Respiratory sounds are produced by the airflow in the respiratory tract and are divided into two categories: Normal or abnormal sound. Normal respiratory sound is made when there is no pulmonary disorder and consist of tracheal, bronchial, bronchovesicular, and vesicular sounds [33]. Abnormal respiratory sounds are caused by diseases at the lung or bronchus [34]. They can be described by the mechanism of production, location they are detected in, characteristics (such as continuity, range of pitch, timing mostly heard), and acoustic features (Table 1) [35].

Crackles are short, discontinuous, explosive sounds heard during inspiration and sometimes on expiration [36, 37]. Fine crackles are generated by inspiratory opening of small airways and associated with interstitial pneumonia or idiopathic pulmonary fibrosis (IPF), and congestive heart failure (CHF) [38]. Coarse crackles are produced by gas passing through intermittent airway opening and are related to secretory diseases such as chronic bronchitis and pneumonia [39].

Wheezes are generated in the narrowed or obstructed airway [36]. They have high frequency (> 100–5000 Hz) and sinusoidal oscillation in sound analysis [40]. They usually occur in obstructive airway diseases like asthma and chronic obstructive pulmonary disease (COPD) [39]. Rhonchi are induced by the narrowing of airways, caused by the production of secretions, so rhonchi can disappear after coughing (Table 1) [36].

Stridor is a high-pitched, continuous sound produced by turbulent airflow through a narrowed airway of upper respiratory tract [36]. It is usually a sign of airway obstruction that requires prompt intervention. In patients with pleural inflammation such as pleurisy or pleural tumor, a visceral pleura becomes rough, and its friction with the parietal pleura generates crackling sounds, a friction rub (Table 1) [41].

Although respiratory sounds are not difficult for a trained clinician to discern in usual cases, some sounds are ambiguous even for an expert to distinguish accurately. In addition, there are some cases where a mixture of several abnormal breathing sounds can be heard. Novel AI-assisted stethoscope can be useful for these challenging situations.

As mentioned earlier, inherent subjectivity is considered as the largest drawback of auscultation. Many studies have been performed to assess the human’s ability to auscultate and identify respiratory sounds (Table 2).

Hafke-Dys et al. conducted a study comparing the skills of doctors and medical students in the auscultation of respiratory sounds. The pulmonologists performed remarkably better than the other groups and there was no significant difference in the rest of the groups [42]. Melbye et al. proceeded a study assessing the inter-observer variation in pediatricians and doctors for adults when classifying respiratory sounds into detailed or broader categories. The results indicated that descriptions of auscultation sounds in broader terms were more steadily shared between participants compared to more detailed descriptions [43]. Mangione et al. conducted a research assessing auscultatory skills of respiratory sounds among doctors and medical students. On average, trainees of internal medicine and family practice did not show significantly better performance than medical students. On the other hand, pulmonary fellows recorded the highest scores in all categories [44]. Mehmood et al. assessed the auscultatory accuracy of health care professionals working in medical intensive care unit (ICU). The sounds presented were wheezes, stridors, crackles, holosystolic murmur, and hyperdynamic bowel sounds. As expected, attending physicians performed best, followed by residents and subsequently nurses [45]. Andres measured the accuracy of medical students’ auscultation and investigated the efficacy of adding visual representation of sounds to support diagnosis and education. The results showed the potential of sound representation for increasing the accuracy of auscultation [46].

Overall, the studies have shown discrepancies in auscultation ability among doctors (especially in detailed classifications of respiratory sounds), suggesting that they may cause inaccurate diagnosis or incorrect treatment [47]. To reduce the subjective interpretation of sounds and complement the gap of auscultation capabilities between doctors, it would be helpful to establish a system which can record and share auscultated sounds.

Another drawback of auscultation is the impossibility of remote care. When doctors examine patients with a stethoscope, auscultation must be implemented by contacting the stethoscope on the body of patients. Many patients with chronic diseases or limited mobility stay in nursing facilities or at home often without a medical practitioner. Also, the demand of patients in hard-to-reach area for telemedicine is increasing nowadays. However, it is difficult for doctors to examine these patients and auscultation is hardly done. If a stethoscope that is easy to use even for non-specialists is developed using data transmission technology, doctors will be able to check the patient's condition from a distance.

Development of a standardized system to analyze respiratory sounds accurately is required to overcome the subjectivity of human auscultation and the discrepancy in auscultation ability between doctors [8]. Recently, machine learning-based AI techniques are applied mainly by deep learning networks in many areas including chest radiograph or electroencephalography (EEG) [48–50]. These AI techniques enable us to obtain a new approach or more accurate analysis of respiratory sounds [9]. In order to satisfy the requirement, there have been many attempts to develop a new method of classifying and interpreting respiratory sounds automatically using deep learning-based analysis [10, 11]. However, because of the black box type algorithmic property of the deep learning algorithm, there is a certain lack of interpretability of detailed information of the analysis [51]. Though interpretability is an important factor for analysis, it is highly related to technical issues and data dependency. Moreover, it is not clearly defined nor stable yet [51]. For this reason, interpretability will be not covered in this review.

From the machine learning perspective, there are two main parts for respiratory sound analysis. The first is to develop predictive algorithms or models based on well-known machine learning methods (support vector machine [SVM], K-nearest neighbors [KNN], artificial neural network [ANN]) and deep learning architectures (convolutional neural networks [CNN], residual networks [ResNet], long short-term memory [LSTM], gated recurrent unit [GRU]) with multi-layers and the second is to define appropriate features explaining respiratory sound characteristics and extract them (short-time Fourier-transformed [STFT], wavelet transform [WT], Mel-frequency cepstrum coefficient [MFCC], singular spectrum analysis [SSA]) from given data and their ensembles. In this point of view, methods and algorithms for respiratory sound classification and prediction are summarized in more detail below (Table 3).

Fraiwan et al. conducted a study to explore the ability of deep learning algorithms in recognizing pulmonary diseases from recorded lung sounds. After several preprocessing steps (wavelet smoothing, displacement artifact removal, and z-sore normalization), two deep learning network architectures including CNN and bidirectional long short-term memory (biLSTM) units were applied. The resulting algorithm (CNN + biLSTM) achieved the highest accuracy [12]. Chen et al. proceeded with research to overcome the limitations of existing classification methods of lung sounds; artifacts and constrained feature extraction methods. The proposed method using optimized S-transform (OST) and deep ResNets outperformed the ensemble of CNN and the empirical mode decomposition (EMD)-based ANN [13]. Meng et al. combined the wavelet signal similarity with the relative wavelet energy and entropy as the feature vector to extract features of lung sounds. Applying the ANN to this system showed higher accuracy than the methods using SVM and KNN [14]. Hsu et al. applied eight kinds of AI-technique models and conducted a performance comparison between them. GRU-based models outperformed the LSTM-based models, and bidirectional models outperformed unidirectional counterparts. Moreover, adding CNN improved the accuracy of lung sounds analysis [15]. Jung et al. proposed a feature extracting process through the depthwise separable-convolution neural network (DS-CNN) to classify lung sounds accurately. Also, they found that the fusion of the STFT and the MFCC features and DS-CNN achieved a higher accuracy than other methods [16]. Grzywalski et al. compared the efficiency of auscultation of doctors and machine learning-based analysis based on neural networks and proposed that the efficiency could be improved by the implementation of automatic analysis [17]. Kevat et al. showed that a neural network-based AI algorithm detected respiratory sounds with a high accuracy [18]. Aykanat et al. found that CNN and SVM machine learning algorithms can be used to classify lung sounds, but the accuracy decreased as the number of sounds to be compared increased, as with humans [19]. Mondal et al. proposed a feature extraction technique based on EMD improving the performance of lung sound classification and the method was compared with WT, MFCC, and SSA method-based classification systems including ANN, SVM, and Gaussian mixture model (GMM) classifier. The proposed method gives a higher accuracy of 94.16 for an ANN classifier [20]. Altan applied deep belief networks (DBN) algorithm to diagnose early COPD and classify the severity of COPD, and the results showed significantly high accuracy. Since COPD is irreversible when it progresses, early diagnosis is important. In this regard, the results of their studies are groundbreaking and useful [21, 22]. Chamberlain et al. applied SVM with a semi-supervised deep learning algorithm and their algorithm achieved receiver operating characteristic (ROC) curves with a relatively high area under the ROC curve (AUC) [52].

