.12 s to 2 s. The dataset has been validated by consulting with an expert clinical cardiologist from TMM Hospital.

3.1.2 Paediatric dataset

The Paediatric dataset was collected from Thiruvalla Medical Mission Hospital (TMM) after receiving approval from the local institutional ethics committee. The data was collected between 26 August 2022 to 20 April 2023. A total of 2322 data samples were collected from 391 subjects in the age range of newborns to 13 years old. Heart sound recordings were done from four auscultation points: the aortic, pulmonary, mitral, and tricuspid. The time span of the recordings is between 3 and 30 seconds. Data collection was conducted by an expert cardiologist in a clinical environment using Eko Littmann’s electronic stethoscope. It is an unbalanced dataset comprising 254 ASD, 218 functional, 821 normal, 674 pathological, and 355 VSD PCG recordings.

3.2 Preprocessing

The audio files in the GitHub dataset vary in duration, ranging from 1.12 seconds to 30 seconds. To make them identical in duration, the PCG signals have been truncated to 1.12 s, the shortest heart sound duration available in the dataset. Then, the files are resampled to 8000 Hz and normalised between $+$ 1 and $-$ 1.

3.3 Feature mapping

The features are represented using MFCC [79]. It is a very efficient representation technique for speech and low-frequency signals. Since heart sound occupies a low-frequency range, MFCC is one of the best methods for representation [80]. 13 MFCC coefficients are considered in this work. The steps to obtain the MFCC features are shown in Fig. 2.

Figure 2.

Steps involved in MFCC feature mapping.

First, the time-domain PCG signal is modified into the frequency-domain signal by the Eq. (1):

$\displaystyle y({f,\tau})=\mathop{\int}\limits_{-\infty}^{\infty}W({t-\tau})P(% t)\text{e}^{-j2\pi f\tau}\text{d}t$ (1)

$x(t)$ is the time domain signal. $y({f,\tau})$ is the frequency domain signal. $W({t-\tau})$ is a Hamming window.

$\displaystyle F_{\textit{mel}}=2595\log_{10}\left(1+\frac{f_{p}}{700}\right)$ (2)

$F_{\textit{mel}}$ represents the Mel frequency extracted from the PCG signal, and ${f_{p}}$ denotes the original frequency of the PCG signal. Transforming the Mel frequency into Log Mel is achieved by taking the logarithm. Subsequently, to derive the Mel Frequency Cepstral Coefficients (MFCC), we apply the Discrete Cosine Transform to the Log Mel data.

$\displaystyle D_{t}(n)=\mathop{\sum}\limits_{m=0}^{N-1}L_{t}(m)\cos\left({% \frac{\pi n({m-0\cdot 5})}{T}}\right)$ (3)

where $n$ denotes the total MFCC coefficient. $D_{t}(n)$ is the $n^{\text{th}}$ coefficient in the $t^{\text{th}}$ frame. $L_{t}(m)$ represents the logarithmic Mel spectrum of the PCG signal; $T$ stands for the number of triangular filters used. The parameters, such as hop length, length of the vector subjected to FFT, and window size, are the same as those in our paper [60].

3.3.1 Statistical significance of MFCC features

A statistical inference test called Analysis of Variance (One-Way ANOVA) was performed to demonstrate the relevance of the obtained MFCC features. The Python statsmodel package was used to run the test and received the F-statistic value of 50 with the associated $P$ value of 6 $\times$ 10^{- 9}. The hypothesis considered for the test was that the mean MFCC features of different classes share the same mean value. The results reject the hypothesis, as the $P$ value is significant at a 99% confidence level. The $F$ and $P$ values achieved for the paediatric dataset are 5.44 and 0.01, respectively. The hypothesis has been rejected, and the features are statistically significant.

3.4 HbNet model

HbNet, a blended ensemble model, can provide consistent predictions with greater accuracy than the individual models. In blended architecture, one base model will be proficient at identifying a set of features, and the other model will recognise a different set of features. Combining them will make the ensemble model benefit from the merits of the base models. The two base models employed in this work are CNN-BILSTM and CNN-LSTM. CNN can model time and frequency-domain components in a signal, whereas BiLSTM and LSTM uncover the long and short-term temporal dependencies of features among and within PCG signals. LSTM considers only past information, while BiLSTM preserves past and future temporal information. Incorporating LSTM and BiLSTM in a single model can facilitate a more accurate classification of PCG. In this study, the initial predictions are generated by applying MFCC outputs to the base models. The meta-learner is Softmax Regression, as HbNet is designed for multiclass classification problems. The architecture is intended for accurately classifying short PCG recordings that are not segmented.

Figure 3.

Base model architecture.

3.4.1 Base model architecture

The base models that constitute the HbNet are CNN-BiLSTM and CNN-LSTM. Rigorous hyperparameter tuning [81] has been performed to determine the optimum base model architecture, which can enhance the total performance of HbNet. The CNN-LSTM comprises three convolutional (Conv) layers subsequently succeeded by max-pooling and 4 LSTM layers, as shown in Fig. 3. Each Conv layer output (C) is represented by Eq. (4).

$\displaystyle C_{i}^{lj}=A(b_{j}+\mathop{\sum}\limits_{k=1}^{K}W_{k}^{j}I_{i+k% -1}^{j}$ (4)

Where $l$ represents the layer index, $A$ denotes the activation function, $b$ signifies the bias of the $j^{\text{th}}$ feature, $k$ represents the kernel size and $W$ weight associated with the filter, and $I$ stands for input after feature representation. The Conv layers of the model are designed with 256, 128, and 64 neurons, respectively. The Conv output (C) is given to LSTM, and LSTM output can be modelled by Eq. (5).

$\displaystyle O_{t}=\sigma(W_{o}\ast[{HD_{t-1},C_{t}}]+b_{o}$ (5)

$o_{t}$ is the output gate, $\sigma$ is the sigmoid function, and HD represents the hidden state. The LSTM module comprises 256, 128, 64, and 32 units. The CNN-BiLSTM model has a similar architecture as the CNN-LSTM, except the third CNN layer consists of only 32 neurons instead of 64, and LSTM is replaced with BiLSTM. The BiLSTM handles data in both forward and backward directions, and the output equation changes as given in Eq. (6). The detailed model is depicted in Fig. 3.

