Cardiac murmurs detection

Detecting cardiac murmurs is a fundamental but crucial task in heart sound analysis. Cardiac murmurs are abnormal sounds produced during the cardiac cycle and can indicate underlying heart conditions. As mentioned previously, identifying heart murmurs through traditional auscultation is challenging because it requires considerable expertise and often results in variability among clinicians.

Heart murmur detection is a more straightforward task than diagnosing a specific disease, which promotes research and development in this field. DL research is based on high-quality data, for which medical professionals and researchers have created rich datasets labeled with murmurs and shared them with the public [8,10,12,119]. Utilizing these datasets and some private datasets, researchers developed several DL models for detecting heart murmurs. PCG signals can be transformed into various types of spectrograms using different transformations and scales, thus converting temporal information into spatial representations. Murmurs display distinct features within these spectrograms, which can be accurately detected. Many models designed for this task primarily use CNN architecture for its effectiveness in processing spatial data representations [32,40,41,44,49,53–55,64,66–68,70–73,75,76,81,82,105–108,111–113]. Other models incorporate more sophisticated NN architectures, such as CNN–RNN hybrids, that leverage both spatial and sequential data processing capabilities [50,69,115]. Other advanced architectures, such as Capsule networks [38], have also been assessed for this task. The most recently developed models also integrate attention mechanisms to enhance feature extraction and improve performance [27,91].

While these models can only identify the presence of heart murmurs and not provide definitive diagnoses, they can play an important role in community screening. This allows people with potential heart problems to be referred to specialists promptly, which is valuable, especially in underprivileged areas with limited medical resources.

Valvular heart disease

VHD is a prevalent condition associated with high mortality rates worldwide [120]. Early screening and follow-up are crucial for managing VHD as most patients remain asymptomatic until the advanced stage, resulting in poor prognosis without timely intervention. While echocardiography is the current gold standard for VHD diagnosis [121,122], its cost and requirement for specialized personnel make it impractical for community screening and self-monitoring.

Cardiac auscultation is a simple and cost-effective diagnostic tool for VHD. However, relying solely on auscultation for diagnosis results in low accuracy due to human errors and environmental disturbances [123]. DL techniques have demonstrated superior recognition capabilities compared to humans. Using the Yaseen et al. dataset [11] consisting of 1,000 audio clips of normal heart sounds and 4 VHDs, AS, MS, MR, and MVP, several DL algorithms have been developed to extract heart sound features and train models for VHD diagnosis: First, Yaseen et al. [11] employed MFCCs combined with DWT features as inputs for deep neural network (DNN) classifiers, achieving an accur acy of 92.1%. Second, Ghosh et al. [90] extracted features from the time–frequency matrix of the heart sound recordings and input them into a Deep Layer Kernel Sparse Representation Network classifier, resulting in an overall accuracy of 96.79%. Third, Ding et al. [84] developed a CNN classifier (GoogLeNet) utilizing heart sound time–frequency scalograms based on CWT, achieving an overall accuracy of 97.5%. Fourth, Roy et al. simultaneously constructed a CNN-based inception network, a hybrid CNN–RNN architecture, and an LSTM-based RNN model and compared their performance. The results demonstrated that the CNN-based inception network outperformed the other 2 architectures for heart sound classification [124]. Fifth, Torre-Cruz et al. proposed an approach combining orthogonal nonnegative matrix factorization (ONMF) with CNN architectures in a 3-stage cascade to improve the learning process by identifying optimal ONMF temporal or spectral patterns for accurate VHD detection. This approach increases accuracy by about 45% compared to architectures not using ONMF spectral features [42]. Besides the Yaseen et al. dataset, in the Tromsø study, a population-based prospective study, Waaler et al. collected heart sound recordings and echocardiography results from 2,124 participants. Among them, 408 (19.2%) participants had at least one significant VHD. They developed an LSTM-based RNN model using MFCC-processed heart sound signals as input. This model had excellent discrimination of AS murmurs (area under the curve [AUC] = 0.979), but its performance was mediocre for aortic regurgitation (AUC = 0.634) and MR (AUC = 0.549) [58].