Many studies have been conducted in collaboration with doctors and machine learning experts, and it has become possible to discriminate lung sounds with a considerable level of accuracy. However, there is still a limitation that the analysis becomes less accurate when noises caused by the stethoscope itself, surrounding environment, other organ activities, and so on are mixed among the recorded sounds or when two or more breathing sounds are present at the same time. This should be resolved through additional research in the future [53].

There are several available electronic stethoscopes: Littmann 3100, Stethee pro, Thinklabs one digital amplified medical stethoscope, Littman core digital stethoscope 8490, and StethoMe (Table 4). These digital stethoscopes overcome the low sound levels by electronically intensifying the respiratory sounds. Most importantly, recording of respiratory sounds with a digital stethoscope has allowed and facilitated the study of automatic respiratory sound analysis. Littmann 3100 is one of the most popular electronic stethoscopes, and many studies using respiratory sounds have been conducted with this stethoscope [54, 55]. It can save multiple sounds and transmit the data via Bluetooth transmission. Interestingly, Stethee Pro uses machine learning algorithms to capture and monitor both heart and lung sounds. This stethoscope can amplify the sound up to 96 times and visualize the sound data on the screen. Thinklabs One is the smallest digital stethoscope, and it can be used for personal protective equipment (PPE) auscultation in patients with infectious diseases such as COVID-19. StethoMe was developed for homecare service and installed AI can analyze the abnormality of respiratory sound. It is particularly specialized for monitoring airway diseases including asthma. These digital stethoscopes are continuously developing and becoming more useful for monitoring and diagnosing pulmonary disease.

In addition, recent innovative advances in battery technology, embedded processors with low power consumption, and integrated sensors have made many medical devices wearable and wireless (Table 4). Some studies have applied these techniques to stethoscopes, and the researchers developed the stethoscopes that monitor cardiorespiratory signals through wireless bio-signal acquisition [26, 27]. Certain airway diseases, such as asthma, often get worse at night or early in the morning, so doctors often cannot detect them during the daytime. Just as in the diagnosis of arrhythmia disease, Holter monitoring is used to monitor a patient's heart rate for 24 h, continuous monitoring of respiratory sound through a wearable device in airway disease will be of great help in diagnosis and emergency treatment. Some groups developed water permeable, adhesive, biocompatible acoustic devices for electrophysiological recording [28, 29]. Technologies of recording of sounds clearly and filtering out noises need further improvement, but wearable stethoscopes are expected to be used to diagnose and monitor chronic pulmonary diseases soon.

There are several clinical studies using distal stethoscopes and AI for respiratory analysis. One study showed that CNN can classify chronic disease, non-chronic disease, and healthy groups by automatically analyzing respiratory sounds. In addition, the CNN is able to subcategorize disease group to different types of diseases including COPD, bronchiectasis, pneumonia, and bronchiolitis (Table 5) [56]. Another study adopted the acoustic characteristics of fine crackles to predict honeycombing on chest computed tomography (CT). They concluded that the presence of honeycombing was independently associated with onset time, number of crackles in the inspiratory phase, and F99 value of fine crackles [57].

Many studies related to digital stethoscope and AI analysis of auscultation sound are still currently in progress. As the need to collect and analyze the auscultation sounds of patients in quarantine facilities increases due to the recent COVID-19 crisis, related research is being conducted more actively. Several studies are trying to find a typical pattern of auditory sounds in COVID-19 patients (Table 5). One study plans to evaluate the AI-aided auscultation with automatic classification of respiratory sounds by using StethoMe stethoscope. If these studies are conducted well and AI-equipped stethoscopes can detect wheezing, rhonchi, and crackle accurately, these stethoscopes will be useful in emergency room treatment, medical screening, and telemedicine fields [58]. These smart stethoscopes will be of great help in monitoring patients with chronic pulmonary diseases, and many studies are underway for patients with idiopathic pulmonary fibrosis (IPF) and COPD (Table 5).

Not applicable.

Not applicable.

The authors declare that they have no competing interests.

Not applicable.