$\displaystyle Ob_{t}=\sigma(W_{o}\ast\{{\overrightarrow{HD_{t-1,}O_{t}],}% \overleftarrow{HD_{t-1},O_{t}]}}\}+b_{o}$ (6)

For both base models, the kernel size employed for CNN layers is 3 $\times$ 3, and max-pooling is 2 $\times$ 2 with stride 1. The max-pooling layer will extract the most significant characteristics from the PCG signal and lower the subsequent layers’ computational complexity. ReLU was selected as the activation function. Leaky ReLU and tanh were also investigated as potential activation functions. The CNN modules extract time-invariant features, which are then flattened and applied to either the BiLSTM module for the CNN-BiLSTM model or the LSTM module for the CNN-LSTM model. The output of BiLSTM/LSTM is given to a dense layer with 32 nodes. A dropout of 0.3 is added to avoid overfitting. Finally, the softmax layer classifies the output from the dropout layer into five VHD classes.

3.4.2 Meta learner

HbNet comprises two hybrid deep learning models that excel in classifying PCG signals, demonstrating remarkable performance. Therefore, as the meta-learner, we require a simple and efficient algorithm that does not involve complex training procedures. Softmax regression is a simple model that can handle multiple classes with minimal training and produce maximum output. Another advantage of softmax regression is its reduced susceptibility to overfitting, especially when working with a small and low-dimensional dataset. This work involves a not-so-large one-dimensional dataset with two effective base models. This characteristic makes Softmax Regression an excellent choice for this work. The softmax regression meta-learner can be modelled by Eq. (7).

$\displaystyle S_{i}=\frac{e^{({\theta_{i}})}}{\mathop{\sum}\nolimits_{k=1}^{c}% e^{\theta_{ik}}}$ (7)

Where $\theta_{i}$ , the linear combination of output from base learners. $C$ is the number of classes and $S_{i}$ , the softmax probability. The predicted output is given by Eq. (8).

$\displaystyle P_{\textit{Out}}=\text{argmax}({S_{i}})\text{ for all class }C$ (8)

The penalty applied in this work to obtain an accurate model is L2 regularisation. L2 penalty also helps in reducing the model complexity. The L-BFGS solver has been selected as it requires less memory space and converges fast for small and medium datasets. The other solvers that have experimented with developing an accurate HbNet model are newton-cg and sag.

3.5 Training and testing of the model

The datasets are divided into train, validation, and testing. The testing dataset takes up 10% of the total, while validation also takes up 10%, and the remaining 80% is used for training. The number of hidden layers and units in each layer were determined by a trial and error approach. The base model’s hyperparameters, including the activation function, optimiser, batch size, and learning rate were selected through careful tuning using a Grid search algorithm. The hyperparameters considered and the optimal parameters selected after the Grid search are consolidated in Table 2.

Table 2
Optimised hyperparameters

Models	Hyperparameters	Grid search range	Optimal parameter
CNN-LSTM/	Activation function	ReLu, tanh, Leaky Relu	ReLu
CNN-BiLSTM	Optimiser	Adam, RMSprop, SGD	Adam
	Batch size	[16, 32, 64]	32
	Learning rate	[0.01, 0.001, 0.0001, 0.00001]	0.0001
Softmax regression	Solver	L-BFGS, newton-cg and sag.	L-BFGS

The model was designed to run for 100 epochs, but the results demonstrated no further improvement in model accuracy after 40 epochs, so the number of epochs engaged was reduced to 40. The output of the base models is concatenated and given to Softmax Regression. Then, the HbNet model is trained for a hold-out set and evaluated using test data to analyse its effectiveness. The entire model is validated using 10-fold cross-validation. The hardware for developing the HbNet model is an Intel Core i7 processor with Nvidia GeForce GTX 1060 Graphics processing and 6 GB RAM. The model architecture was designed using Python 3.7.10 and built using TensorFlow and Keras.

3.6 Model evaluation parameters

The effectiveness of the HbNet model is demonstrated through evaluation metrics such as average class accuracy, precision, recall, specificity, F1 score, Mattews Correlation Coefficient (MCC), AUC-ROC score and AUPRC score. The ROC curve was plotted to comprehend how well HbNet performed.

4. Results and discussion

4.1 GitHub dataset

The proficiency of the HbNet for the GitHub dataset has been compared with existing methodologies, the soft voting ensemble model and a weighted average ensemble model. HbNet outperformed all the existing models with an average class accuracy of 99.72%, precision, recall, and F1 scores of 99.3%, specificity of 99.8%, MCC of 99%, an AUC-ROC score of 99.9%, and an AUPRC score of 99.65%. The performance metric values of HbNet are shown in Table 3. Based on the data presented in Table 3, it is clear that HbNet consistently demonstrated exceptional performance across all 10 folds. The findings demonstrate the efficiency of the HbNet model.

Table 3
Performance evaluation of HbNet for GitHub dataset

	Precision	Recall	F1-score	Specificity	Average class accuracy	MCC	AUC-ROC score	AUPRC
1	99.2	99.2	99.2	99.72	99.6	98.6	99.72	99.24
2	100	100	100	100	100	100	100	100
3	99.2	99	99.2	99.92	99.6	98.5	99.88	99.3
4	98.8	98.8	98.8	99.8	99.6	98.7	99.96	99.24
5	100	100	100	100	100	100	100	100
6	100	100	100	100	100	100	100	100
7	96.8	97.2	97	99.4	98.8	95.9	99.98	99.3
8	100	100	100	100	100	100	100	100
9	100	100	100	100	100	100	100	100
10	99	99.4	99	99.8	99.6	98.8	99.98	99.42
Average	99.3	99.36	99.32	99.86	99.72	99.05	99.95	99.65

Table 4

The accuracy of different models experimented with in this work using GitHub

Models	Average class accuracy
CNN	99.2
LSTM	98.79
BiLSTM	99
GRU	98.79
CNN-LSTM	99.4
CNN-BiLSTM	99.6
CNN-GRU	99.2
Soft voting	99.56
Weighted average	99.68
HbNet	99.72

Table 4 contains the accuracy of the base models, the soft voting, the weighted average ensemble, and HbNet. Since the CNN-LSTM and CNN-BiLSTM showed the top accuracies, these models were chosen as base models for HbNet. The CNN-GRU model has the same architecture as CNN-BiLSTM with 3 CNN layers (256,128,32) and 4 GRU (256,128,64,32) layers. Table 4 proves that the soft-voting ensemble could not enhance the overall performance compared to the base model CNN-BiLSTM.

