Asthma is one of the most common chronic pediatric conditions in the world, affecting almost 10% of children and causing significant morbidity and mortality [1], as well as putting a considerable burden on health care systems [1]. It is the third leading cause of admission among children aged <15 years [2]. In total, 7% of children with asthma (aged <18 years) require hospital admission each year due to exacerbation episodes, 47% of children have sleep disturbances or show limitation in daily activities [3], and 25% of children report increased school absences [4]. Mismanagement of the condition can cause long-lasting pulmonary sequelae with decreased lung function persisting into adulthood [5].

Asthma management is often delegated to patients and caregivers, who estimate treatment requirements according to their subjective understanding of symptom severity. The presence of various inevitable confounding factors may lead to poorly adapted and inequitable treatment. Although a written, symptom-based action plan for patients and caregivers is recommended to improve asthma control and avoid hospitalization, it is not sufficient to significantly reduce the risk for subsequent pediatric emergency department visits [6]. Careful monitoring and management of acute exacerbation episodes and long-term treatment are therefore essential to decrease morbidity and mortality and to improve the quality of life of children with asthma and their families.

Since the invention of the stethoscope by René Laennec in 1816, pulmonary auscultation has become a fundamental part of routine physical examination. However, the interpretation of sounds is ultimately limited by the human capacity for auditory perception and cognitive pattern detection. Recent efforts to digitalize auscultation audio allow the recording of sounds that can then be exploited by various software to assist interpretation [7]. A particular advantage is that recordings capture a higher resolution and wider range of frequencies, which are not usually perceptible to the average (and especially, the aging) human ear (ie, outside 20-20,000 Hz). The digital information obtained from the sounds recorded by electronic auscultation can, in turn, inform computer software, aiming to overcome standard lung auscultation limitations by removing subjectivity, standardizing interpretation through objective, nuanced acoustic pattern recognition, and automating the identification of abnormal auscultations indicating respiratory diseases. Recent advances in signal processing and computational efficiency offer the promise of accessible high-performance analysis for many physiological measurements [7]. In particular, deep learning (commonly referred to as artificial intelligence [AI]) opens up new perspectives to improve diagnostic accuracy and enable the automated monitoring of respiratory diseases by nonphysicians [8,9]. Studies showing this potential have mostly focused on adults [10]. The few studies that focused on pediatric auscultation [11-13] highlighted the high potential impact of standardizing the identification and assessment of obstructive pulmonary sounds, which are known reliable estimators of the severity and treatment response in asthma and wheezing disorders. Promising results have emerged in the automated detection of wheezing disorders and pneumonia [12,13]. Even early iterations in 2008 showed automated recognition of wheezing to be superior to traditional lung auscultation performed by untrained or trained medical staff in outpatient or intensive care units [14]. Beyond improved diagnosis and management, there is the potential to decentralize lung auscultation to home-based monitoring and remote care for earlier detection of asthma exacerbations, potentially reducing morbidity and health care costs [15].

The purpose of this scoping review is to assess the state of the art of the research conducted in AI applied to digital lung sound analysis in children with asthma and provide a balanced assessment of its strengths, weaknesses, opportunities, and threats (SWOT). To our knowledge, this is the first such review and critical assessment on the topic.

On the basis of the Population targeted, Concept under investigation, Context parameters framework [16], the review question driving the search strategy was “How has AI-driven automated lung sound analysis been used to detect or classify asthma and wheezing disorders in children over the past 23 years (2000-2023), considering its strengths, weaknesses, opportunities, and threats?” (Multimedia Appendix 1).

The review complies with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [17]. An electronic literature search was performed independently by 2 of the authors with expertise in pediatric emergency medicine and pulmonology (JNS and IR-M). The search was initially focused on academic literature in PubMed, MEDLINE Ovid, Embase, and Web of Science databases, with the date range restricted between January 1, 2000, and May 23. In addition, to broaden the scope, gray literature was sought using Google Scholar with the search limited to the first 10 pages of results (ie, first 250 occurrences) [18,19]. Search terms were chosen to filter out articles on AI-based lung sound analysis for the detection of asthma in humans. The following medical subject headings were explored and connected by Boolean operators with adjustments in syntax adapted to individual databases: artificial intelligence; machine learning; deep learning; neural networks, computer; respiratory sounds; asthma; wheezing disorder. The articles retrieved were further refined for pediatric studies using the following keywords: “adolescent” OR “children” OR “child, preschool” [20] OR “infant” (Table 1).