Recently, researchers have worked on constructing automatic “end-to-end” models that do not require signal pre-processing or feature engineering. Also based on the Yaseen et al. dataset [11], Baghel et al. [57] employed a CNN with data augmentation and a Gaussian filter for noise removal, achieving an accuracy of 98.6%. Alkhodari et al. utilized a CNN–RNN model for direct classification using heart sound recordings with one-dimensional wavelet smoothing, resulting in a 99.32% accuracy in a 10-fold cross-validation scheme and 87.31% in external validation based on CinC 2016 dataset [16]. Khan et al. [43] developed a novel Cardi-Net architecture based on a CNN structure to extract discriminative PCG features from the power spectrogram for VHD identification, achieving an accuracy of 98.88%. Jamil and Roy [117] implemented a vision transformer leveraging the self-attention mechanism on PCG signals, further improving the performance in terms of accuracy and computational efficiency. Ma et al. [34] proposed a lightweight CNN architecture, achieving a 92% to 97% reduction in parameters compared to other comparable DL models, with an average accuracy of 98.6% on the Yaseen et al. dataset. In a large-scale study, Chorba et al. [20] trained a CNN model on over 34 h of heart sound recordings from 5,318 patients to detect VHD-related murmurs, yielding a promising performance with a sensitivity of 76.3% and a specificity of 91.4%. Makimoto et al. further improved the interpretability and usability of DL models. They developed a lightweight CNN model to detect severe AS using 1,668 heart sound recordings at 3 auscultation locations from 556 patients. Based on this model, a smartphone application was established, achieving a 95.7% accuracy and 0.93 F1 score. Additionally, they employed Gradient-based Class Activation Maps to identify the specific heart sound features that the DL model focused on when distinguishing the severity of AS [45].

Congenital heart disease

CHD is a prevalent cardiovascular disease in children, affecting approximately 0.8% to 1% of the global population [125]. The most common type is left-to-right shunt CHD, including atrial septal defects (ASDs), ventricular septal defects (VSDs), and patent ductus arteriosus (PDA). This condition can cause chronic volume overload, resulting in HF and pulmonary hypertension (PH) [126,127]. Imaging techniques, including echocardiography, magnetic resonance imaging, and computerized tomography, are crucial for CHD evaluation [128]. However, their limited availability and high costs pose challenges, particularly in underdeveloped regions. The delay in diagnosis can lead to irreversible complications and even death [129].

Auscultation plays a vital role in screening and diagnosing CHD as these patients often present with heart murmurs caused by abnormal blood flow through malformed heart structures [130]. However, the accuracy of this method heavily relies on the physicians’ experience, and not all heart murmurs can be accurately identified [131]. To enhance the diagnostic efficiency of heart sound, DL algorithms have been increasingly employed. Wang et al. developed a temporal attentive pooling-convolutional RNN model for VSD detection using heart sound recordings from 51 patients with VSD and 25 healthy individuals, with a sensitivity of 96.0% and a specificity of 96.7% [36]. Huang et al. converted heart sound recordings from 184 participants, including 46 with VSDs, 50 with ASDs, and 88 with a normal heart structure, into bispectrum signals. These signals were then utilized to train an advanced optical coherence tomography network model for heart sound classification. Remarkably, this model outperformed experienced cardiologists in detecting VSD and ASD with an accuracy of 93.4% and 85.3%, respectively [102]. Liu et al. [21] developed a residual convolution RNN model to detect ASD, VSD, PDA, and combined CHD using 884 heart sound recordings from children with left-to-right shunt CHD, with an overall accuracy of 96.79%. Also, for the task of ASD, VSD, and PDA detection, Wang et al. used a fusion of Mel-Frequency Spectral Coefficients and envelope features of heart sounds as input for a locally concatenated fusion approach combined with a CNN based on coordinate attention (LCACNN). This method achieved classification accuracies of 91.78% and 94.79% on the PhysioNet and private databases, respectively [116]. Hassanuzzaman et al. were the first to apply a transformer model to diagnose CHD. They proposed a DL model that classifies raw PCG signals for CHD diagnosis using a one-dimensional CNN combined with an attention transformer. This model, built on raw PCG data from 484 patients, achieved an accuracy of 0.923 and an AUC of 0.964 [23].

In clinical practice, distinguishing between VSD and bicuspid/tricuspid regurgitation through auscultation can be challenging, as both conditions manifest as systolic murmurs in the mitral and tricuspid areas. Gharehbaghi et al. [91] tackled this issue by training a TGNN model to differentiate VSD from valvular regurgitation and healthy subjects using heart sound recordings from 90 individuals, achieving an accuracy of 86.7%. Furthermore, innocent murmurs are present in approximately 50% of children, leading to many unnecessary referrals to pediatric cardiologists [132]. To address the issue, Gharehbaghi et al. [22] developed a TGNN model capable of distinguishing ASD and VSD from valvular regurgitation and innocent murmur using heart sound recordings from 115 children, resulting in an accuracy of 91.6%.

Heart failure

HF is a global epidemic with high mortality, affecting over 26 million individuals worldwide, and its prevalence continues to rise due to an aging population [133]. Early detection and timely treatment of HF are crucial for long-term prognosis, as the progression of HF can lead to irreversible myocardial remodeling and functional impairment [134]. Current guidelines outline specific conditions for the diagnosis of HF, including typical symptoms and signs, reduced or preserved LVEF, elevated brain natriuretic peptide levels, and the presence of structural heart disease and diastolic dysfunction [135,136]. However, the symptoms or signs may be nonspecific at the early stages of HF [135,136], and echocardiography and blood biomarker tests are unsuitable for screening purposes.

Heart sounds, as a physiological signal generated by myocardial contraction, can provide direct insights into the mechanical dysfunctions of the heart [137]. However, the heart sounds specific to HF, such as gallop rhythm, usually become apparent at the later stages of HF and require sufficient expertise to identify. With DL algorithms, Gao et al. first proposed an HF screening framework based on a gated recurrent unit (GRU) model, distinguishing between the normal subjects, HF with preserved ejection fraction (HFpEF), and HF with reduced ejection fraction (HFrEF) using heart sounds, with an average accuracy of 98.82% [18]. Wang et al. [17] employed CNN and RNN to build a heart sound diagnostic model that accurately differentiated between normal individuals, HFpEF, and HFrEF, achieving an accuracy of 97.64%. Chen et al. further employed a CNN to classify HF patients’ cardiac function according to the New York Heart Association functional classification. They used the CWT to pre-process heart sound signals into spectra as input for the CNN, achieving an accuracy of 94.34% [77].

In addition to left ventricular systolic dysfunction as mentioned above, diastolic dysfunction is also a common type of HF. Yang et al. developed a CNN model to diagnose left ventricular diastolic dysfunction using heart sounds. They applied data augmentation techniques with deep convolutional generative adversarial networks to enhance the model‘s performance, resulting in an accuracy of 98.7% [37]. Similarly, Zheng et al. proposed a PCG transfer learning-based CatBoost model to detect diastolic dysfunction. They utilized multiple domain-specific deep features extracted from PCG spectrograms using pre-trained CNNs, and fed them into CatBoost for classification and performance comparison, yielding an accuracy of 88.2%. [103]

Once HF is diagnosed, accurately classifying the stages of HF is critical for guiding clinical practice. The American Heart Association (AHA)/ American College of Cardiology (ACC) guidelines define 4 stages of HF (stages A, B, C, and D), ranging from developing HF without symptoms to advanced HF [136]. Zheng et al. utilized heart sound recordings from 275 subjects and employed a deep belief network model that incorporated multi-scale (original signal, sub-sequences, and sub-band signals) and multi-domain (time domain, frequency domain, and nonlinear) features. Their approach achieved 74.3% accuracy in automatically HF staging [19].