The soft voting ensemble performs predictions on base models, sums all the output class probabilities, and chooses the category having the highest summed probability as the outcome. Each base model is an equal part of the final prediction. The model can be depicted by an equation.

$\displaystyle l=\text{argmax}_{j}\mathop{\sum}\limits_{i=1}^{K}b_{i}^{j}(x)=% \text{argmax}_{j}\mathop{\sum}\limits_{i=1}^{K}P\left({\frac{l_{j}}{x}}\right)$ (9)

Where $b_{i}$ represents the $i^{\text{th}}$ model in an ensemble of $K$ distinct classifiers, $N$ is the label count. $b_{i}^{j}(x)$ is the output of the $i^{\text{th}}$ model for the $j^{\text{th}}$ label. $P\left({\frac{l_{j}}{x}}\right)$ denotes the posterior probability of the output generated by the $i^{\text{th}}$ classifier. The average class accuracy obtained is 99.56%, roughly equal to the accuracy of the CNN-BiLSTM. The results conclude that the soft voting ensemble with CNN-BiLSTM and CNN-LSTM is not an excellent model for classifying heart sounds.

Examining the outcomes of the weighted average ensemble model from Table 4, it is apparent that the model’s efficiency has been enhanced compared to base learners. The weighted average ensemble model combines the soft voting ensemble with a weighted score, in which the base model’s performance weights each member’s contribution to the final results. Instead of summing with equal weights, different weights of importance were applied to each base model in the ensemble. Then, the output probabilities are multiplied with weights and summed. The model can be described by the equation.

$\displaystyle l=\text{argmax}_{j}\mathop{\sum}\limits_{i=1}^{K}w_{i}.b_{i}^{j}% (x)=\text{argmax}_{j}\mathop{\sum}\limits_{i=1}^{K}w_{i}.P\left({\frac{l_{j}}{% x}}\right)$ (10)

The weights (w) are selected by the grid search method. For each base model, the weights are varied from 0 to 0.5. All combinations of weights are considered so that total weights are added up to 1. The final model will select the weight combination that provides the highest accuracy. The weighted ensemble model gained an average class accuracy of 99.68%, surpassing the accuracy of the base models and outperforming other established methodologies.

Table 5

Class-wise analysis of evaluation metrics on the Github dataset

Classes	Precision	Recall	F1-score	Specificity	Accuracy	AUC-ROC score	AUPRC
AS	99.2	99.2	99.4	99.86	99.8	99.96	99.64
MR	100	99.2	99.2	100	99.6	99.94	99.62
MS	98.6	99.4	99	99.64	99.6	99.92	99.38
MVP	99.2	99.2	99.4	99.86	99.6	99.96	99.9
N	99.6	99.8	99.6	99.86	99.9	99.94	99.72

Figure 4.

AUC-ROC Curve for Github dataset. 4(a) AUC-ROC curve for highest and 4(b) for lowest yielded iterations.

Table 5 demonstrates the class-wise performance of the model. HbNet could recognise each class with great accuracy and precision. The improved accuracy can be attributed to the fact that the blended ensemble model undergoes two rounds of training.

The AUC-ROC curve for the iterations that yielded the highest and lowest accuracy is demonstrated in Fig. 4a and b. The very high AUC-ROC score, almost equal to one, confirms that the model can effectively discriminate the different classes.

4.1.1 Comparison with existing works on the GitHub dataset

The paper presents a detailed comparison of the HbNet with all the state-of-the-art studies that used the same GitHub dataset employed in this study, as demonstrated in Table 6. Compared with the existing work, HbNet achieved better results for all the metrics evaluated. None of the existing papers has analysed the precision-recall score. This work has achieved a very high precision-recall score of 99.65%. It is evident from Table 6 that HbNet outperformed all the existing works in all aspects.

Table 6
Comparison of HbNet with existing works on GitHub dataset

Author	Acc	Sens	Spec	Precision	F1-score	AUC-ROC	MCC	Validation
Soonil et al., 2018 [5]	97.9	98.2	99.4	–	–	–	–	5 fold
Jahmunah et al., 2020 [61]	97	92.5	98.1	–	–	–	–	10 fold
Neeraj et al., 2020 [49]	98.6	–	–	–	–	–	–	10 fold
Samiul et al., 2021 [59]	99.6	98.6	–	98.6	98.5	–	–	10 fold
Mohanad et al., 2021 [62]	99.32	98.3	99.58	98.5	98.32	99.8	–	10 fold
J S Khan et al., 2022 [52]	98.87	97.3	99.2	96.7	96.7	–	–	10 fold
Al-Issa et al., 2022 [63]	98.5	–	–	–	98.5	99.78	–	10 fold
Pachori et al., 2022 [23]	98.5	96.3	99.05	–	96.29	99.74	–	10 fold
Qaisar et al., 2022 [67]	99	96.7	96.3	96.3	96.3	–	–	10 fold
Karhan et al. 2022 [48]	98.9	98.7	–	98.5	98.9	98.8	97.7	5 fold
Roy et al., 2023 [73]	99.37	99.5	98.9	98.5	–	–	–	5 fold
This study	99.72	99.3	99.8	99.3	99.3	99.95	99.05	10 fold

4.2 Performance of the HbNet model on the paediatric dataset

The HbNet model performed well on the paediatric dataset as well. In order to optimise the results, the number of epochs has been increased to 50 from 40 while training the paediatric dataset. The HbNet achieved a maximum accuracy of 87.6% with a mean accuracy of 86.5% across the 10 folds. The detailed performance evaluation of the HbNet on the paediatric PCG is shown in Table 7. The class-wise analysis of evaluation metrics is given in Table 8. Table 8 shows that the most misdiagnosed functional murmur is identified with 100% precision and specificity.