Identified articles were submitted to inclusion criteria for eligibility: articles should have described automated lung sound analysis in children from birth to 18 years of age with asthma (Textbox 1). Articles should have been written in English. Articles were excluded if the study was conducted on animals. Studies concerning cough rather than lung sounds were excluded. Comments on articles and unpublished work were not considered suitable and were excluded from the review.

Data from the selected articles obtained through academic literature search were extracted by one of the authors (JNS) in the following structured format: publication reference; number of participants and conditions when available, age, number of recordings; study design; number of classes; methodology; recording device; AI model architecture used; findings and performance (metrics) of the algorithm; advantages and limitations of the study; and sharing algorithm code or data as open source. As for the articles from the gray literature, they are provided in the following format: publication reference; article type; and a brief summary of the article.

Quality assessment of selected studies from the academic literature search was performed by JNS using the Checklist for the Assessment of Medical Artificial Intelligence (ChAMAI) [21] and then reviewed by LL. An agreement was reached through discussion. The ChAMAI is a 30-item practical assessment-oriented checklist that is recommended to be reported for medical AI studies to encourage reproducible results and evaluation. It aims to distinguish high-quality medical AI studies from the simple application of AI techniques to medical data. The checklist is organized into six dimensions according to the Cross Industry Standard Process for Data Mining methodology [22]: (1) problem understanding, (2) data understanding, (3) data preparation, (4) modeling, (5) validation, and (6) deployment. Each item can be rated as OK (adequately addressed), minor revision (mR; sufficient but requiring mR to improve), or major revision (MR; insufficiently addressed and necessarily requiring MR or justifying rejection of the manuscript). A total of 20 (67%) of the 30 high-priority items can be assigned a value of 2, 1, and 0 depending on their respective rating (OK, mR, or MR). For the remaining 10 (30%) low-priority items, the values are halved (ie, 1, 0.5, and 0). The maximum score is 50 points. Study quality was classified as low (0-19.5), medium (20-34.5), or high (35-50) [23].

A SWOT analysis categorizes factors describing a research subject into 4 distinct groups: strengths, weaknesses, opportunities, and threats. Strengths and weaknesses are internal factors controlled or influenced by the entity itself, such as its resources, capabilities, and organizational structure, while opportunities and threats are external factors beyond the entity’s direct control that can significantly impact its activity [24]. Although primarily used in organizational studies [25], we used SWOT as a structured analysis framework to conduct a balanced assessment of the SWOT associated with AI-assisted lung auscultation for pediatric asthma, drawing from the academic literature retrieved in this scoping review.

The search strategy selected 82 articles from academic database searches, of which 64 (78%) were only screened, and 18 (22%) were assessed for full-text eligibility. In addition, 250 records were retrieved from gray literature search, with 183 (73%) assessed for eligibility, resulting in a total of 201 unique records (Figure 1). Of the 46 articles not included at the screening stage, 9 (20%) pertained to AI-based asthma or wheezing recognition algorithms using cough sounds [26-34]. Of the 18 studies, 11 (61%) identified through the academic literature search were excluded due to ineligible study design (n=4, 36% in adults; n=4, 36% not relying on AI or lung sounds; and n=3, 27% other study types). Among them were 3 quasi-experimental studies [35-37] focused on computerized wheeze detection through respiratory spectrum analysis in children with asthma, not involving the support of AI algorithms. Along with records from the gray literature search, a total of 9% (7/82) and 3.6% (9/250) of studies were respectively included in the scoping review.