Coronary artery disease

CAD is a major cause of mortality and morbidity worldwide and substantially burdens the medical system [138]. While coronary angiography is considered the gold standard for CAD diagnosis, its invasive nature and requirement for specialized catheterization laboratories restrict its availability. ECG is another commonly used diagnostic tool, but it has limitations in terms of sensitivity, particularly in stable and asymptomatic patients, and its accuracy highly depends on the expertise of the interpreting physicians [139]. Consequently, there may be a considerable number of undiagnosed CAD cases in underdeveloped regions.

Previous studies have shown that turbulence in stenotic coronary arteries can produce faint high-frequency murmurs [140–142]. However, these faint murmurs are often not discernible during auscultation, and recognizable changes in heart sounds typically occur only after the development of severe structural complications, such as papillary muscle dysfunction, septal perforation, or ventricular dilatation [1]. Given that machine recording can capture faint murmurs, a DL-based diagnostic model shows promise for CAD detection. In 2020, Li et al. developed a CAD detection model using heart sounds. They extracted 110 multi-domain features and MFCCs from the heart sound recordings of 175 subjects. The fusion framework, combining selected multi-domain and DL features, served as input for a CNN classifier, achieving an accuracy of 90.43% [104]. In 2021, Li et al. further improved their approach by developing a multi-input CNN framework that integrated time, frequency, and time–frequency domain deep features from simultaneous ECG and PCG signals of 195 subjects for CAD detection. The model, which combined multi-domain deep features from the 2 modalities, showed high performance in CAD identification, with an accuracy of 96.51% [59]. To address the challenge of limited sample size for training CNN models, Pathak et al. explored transfer learning for CAD detection using heart sounds. They employed a CNN pre-trained on the ImageNet database, consisting of 1 million training images, and transferred its feature representation for CAD detection. Multiple kernel learning was then used to fuse the embeddings of the CNN with handcrafted features, including the heat map of Synchrosqueezing Transform and time-varying Shannon and Renyi Entropy in sub-bands of Synchrosqueezing Transform. Despite having only 40 CAD and 40 normal subjects’ heart sound data, their diagnostic model achieved an accuracy of 89.25% [143]. Ainiwaer et al. recorded heart sounds from 319 patients who underwent coronary angiography, of whom 201 were diagnosed with CAD and 118 were not. They employed state-of-the-art DL architectures (VGG-16, 1D CNN, and ResNet18) to build models for identifying CAD. They found that VGG-16 demonstrated the highest performance, achieving an AUC of 0.834 and outperforming ResNet-18 and CNN-7, which had AUCs of 0.755 and 0.652, respectively [24].

Rheumatic heart disease

RHD remains a significant public health issue in developing countries, impacting a minimum of 33 million individuals and contributing to at least 345,000 deaths annually [144]. RHD is caused by an abnormal immune response to beta-hemolytic streptococcal pharyngitis infection and primarily affects the mitral valve. Typically, echocardiography is used to diagnose RHD by evaluating valve morphology and severity of valve dysfunction [145]. However, given the high prevalence of RHD in underdeveloped regions, there is an urgent need to develop a cost-effective screening method for RHD.

The RHD-related damage on the valves disrupts the normal blood flow in the heart chambers and causes murmurs, which presents a possibility of creating a DL-based model for RHD detection using heart sounds. In 2020, Asmare et al. collected 33,453 heart sound clips from 124 RHD patients and 46 healthy individuals. They trained a CNN model using the Mel Spectro-temporal representation of un-segmented PCG, achieving an overall accuracy of 96.1% with 94.0% sensitivity and 98.1% specificity [47].

Compared to intervention after RHD has developed, early detection of subclinical RHD in susceptible populations and providing penicillin for prophylaxis may be a more cost-effective strategy for individuals and healthcare systems [146]. Based on this principle, Ali et al. proposed a study plan to recruit 1,700 children (5 to 15 years) from underprivileged schools in Pakistan and collect clinical data, including heart sound recordings and echocardiograms. They aimed to train a DNN to automatically identify patients with subclinical RHD and definite RHD [25]. This study is currently ongoing, and we look forward to its results.