Table 7
Evaluation results of the HbNet model on the paediatric dataset

Iteration number	Precision	Recall	F1-score	Specificity	Accuracy	MCC	AUC-ROC score	AUPRC
1	87.4	86	86.6	96	86.3	83	96.9	92.86
2	89	87	88	96.4	87.6	84.8	97.64	94.6
3	85.8	83.8	84.6	96.12	86.3	80.9	97.36	94.44
4	88.2	85.2	86.4	96.12	85.4	82.5	97.36	94.44
5	87.8	82.8	84.4	95.94	85.4	81.2	96.38	91.58
6	88.4	84.4	86	96.44	85.8	82.3	96.94	92.02
7	85.6	85.6	85.4	96.32	86.7	81.8	97.86	93.16
8	86.6	87.2	86.8	96.78	87.6	83.5	97.8	93.24
9	87.4	83.2	85	96	85.8	81.3	95.98	91.38
10	90.4	83.6	86.2	96.32	87.6	83.1	97.44	93.14
Average	87.66	84.88	85.94	96.24	86.5	82.44	97.16	93.08

Table 8

Class-wise evaluation of the paediatric dataset

Classes	Precision	Recall	F1-score	Specificity	AUC-ROC score	AUPRC
ASD	82.6	68.9	74.5	98	96	85.95
F	100	96	97.8	100	99.94	99.93
N	87	92.6	89.6	92.12	97.07	95.24
P	82.9	83.6	83.2	93.8	95.11	91.4
VSD	85.8	83.3	84.6	97.28	97.71	92.91

Figure 5.

AUC-ROC Curve for the paediatric dataset. 5(a) AUC-ROC curve for highest and 5(b) for lowest performing iteration.

Table 9 displays the performance of ensemble models on the paediatric dataset. The accuracy of all three ensemble models was higher than that of the base models, out of which the HbNet model excelled with an average accuracy of 86.5%.

Table 9

Accuracy of different ensemble models for the paediatric dataset

Models	Average accuracy
CNN	60
GRU	57.9
LSTM	58.2
BiLSTM	64
CNN-LSTM	77.6
CNN-BiLSTM	83
CNN-GRU	72.4
Soft voting	84.9
Weighted average	85.8
HbNet	86.5

The AUC-ROC curves for the iterations with maximum and minimal accuracy are presented in Fig. 5a and b.

From the result analysis of the HbNet performance on Github and paediatric datasets, it is evident that there is a significant performance difference. The difference is that both datasets handle a distinct set of diseases, and the real-time paediatric dataset is much noisier than the GitHub dataset due to the crying and movements of paediatric subjects. Also, the heart rate is a varying parameter for paediatric subjects.

4.2.1 Comparison with existing works on the paediatric dataset

Comparing our study directly with other paediatric studies is challenging for significant reasons. Firstly, most earlier studies have focused on binary classification scenarios and relied on their self-gathered datasets, which are not publicly available. In contrast, this study is designed to classify PCG into five distinct categories. The comparison is given in Table 10.

Table 10
Comparison of HbNet with existing works on paediatric pcg classification

Author	Acc	Sens	Spec	Precision	F1-score	AUC-ROC	MCC	Validation
2 Class
Gharehbaghi et al., 2020 [47]	91.6	88.4	–	–	–	–	–	5 fold
Wang et al., 2020 [25]	93	93.5	91.7	–	–	–	–	–
Sergi et al., 2021 [85]	–	–	–	–	–	78	–	5 fold
Ding et al., 2022 [58]	85	87	82	–	–	99.8	–	–
Weize et al., 2022 [26]	95	94.6	96.1	–	95.3	–	–	5 fold
Shekhar et al., 2022 [54]	–	90	98	–	–	94	–	5 fold
Nurul et al., 2022 [86]	92.3	97.3	83.3	–	93.9	96.4	84.1	–
Zafar et al., 2023 [75]	98.5	99	98	98.5	98.56	–	–	–
Sabouri et al., 2023 [76]	95.45	97.4	93.6	95.54	95.45	–	–	–
3 Class
Aziz et al., 2020 [22]	95.24	95.24	9	–	–	–	–	10 fold
Kai et al., 2022 [82]	85.3	76.4	90	–	–	–	–	–
5 Class
This study	86.5	84.8	96.2	87.6	85.9	97.1	82.4	10 fold

Although Aziz et al. [22] achieved a remarkable accuracy rate of 95% for 3 class classifications, it’s important to note that the study considered a quite small sample size of only 56 participants. This limited sample size may raise questions about the broader reliability and generalizability of the results. Kai et al. [82] also performed 3 class classifications but could only achieve 85% accuracy and 76% sensitivity. As the number of classes increases, the PCG classification gets tricky as functional murmurs and many other pathologies such as ASD, VSD, and Pulmonary Stenosis (PS) have systolic murmurs, and the chances of misdiagnose are high. Despite these limitations, the proposed model achieved 86.5% accuracy, 85% sensitivity and 87.6% precision, much higher than the cardiologist’s accuracy (less than 80%).

4.3 Limitations and merits

The limitation of HbNet is that it is a model with large computational complexity and demands an extensive training period than CardioXNet [59]. HbNet requires 102 seconds to train the model and comprises 3M parameters. However, it’s essential to note that this limitation is somewhat mitigated by the substantial performance enhancements achieved compared to existing methods. Another drawback is that it fails to recognise the noises that fall in the same frequency range as heart murmurs. AI models for medical applications prioritise the model’s proficiency in accurately distinguishing the various diseases over the complexity of the model, as misdiagnosis can cause life-threatening situations and even fatalities [83]. In that perspective, HbNet is superior to all the existing works that employed the GitHub dataset, as it gained high accuracy and a remarkably low miss classification rate.

5. Conclusion and future scope

This study introduces a novel blended ensemble model, HbNet, for categorising heart sounds into normal and pathological categories. The innovative aspect of HbNet lies in its application of a unique blended ensemble model featuring the integration of CNN-BiLSTM and CNN-LSTM for the classification of heart signals. This research also explores the performance of soft voting and weighted average ensembles using the same base models. Upon analysing results and comparison with existing methods, soft voting and weighted average model, it is evident that HbNet excels in classifying PCG signals with remarkable accuracy (99.72%) and precision (99.3%). The weighted ensemble model also outperformed the existing methodologies. The HbNet has been evaluated using a real-time paediatric dataset, and the HbNet model delivered adequate results even without denoising and segmentation. Functional murmur, the category which poses the most diagnostic challenges among physicians and paediatricians, was identified with cent-percent precision. The results indicate that the model holds up well across various datasets.