The characteristics of the 7 studies included from the academic literature search are presented in Table 2. They were conducted between 2015, and 2022, across 8 different countries (Croatia, Poland, Australia, China, Singapore, the Republic of Korea, and Russia). With the exception of the study by Mazić et al [38], which was performed in a general hospital, all AI models were implemented in children’s hospitals, either in respiratory clinics or inpatient wards. Study participants varied in number, ranging from 16 to 112. The participant count was not specified in 1 study by Gelman et al [39], involving individuals aged 28 days to 18 years. The design of the studies was also variable, with 4 quasi-experimental studies and 3 prospective cohort studies. Lung sounds were collected using a variety of custom-built [38,40] and commercial (StethoMe, Littmann 3200, CliniCloud, Yunting model II, and Jabes) digital stethoscopes (including a mix of various stethoscopes in 1 of the studies) [11,12,41,42], or with built-in or external cell phone microphones [39]. Although most of the studies used patients’ lung sounds under realistic conditions in their AI models, the study by Cheng et al [40] also incorporated breath sounds from a commercially available database.

Among the identified 7 studies, 4 (57%) [12,39,40,42] developed customized models for their classification tasks, while 3 (43%) [11,38,41] chose a preexisting ready-to-use model (ie, StethoMe AI, and an AI algorithm developed by Tuoxiao Intelligent Technology Company, Shanghai, China). Various model architectures were used for respiratory sound classification, such as support vector machines, neural networks (NNs), and convolutional NNs (Table 2). The algorithms were used for automatic classification of wheezes, crackles, and normal breath sounds. The performance measures varied across studies, with accuracies ranging from 57.7% to 97.7%, sensitivities ranging from 78.2% to 91%, specificities ranging from 82.2% to 97%, and F₁-scores ranging from 66.4% to 87.2% (Table 2). None of the studies shared the algorithm code or data with the community, making reproducibility impossible.

Of the 9 records from gray literature search, the final sample consisted of 6 (67%) conference papers [44-49], 2 (22%) review articles [50,51], and 1 (11%) research article [52] (Multimedia Appendix 2 and Multimedia Appendix 3 [44-53]). Among the total of 201 articles assessed for eligibility across all literature resources, 120 (59.7%) were partially or fully composed of public data set repositories or multimedia resources of breath sounds to train and benchmark their AI-based lung sound algorithms or discuss their relevance in this field. These data sets and their primary features are detailed in Multimedia Appendix 4 [43,53-71]. The International Conference on Biomedical and Health Informatics 2017 Respiratory Sound Database and the Respiratory Acoustic Laboratory Environment represented the 2 most common data sets, used in 61 (50.8%) of 120 studies.

Table 3 summarizes the quality assessment scores of each dimension of the ChAMAI and the total score in each study. The average score of the included studies was 20.8 (SD 6.7), indicating overall medium quality within the lower range. All studies exhibited scores ranging from 13.5 to 31.5 (low to medium quality). Multimedia Appendix 5 [11,12,21,36-40] shows scores for each of the high- and low-priority items.

A visual representation of the SWOT elements related to AI-assisted lung auscultation for pediatric asthma is presented in Figure 2. It provides a comprehensive understanding of the multifaceted impact of integrating AI into pediatric asthma care, highlighting strength and growth potential, while addressing existing challenges and threats. A detailed breakdown of the SWOT identified in the retrieved academic literature is provided in Multimedia Appendix 6 [11,12,38-42].

This scoping review provides an overview of the current state of research on AI models for lung sound analysis in pediatric asthma and wheezing disorders. Only 7 (11%) of the 64 studies qualified for review, highlighting the nascent stage of research in this field. Variations among studies can be attributed to heterogeneity in study populations, design, sample sizes, lung sound recording methods, and models used. Notably, although these models showed promising results in distinguishing between wheezing, crackles, and normal breath sounds, none of the studies conducted external validation to assess the reproducibility and generalizability of the model. Furthermore, the lack of open-source code makes reproducibility or external validation impossible.

In terms of quality, significant disparities were observed among the identified studies, which can be characterized as average quality overall. While problem understanding and modeling were satisfying across all studies, with the exception of the study by Mazić et al [38], there were shortcomings in adequately defining participant inclusion and exclusion criteria. Moreover, descriptions related to data preparation, validation, and deployment exhibited areas that could benefit from improvement. