Valvular heart disease (VHD) is a condition in which the valves of the mitral, tricuspid, aortic and pulmonary valves become diseased due to rheumatic fever, mucus degeneration, degenerative changes, congenital malformations, ischemic necrosis, infection or trauma, which affects the normal flow of blood and thus causes abnormal heart function [1]. Approximately 2 million people in China suffer from VHD, and 150,000 new cases of VHD are diagnosed each year [2]. Congenital heart disease (CHD) is defined as a gross structural abnormality of the heart or great vessels [3], common diseases in this category include atrial septal defects (ASDs), patent foramen ovale, ventricular septal defects (VSDs), and patent ductus arteriosus (PDA).

Although imaging tools are the primary methods for diagnosing VHD and CHD, physical examination, which includes cardiac auscultation, is a screening tool for VHD and CHD. Auscultation plays a key role in the diagnosis of VHD and CHD [4, 5, 6]. In the context of analysing heart sound signals, computer-aided detection technology can be a useful and cost-effective tool for acquiring and analysing these signals in a quantitative manner, with the added benefits of speed and efficiency [7].

We performed a narrative literature review, and here, we review the recent progress achieved using machine learning applications with heart sound signals derived from VHD and CHD. We examine the advantages and limitations of using artificial intelligence (AI) techniques in the field of VHD and CHD auscultation and suggest some promising future research directions in this field.

Heart sounds are formed by vibrations caused by cardiovascular activities such as heart contractions, heart valve closures, and ventricular wall compressions. According to the order of occurrence in the cardiac cycle, heart sounds are divided into four components: the first heart sound (S1), the second heart sound (S2), the third heart sound (S3) and the fourth heart sound (S4) [8, 9]. In cardiac physiology, the period between S1 and S2 in the same cardiac cycle is referred to as systole, while the period between S2 and the S1 in the subsequent cycle is referred to as diastole. Heart murmur refers to the abnormal sound produced by the vibration of the ventricular wall, valves or blood vessels due to the turbulence of blood in the heart or blood vessels during systole or diastole, in addition to heart sounds, which are noises with different frequencies, different intensities and longer durations. Fig. 1 shows phonocardiograms (PCGs) of different diseases.

Fig. 1.

Normal phonocardiogram and different heart disease phonocardiograms. Note: The figure shows the phonocardiogram of a normal heart, some valvular heart disease and some congenital heart disease. The red arrow points to the murmur. S1: First Heart Sound; S2: Second Heart Sound.

Cardiac auscultation is a qualitative method for assessing heart sounds, heart rate, pericardial friction sounds, etc. The advent of the digital stethoscope, augmented with analytical software, has revolutionized its utility, enabling objective and quantitative assessments of cardiac function in a clinical setting. Automated heart sound analysis in clinical applications usually consists of three steps: pre-processing, segmentation and classification. Pre-processing includes denoising, down-sampling, and normalizing data; Segmentation includes audio cutting and feature extraction; and classification model construction includes network building and model training.

Traditional heart sound classification algorithms require that the feature extraction operators be manually set (Fig. 2), and such methods generally lack model generalizability and have limitations in terms of nonlinear data feature extraction. In recent years, scholars have proposed transforming the original heart sound signal into a two-dimensional heart sound time-frequency map with some transformations, such as the short-time Fourier transform, wavelet transform, and mel-frequency cepstral coefficients (MFCCs) [5, 22], and training deep convolutional networks in the frequency domain for classification purposes [23, 24].

Fig. 2.

Structural diagram of the heart sound identification process. Note: HSMM, Hilbert transform, hidden semi-Markov model; CLSTM, convolutional long short-term; CNN, convolutional neural network; MFCC, Meier spectral coefficient; GAP, global average pooling; MLP, multilayer perceptron; LR, logistic regression; CNN, convolutional neural network; RNN, recurrent neural network; VSM, vector space model.

Recently, the transformer and multilayer perceptron (MLP) techniques in deep learning have attracted the attention of many researchers. Traditional convolutional neural networks (CNNs) can automatically extract local information from data by conducting local weight sharing among the convolutional kernels; however, these models have difficulty accessing global information. Analyses of the second-order spectral feature maps of heart sound data have shown that in addition to certain local saliency characteristics exhibited by heart sound signals, global distribution characteristics are crucial for heart sound classification. An attempt has been made to design a hybrid model by using an MLP algorithm, global average pooling (GAP), and the convolutional technique instead of self-attention when constructing the training model [22]. The network was divided into three parts with different perceptual abilities, i.e., a global perceptron (GP), a partition perceptron (PP), and a local perceptron (LP), to classify the heart sound signals in all directions. Deploying deep learning algorithms to the cloud is a major trend in future research.

The scarcity of heart sound data, particularly the unavailability of publicly accessible and high-quality heart sound databases, poses a significant challenge to the development and evaluation of AI auscultation algorithms, intelligent heart sound diagnosis and analysis technology, and auscultation screening applications [20]. A review of the commonly used heart sound databases that are available on the internet is presented in Table 1. Current heart sound databases have low applicability due to differences in electronic stethoscopes, timing, location, etc., and establishing a standardized heart sound database is the cornerstone of future intelligent heart sound research.

An electronic stethoscope is an important tool for achieving AI-based heart sound diagnosis. The stethoscope is an instrument that is commonly used by all doctors. While its invention dates as far back as 1816, several technological evolutions have occurred over the past 200 years, the latest of which is the electronic stethoscope, which was first developed in the early 1990s [25]. Since then, electronic stethoscopes with various functions have been proposed. Table 2 shows the advantages and disadvantages of different stethoscopes in terms of their recording frequencies, communication techniques, data losses, filtering techniques and environmental filtering techniques [26, 27, 28, 29].

The high incidence of CHD and the dangers leading to poor prognosis are widely recognized. For young children, early detection and treatment are important to reduce the mortality rate of CHD. At present, the CHD diagnosis process is divided into two steps: first, for the initial diagnosis, the doctor makes a preliminary judgement on whether the patient has CHD through cardiac auscultation; then, the initial diagnosis is confirmed by echocardiography in suspected cases. Most patients with CHD are not diagnosed in the early stages of life due to the lack of specific symptoms [30]. This prevents the infant from receiving timely and effective surgical repair or palliative care. Although echocardiography is the gold standard for verifying CHD cases, it usually takes more than 10 minutes to complete [30]. Therefore, in resource-limited areas, it is impractical to perform an echocardiogram on every screening subject. In many areas, community screening for congenital heart disease is often performed by auscultation [31]. With the development of imaging technology, an increasing number of physicians are losing the skill of auscultation [32], and in the absence of symptoms of congenital heart disease, physicians will not perform ultrasound and other tests on patients. In addition to these reasons, the unequal distribution of medical resources is also an important one, thus creating a paradox: the number of young children who need to be screened is large, but few doctors have the required clinical auscultation experience [33]. Some AI-based auscultation applications for CHD diagnosis are summarized in Table 3 (Ref. [4, 5, 21, 22, 23, 24, 34, 35, 36, 37, 38, 39, 40]).

VHD is usually a slowly progressive, chronic disease that may be asymptomatic in its initial stages. The collected data have repeatedly shown that most patients are diagnosed with advanced-stage disease when they are symptomatic or have complications (e.g., ejection dysfunction). Several factors may lead to the delayed diagnosis of VHD, including patients’ inadequate knowledge of the condition and clinicians’ underutilization of cardiac auscultation. Even with experienced clinicians, the sensitivity (up to 43%) and specificity (69%) of physician auscultation for the diagnosis of significant VHD are inadequate [6]. Digital stethoscopes improve murmur detection by converting sounds into electronic signals that can be further amplified, filtered and digitized [50, 51] (Table 4, Ref. [12, 13, 22, 31, 35, 52, 53, 54, 55, 56, 57]).

Thompson et al. [52] selected 3180 heart sound recordings from 603 clinic visits from the Johns Hopkins Cardiac Auscultatory Recording Database. The detection sensitivity and specificity of patients with pathological murmurs were 93% and 81%, respectively, with an accuracy of 88%. However, data that were considered “noisy” or lacking audible murmurs were removed prior to testing, which artificially improved the performance of the algorithm compared to that achieved in real-world settings [52]. In addition to diagnosing valvular disease, computer-assisted auscultation appears to be a relevant support tool for detecting pathological murmurs and appropriately referring patients for further evaluation (93% referral sensitivity and 79% specificity) according to Watrous et al. [58]. The performance varied according to the deterministic measurements of the algorithm, patient ages, heart rates, murmur intensities, and chest recording locations [52]. In a separate investigation [59], Gharehbaghi et al. [22] employed a combination of two deep learning methods, static and dynamic time-varying neural networks, to analyse phonocardiogram (PCG) data. The model was applied to evaluate 140 children with congenital heart disease (CHD) and 50 elderly patients with aortic stenosis (AS), achieving an accuracy of 84.2% and a sensitivity of 82.8%.

Although AI-based cardiac auscultation can help in the diagnosis of VHD, obesity and diseases that affect auscultation (e.g., chronic lung disease) may affect the quality of the obtained sound, leading to inaccurate results. However, because screening large populations is much less expensive than using echocardiography data, this screening process reduces the need for trained health professionals and does not require specialized health care facilities.

Some limitations still need to be addressed before the technology may be used more widely: First, algorithms are often sensitive to the type of stethoscope used and the quality and range of data obtained, and the same algorithm often produces different results for the interpretation of signals obtained from different stethoscopes [20]. In addition, AI-based stethoscopy algorithms should be conducted in collaboration between researchers and medical experts to avoid research compartmentalization [60]. Most importantly, intelligent auscultation should ultimately be used for clinical purposes, yet most of the existing studies have focused on theoretical algorithms rather than practical applications [32]. Another important issue is the lack of a common, authoritative and comprehensive database to compare algorithms and address data imbalances, as each study is relatively independent and there are few systematic and objective evaluations of acquisition environments, parameters and methods [61, 62]. The field will also involve the concept of ethics, as the black-box nature of AI methods leads to unexplainable algorithms without sufficient theory to support their widespread use in clinical settings [63], and one of the biggest challenges is the decreasing frequency of stethoscope use in actual clinical practice, many imaging tests having long since replaced acoustically driven stethoscopes [32].

In terms of recent research, AI-based methods have rarely been applied in clinical settings, and because AI lacks the human-like ability to think about and explore different diseases, AI-based approaches cannot yet replace clinicians and independently complete treatments. The following are future research directions in this field: PCG data can be applied when differentiating between innocent and pathological murmurs is difficult. In such cases, the use of PCGs may increase or decrease the level of suspicion and prompt further investigation or reassurance [19]. In settings with limited access to diagnostic tools, PCG signals may be used to confirm the clinical presentation of VHD for referral to centres with more advanced diagnostic capabilities, rather than for screening purposes [27]. The establishment of high-quality heart sound databases for multiple cardiac diseases and the creation of uniform standards for this purpose are important directions for the future development of this field [52]. Crucially, establishing a unified heart sound processing scheme and solving the problem of interpretability of intelligent models is one of the biggest problems in translating intelligent auscultation into clinical applications.

Smart VHD and CHD auscultation has been used initially in some studies with good results, but currently faces some problems that need to be solved by conducting more studies in the future.

AI, artificial intelligence; ASD, atrial septal defect; VSD, ventricular septal defect; PDA, patent ductus arteriosus; VHD, valvular heart disease; CHD, congenital heart disease; PCG, phonocardiogram; HSMM, Hilbert transform, hidden semi-Markov model; CLSTM, convolutional long short-term; CNN, convolutional neural network; EEG, electroencephalogram; ECG, electrocardiogram; EMG, electromyography; MFCC, Meier spectral coefficient; GAP, global average pooling; GP, global perceptron; PP, partition perceptron; LP, local perceptron; ANC, active noise cancellation; LED, light emitting diode; HZ, hertz; HMM, Mel-frequency cepstral coefficients-hidden Markov model; STGNN, spatial-temporal graph neural network; SVM, support vector machine; XGBoost, extreme gradient boosting; DNN, dynamic neural network; KNN, K-nearest neighbours; LPCC, linear prediction cepstral coefficient; LS-SVM, least-squares support vector machine; CNN, convolutional neural networks; ACC, accuracy; ROC, receiver operating characteristic; AS, aortic stenosis; PS, pulmonary stenosis; MR, mitral regurgitation.

Among the authors in the list, XM and YTM designed the research study and revised it critically for important intellectual content. AA and KK searched and organized the literature, was the main drafter of the manuscript and critically revised the important content. LQ drafted the content of the forms and participated in revising important content of the manuscript. RR checked the fluency of the language and contributed to the manuscript design. AA and RR participated in the collation and analysis of the literature and provided advice on revising the structure of the manuscript. KK and LQ assisted in literature retrieval and participated in revising important content of the manuscript. XM contributed to funding acquisition. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Not applicable.

Not applicable.

This research was supported by Key Research and Development Task of Xinjiang Uygur Autonomous Region Research on Key Technologies and Optimization Strategies for Individualized Precision Diagnosis and Treatment System of Cardiovascular Diseases (2022B03022-3) and Grant 2020XS13 from Tianshan Cedar Program.

The authors declare no conflict of interest.