In future, it is planned to incorporate other clinical features, such as auscultating points where the murmur is predominant and the subject’s heart rate, along with MFCC features, to improve accuracy. To address the challenge of reducing model complexity and the number of parameters, incorporating an attention mechanism is currently under investigation. It is also envisioned to develop a mobile application that can be used by the cardiologist as a diagnostic aid during auscultation.

Funding

The Centre for Engineering Research and Development (CERD), A.P.J Abdul Kalam Technological University, Kerala, India, funded this study (Fellowship Grant Number: 547/2019/KTU).

Author contributions

Ann Nita Netto: Conceived and designed the study and methodology, implemented the software, analysed the results and prepared the draft manuscript.

Lizy Abraham: Supervised the study and critically reviewed the manuscript.

Saji Philip: Data collection, interpretation and guidance in the research area as a medical expert.

Ethical approval

The ethical principles of this research were upheld by adhering to the Declaration of Helsinki. Ethical committee approval was obtained from the local ethical committee ECR/975/Inst/KL/2017 with study reference number 05/08/2022.

Informed consent

Informed parental consent was obtained from all participants included in the study.

Footnotes

Acknowledgments

The first author acknowledges CERD, A.P.J Abdul Kalam Technological University, Kerala, India, for supporting the work.

Conflict of interest

The authors report no conflict of interest.

References

Cardiovascular diseases (CVDs) [Internet]. [cited 2022 Oct 9]. Available from: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds).

Coffey

Roberts-Thomson

Brown

Carapetis

Chen

Enriquez-Sarano

, et al. Global epidemiology of valvular heart disease. Nature Reviews Cardiology. 2021 Dec 1; 18(12): 853-64. doi: 10.1038/S41569-021-00570-Z.

Dou

Yunus

Bui

Merghadi

Sahana

Zhu

, et al. Improved landslide assessment using support vector machine with bagging, boosting, and stacking ensemble machine learning framework in a mountainous watershed, Japan. Landslides. 2020; 17(3). doi: 10.1007/s10346-019-01286-5.

Dong

Cao

Shi

. A survey on ensemble learning. Vol. 14, Frontiers of Computer Science. 2020. doi: 10.1007/s11704-019-8208-z.

Yaseen Son

Kwon

. Classification of heart sound signal using multiple features. Applied Sciences (Switzerland). 2018; 8(12). doi: 10.3390/app8122344.

Romiti

Vinciguerra

Saade

Anso Cortajarena

Greco

. Artificial Intelligence (AI) and Cardiovascular Diseases: An Unexpected Alliance. Cardiology Research and Practice. 2020; 2020(Ml). doi: 10.1155/2020/4972346.

Gradolewski

Tojza

Redlarski

. Adaptive Neural Network Filter for denoising the phonocardiography signal. Przeglad Elektrotechniczny. 2014 Nov 1; 90: 227-30. doi: 10.12915/PE.2014.11.58.

Belloni

Della Giustina

Riva

Malcangi

. A New Digital Stethoscope with Environmental Noise Cancellation.

Khan

Onoue

Hashiodani

Fukumizu

Yamauchi

. Signal and noise separation in medical diagnostic system based on independent component analysis. IEEE Asia-Pacific Conference on Circuits and Systems, Proceedings, APCCAS. 2010. pp. 812-5. doi: 10.1109/APCCAS.2010.5775018.

10.

Tsalaile

Sameni

Sanei

Jutten

Chambers

. Sequential blind source extraction for quasi-periodic signals with time-varying period. IEEE Transactions on Biomedical Engineering. 2009 Mar; 56(3): 646-55. doi: 10.1109/TBME.2008.2002141.

11.

Almasi

Shamsollahi

Senhadji

. Bayesian denoising framework of phonocardiogram based on a new dynamical model. IRBM. 2013 Jun 1; 34(3): 214-25. doi: 10.1016/J.IRBM.2013.01.017.

12.

Liu

Wang

. Research and implementation of heart sound denoising. Physics Procedia. 2012 Jan 1; 25: 777-85. doi: 10.1016/J.PHPRO.2012.03.157.

13.

Chourasia

Tiwari

. Design methodology of a new wavelet basis function for fetal phonocardiographic signals. The Scientific World Journal. 2013; 2013. doi: 10.1155/2013/505840.

14.

Deng

Han

. Adaptive overlapping-group sparse denoising for heart sound signals. Biomedical Signal Processing and Control. 2018 Feb 1; 40: 49-57. doi: 10.1016/J.BSPC.2017.08.027.

15.

Jain

Tiwari

. An adaptive thresholding method for the wavelet based denoising of phonocardiogram signal. Biomedical Signal Processing and Control. 2017 Sep 1; C(38): 388-99. doi: 10.1016/J.BSPC.2017.07.002.

16.

Eslamizadeh

Barati

. Heart murmur detection based on wavelet transformation and a synergy between artificial neural network and modified neighbor annealing methods. Artificial Intelligence in Medicine. 2017 May 1; 78: 23-40. doi: 10.1016/J.ARTMED.2017.05.005.

17.

Gradolewski

Magenes

Johansson

Kulesza

. A wavelet transform-based neural network denoising algorithm for mobile phonocardiography. Sensors (Switzerland). 2019 Feb 2; 19(4). doi: 10.3390/S19040957.

18.

Nazemi

Hakimnejad

Azimifar

. PCG denoising using AR-based Kalman filter. 2021 29th Iranian Conference on Electrical Engineering, ICEE 2021. 2021 May 18. pp. 902-6. doi: 10.1109/ICEE52715.2021.9544365.

19.

Chen

Yang

Tsai

Chen

Chang

, et al. S1 and S2 heart sound recognition using deep neural networks. IEEE Transactions on Biomedical Engineering. 2017 Feb 1; 64(2): 372-80. doi: 10.1109/TBME.2016.2559800.

20.

Milani

MGM

Abas

De Silva

. A critical review of heart sound signal segmentation algorithms. Smart Health. 2022 Jun 1; 24: 100283. doi: 10.1016/J.SMHL.2022.100283.

21.

Jain

Tiwari

. A robust algorithm for segmentation of phonocardiography signal using tunable quality wavelet transform. Journal of Medical and Biological Engineering 2017 38: 3. 2017 Sep 25; 38(3): 396-410. doi: 10.1007/S40846-017-0320-7.

22.

Aziz

Khan

Alhaisoni

Akram

Altaf

. Phonocardiogram signal processing for automatic diagnosis of congenital heart disorders through fusion of temporal and cepstral features. Sensors (Switzerland). 2020; 20(13): 1-20. doi: 10.3390/s20133790.

23.

Khan

Qaisar

Pachori

. Automated classification of valvular heart diseases using FBSE-EWT and PSR based geometrical features. Biomedical Signal Processing and Control. 2022 Mar 1; 73: 103445. doi: 10.1016/J.BSPC.2021.103445.

24.

Nivitha Varghees

Ramachandran

. Effective heart sound segmentation and murmur classification using empirical wavelet transform and instantaneous phase for electronic stethoscope. IEEE Sensors Journal. 2017 Jun 15; 17(12): 3861-72. doi: 10.1109/JSEN.2017.2694970.

25.

Wang

You

Gong

Xie

, et al. Intelligent Diagnosis of Heart Murmurs in Children with Congenital Heart Disease. Journal of Healthcare Engineering. 2020; 2020. doi: 10.1155/2020/9640821.

26.

Chen

Yin

, et al. Automatic pediatric congenital heart disease classification based on heart sound signal. Artificial Intelligence in Medicine. 2022 Apr 1; 126: 102257. doi: 10.1016/J.ARTMED.2022.102257.

27.

Meintjes

Lowe

Legget

. Fundamental Heart Sound Classification using the Continuous Wavelet Transform and Convolutional Neural Networks. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS. 2018 Oct 26; 2018-July: 409–12. doi: 10.1109/EMBC.2018.8512284.

28.

Cherif

Debbal

. Algorithm for the estimation of pulmonary hypertension by the heart sounds using a Hilbert transform. International Journal of Biomedical Engineering and Technology. 2016; 20(4): 356-68. doi: 10.1504/IJBET.2016.076609.

29.

Saini

Nath

Tripathi

Saini

. Computation and Analysis of Heart Sound Signals using Hilbert Transform and Hilbert-Huang Transform Article in. International Journal of Engineering and Technology. 2017. doi: 10.21817/ijet/2017/v9i2/170902219.

30.

Milani

MGM

Abas

De Silva

. Identification of normal and abnormal heart sounds by prominent peak analysis. ACM International Conference Proceeding Series. 2019 Sep 20. pp. 31-5. doi: 10.1145/3364908.3364924.

31.

Alexander

Nallathambi

Selvaraj

. Screening of Heart Sounds Using Hidden Markov and Gammatone Filterbank Models. Proceedings – 17th IEEE International Conference on Machine Learning and Applications, ICMLA 2018. 2019 Jan 15. pp. 1460-5. doi: 10.1109/ICMLA.2018.00237.

32.

Kamson

Sharma

Dandapat

. Multi-centroid diastolic duration distribution based HSMM for heart sound segmentation. Biomedical Signal Processing and Control. 2019 Feb 1; 48: 265-72. doi: 10.1016/J.BSPC.2018.10.018.

33.

Oliveira

Renna

Mantadelis

Coimbra

. Adaptive sojourn time HSMM for heart sound segmentation. IEEE Journal of Biomedical and Health Informatics. 2019 Mar 1; 23(2): 642-9. doi: 10.1109/JBHI.2018.2841197.

34.

Renna

Oliveira

Coimbra

. Deep convolutional neural networks for heart sound segmentation. IEEE Journal of Biomedical and Health Informatics. 2019 Nov 1; 23(6): 2435-45. doi: 10.1109/JBHI.2019.2894222.

35.

Messner

Zöhrer

Pernkopf

. Heart sound segmentation – An event detection approach using deep recurrent neural networks. IEEE Transactions on Biomedical Engineering. 2018 Sep 1; 65(9): 1964–74. doi: 10.1109/TBME.2018.2843258.

36.

Fernando

Ghaemmaghami

Denman

Sridharan

Hussain

Fookes

. Heart sound segmentation using bidirectional LSTMs with attention. IEEE Journal of Biomedical and Health Informatics. 2020 Jun 1; 24(6): 1601-9. doi: 10.1109/JBHI.2019.2949516.

37.

Chen

Sun

Jia

Huang

. End-to-end heart sound segmentation using deep convolutional recurrent network. Complex & Intelligent Systems 2021 7: 4. 2021 Mar 29; 7(4): 2103-17. doi: 10.1007/S40747-021-00325-W.

38.

Nogueira

Ferreira

Gomes

Jorge

. Classifying heart sounds using images of motifs, MFCC and temporal features. Journal of Medical Systems. 2019; 43(6). doi: 10.1007/s10916-019-1286-5.

39.

Singh

Majumder

. Classification of unsegmented heart sound recording using knn classifier. https://doi.org/ 101142/S0219519419500258. 2019 Jun 25; 19(4). doi: 10.1142/S0219519419500258.

40.

Whitaker

Suresha

Liu

Clifford

Anderson

. Combining sparse coding and time-domain features for heart sound classification. Physiological Measurement. 2017 Jul 31; 38(8): 1701. doi: 10.1088/1361-6579/AA7623.

41.

Hamidi

Ghassemian

Imani

. Classification of heart sound signal using curve fitting and fractal dimension. Biomedical Signal Processing and Control. 2018 Jan 1; 39: 351-9. doi: 10.1016/J.BSPC.2017.08.002.

42.

. Classification of heart sounds based on the wavelet fractal and twin support vector machine. Entropy 2019, Vol 21, Page 472. 2019 May 6; 21(5): 472. doi: 10.3390/E21050472.

43.

Hazeri

Zarjam

Azemi

. Classification of normal/abnormal PCG recordings using a time-frequency approach. Analog Integrated Circuits and Signal Processing. 2021; 109(2): 459-65. doi: 10.1007/s10470-021-01867-2.

44.

Antink

Becker

Leonhardt

Walter

. Nonnegative Matrix Factorization and Random Forest for Classification of Heart Sound Recordings in the Spectral Domain. doi: 10.22489/CinC.2016.235-109.

45.

Beritelli

Capizzi

Lo Sciuto

Napoli

Scaglione

. Automatic heart activity diagnosis based on Gram polynomials and probabilistic neural networks. Biomedical Engineering Letters 2017 8: 1. 2017 Aug 22; 8(1): 77-85. doi: 10.1007/S13534-017-0046-Z.

46.

Zeng

Yuan

Wang

Liu

Wang

. A new approach for the detection of abnormal heart sound signals using TQWT, VMD and neural networks. Vol. 54, Artificial Intelligence Review. Springer Netherlands. 2021; 1613-1647. doi: 10.1007/s10462-020-09875-w.

47.

Gharehbaghi

Sepehri

Babic

. Distinguishing septal heart defects from the valvular regurgitation using intelligent phonocardiography. Studies in Health Technology and Informatics. 2020; 270: 178-82. doi: 10.3233/SHTI200146.

48.

Arslan

Karhan

. Effect of Hilbert-Huang transform on classification of PCG signals using machine learning. doi: 10.1016/j.jksuci.2021.12.019.

49.

Baghel

Dutta

Burget

. Automatic diagnosis of multiple cardiac diseases from PCG signals using convolutional neural network. Computer Methods and Programs in Biomedicine. 2020; 197. doi: 10.1016/j.cmpb.2020.105750.

50.

Liu

Zhao

Kong

Dong

Liu

, et al. Feature extraction and classification of heart sound using 1D convolutional neural networks. Eurasip Journal on Advances in Signal Processing. 2019; 2019(1). doi: 10.1186/s13634-019-0651-3.

51.

Sujadevi

Soman

Vinayakumar

Prem Sankar

. Anomaly Detection in Phonocardiogram Employing Deep Learning. Vol. 711, Advances in Intelligent Systems and Computing. Springer Singapore. 2019; 525-534. doi: 10.1007/978-981-10-8055-5_47.

52.

Khan

Kaushik

Chaurasia

Dutta

Burget

. Cardi-Net: A deep neural network for classification of cardiac disease using phonocardiogram signal. Computer Methods and Programs in Biomedicine. 2022 Jun 1; 219: 106727. doi: 10.1016/J.CMPB.2022.106727.

53.

Bhardwaj

Singh

Joshi

. Explainable Deep Convolutional Neural Network for Valvular Heart Diseases Classification Using PCG Signals. IEEE Transactions on Instrumentation and Measurement. 2023; 72. doi: 10.1109/TIM.2023.3274174.

54.

Shekhar

Vanama

John

Issac

Arjoune

Doroshow

. Automated identification of innocent Still’s murmur using a convolutional neural network. Frontiers in Pediatrics. 2022; 10. doi: 10.3389/fped.2022.923956.

55.

Raza

Mehmood

Ullah

Ahmad

Choi

. Heartbeat sound signal classification using deep learning. Sensors (Switzerland). 2019; 19(21): 1-15. doi: 10.3390/s19214819.

56.

Khan

Abid

Khan

. Automatic heart sound classification from segmented/unsegmented phonocardiogram signals using time and frequency features. Physiological Measurement. 2020; 41(5). doi: 10.1088/1361-6579/ab8770.

57.

Oliveira

Nogueira

Ramos

Renna

Ferreira

Coimbra

. Using Soft Attention Mechanisms to Classify Heart Sounds. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS. 2019. pp. 6669-72. doi: 10.1109/EMBC.2019.8856748.

58.

Chen

Xuan

Liu

Chen

Xia

, et al. Automatic classification of normal-abnormal heart sounds using convolution neural network and long-short term memory. Electronics. 2022; 11(8): 1246. doi: 10.3390/electronics11081246.

59.

Shuvo

Ali

Swapnil

Al-Rakhami

Gumaei

. CardioXNet: A novel lightweight deep learning framework for cardiovascular disease classification using heart sound recordings. IEEE Access. 2021; 9: 36955-67. doi: 10.1109/ACCESS.2021.3063129.

60.

Netto

Abraham

. Detection and Classification of Cardiovascular Disease from Phonocardiogram using Deep Learning Models. Proceedings of the 2nd International Conference on Electronics and Sustainable Communication Systems, ICESC 2021. 2021 Aug 4. pp. 1646-51. doi: 10.1109/ICESC51422.2021.9532766.

61.

Jahmunah

Ooi

Tan

Ciaccio

Yamakawa

, et al. Classification of heart sound signals using a novel deep WaveNet model. Computer Methods and Programs in Biomedicine. 2020; 196: 105604. doi: 10.1016/j.cmpb.2020.105604.

62.

Alkhodari

Fraiwan

. Convolutional and recurrent neural networks for the detection of valvular heart diseases in phonocardiogram recordings. Computer Methods and Programs in Biomedicine. 2021 Mar 1; 200: 105940. doi: 10.1016/J.CMPB.2021.105940.

63.

Al-Issa

Alqudah

. A lightweight hybrid deep learning system for cardiac valvular disease classification. Scientific Reports 2022 12: 1. 2022 Aug 22; 12(1): 1-20. doi: 10.1038/s41598-022-18293-7.

64.

Ismail

Siddiqi

Akram

. PCG classification through spectrogram using transfer learning. Biomedical Signal Processing and Control. 2023 Jan 1; 79: 104075. doi: 10.1016/J.BSPC.2022.104075.

65.

Boulares

Alafif

Barnawi

. Transfer learning benchmark for cardiovascular disease recognition. IEEE Access. 2020; 8: 109475-91. doi: 10.1109/ACCESS.2020.3002151.

66.

Ismail

. PCG signal classification using a hybrid multi round transfer learning classifier. Biocybernetics and Biomedical Engineering. 2023 Jan 1; 43(1): 313-34. doi: 10.1016/J.BBE.2023.01.004.

67.

Abbas

Hussain

Baig

. Automatic detection and classification of cardiovascular disorders using phonocardiogram and convolutional vision transformers. Diagnostics. 2022; 12(12). doi: 10.3390/diagnostics12123109.

68.

Singh

Meitei

Majumder

. Short PCG classification based on deep learning. Deep Learning Techniques for Biomedical and Health Informatics. Elsevier Inc.; 2020; 141-164. doi: 10.1016/B978-0-12-819061-6.00006-9.

69.

Daniel

Durga

Iwin Thanakumar Joseph

Angelin

Raj

SBE

. IoT-Based Ensemble Method on PCG Signal Classification to Predict Heart Diseases. EAI/Springer Innovations in Communication and Computing. 2022. doi: 10.1007/978-3-030-79766-9_7.

70.

Baydoun

Safatly

Ghaziri

El Hajj

. Analysis of heart sound anomalies using ensemble learning. Biomedical Signal Processing and Control. 2020; 62. doi: 10.1016/j.bspc.2020.102019.

71.

Noman

Ting

Salleh

Ombao

. Short-segment Heart Sound Classification Using an Ensemble of Deep Convolutional Neural Networks. ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing – Proceedings. 2019. doi: 10.1109/ICASSP.2019.8682668.

72.

Gómez-Quintana

Schwarz

Shelevytsky

Shelevytska

Semenova

Factor

, et al. A framework for ai-assisted detection of patent ductus arteriosus from neonatal phonocardiogram. Healthcare (Switzerland). 2021; 9(2): 1-20. doi: 10.3390/healthcare9020169.

73.

Roy

Mandal

. Conv-Random Forest-Based IoT: A Deep Learning Model Based on CNN and Random Forest for Classification and Analysis of Valvular. IEEE Open Journal of Instrumentation and Measurement. 2023; 2(September): 1-17. doi: 10.1109/OJIM.2023.3320765.

74.

Zhang

Wang

. Classification of Children’s Heart Sounds With Noise Reduction Based on Variational Modal Decomposition. 2022; 4(May). doi: 10.3389/fmedt.2022.854382.

75.

Ahmad

Zeeshan

Sohail

, Fazal-E-Haq Haris

Khan

, et al. Automatic Detection of Paediatric Congenital Heart Diseases from Phonocardiogram Signals. 3rd IEEE International Conference on Artificial Intelligence, ICAI 2023. 2023. pp. 188-95. doi: 10.1109/ICAI58407.2023.10136668.

76.

Sabouri

Ghadimi

Kiani-Sarkaleh

Khoshhal Roudposhti

. Effective features in the diagnosis of cardiovascular diseases through phonocardiogram. Multidimensional Systems and Signal Processing. 2023 Sep 1. doi: 10.1007/S11045-023-00876-W.

77.

Saqlain

Jargalsaikhan

Lee

. A voting ensemble classifier for wafer map defect patterns identification in semiconductor manufacturing. IEEE Transactions on Semiconductor Manufacturing. 2019 May 1; 32(2): 171-82. doi: 10.1109/TSM.2019.2904306.

78.

Manconi

Armano

Gnocchi

Milanesi

. A soft-voting ensemble classifier for detecting patients affected by COVID-19. Applied Sciences 2022, Vol 12, Page 7554. 2022 Jul 27; 12(15): 7554. doi: 10.3390/APP12157554.

79.

Peng

Chen

Zhou

Chen

Zhang

Liu

, et al. Environment sound classification based on visual multi-feature fusion and GRU-AWS. IEEE Access. 2020; 8: 191100-14. doi: 10.1109/ACCESS.2020.3032226.

80.

Delgado-Trejos

Quiceno-Manrique

Godino-Llorente

Blanco-Velasco

Castellanos-Dominguez

. Digital Auscultation Analysis for Heart Murmur Detection. 2008. doi: 10.1007/s10439-008-9611-z.

81.

Feurer

Hutter

. Hyperparameter Optimization. 2019; 3-33. doi: 10.1007/978-3-030-05318-5_1.

82.

Huang

Yang

Wang

Huang

Lin

Pan

, et al. Augmented detection of septal defects using advanced optical coherence tomography network-processed phonocardiogram. Frontiers in Cardiovascular Medicine. 2022 Nov 29; 9: 1041082. doi: 10.3389/FCVM.2022.1041082/BIBTEX.

83.

Seckanovic

Sehovac

Spahic

Ramic

Mamatnazarova

Pokvic

, et al. Review of Artificial Intelligence Application in Cardiology. 2020 9th Mediterranean Conference on Embedded Computing, MECO 2020. 2020 Jun 1. doi: 10.1109/MECO49872.2020.9134333.

84.

Begic

Pokvic

Begic

Dedic

Mrsic

, et al. From heart murmur to echocardiography congenital heart defects diagnostics using machine-learning algorithms. Medicina Academica Mostariensia. 2021; 33(2): 236-46.

85.

Gómez-Quintana

Schwarz

Shelevytsky

Shelevytska

Semenova

Factor

, et al. A framework for ai-assisted detection of patent ductus arteriosus from neonatal phonocardiogram. Healthcare (Switzerland). 2021; 9(2): 1-19. doi: 10.3390/healthcare9020169.

86.

Hasan

Abdullah

Mamun

Alkhodari

Ahmed

Khandoker

, et al. Recognition of Pediatric Congenital Heart Diseases by Using Phonocardiogram Signals and Transformer-Based Neural Networks.

HBNET: A blended ensemble model for the detection of cardiovascular anomalies using phonocardiogram

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

Table 1 Summarised studies on the recent adult and paediatric PCG

3.1.1 GitHub dataset

3.1.2 Paediatric dataset

3.2 Preprocessing

3.3 Feature mapping

3.4 HbNet model

Table 2 Optimised hyperparameters

4. Results and discussion

4.1 GitHub dataset

Table 3 Performance evaluation of HbNet for GitHub dataset

Table 6 Comparison of HbNet with existing works on GitHub dataset

Table 7 Evaluation results of the HbNet model on the paediatric dataset

Table 10 Comparison of HbNet with existing works on paediatric pcg classification

5. Conclusion and future scope

Funding

Author contributions

Ethical approval

Informed consent

Footnotes

Acknowledgments

Conflict of interest

References

Table 1
Summarised studies on the recent adult and paediatric PCG

Table 2
Optimised hyperparameters

Table 3
Performance evaluation of HbNet for GitHub dataset

Table 6
Comparison of HbNet with existing works on GitHub dataset

Table 7
Evaluation results of the HbNet model on the paediatric dataset

Table 10
Comparison of HbNet with existing works on paediatric pcg classification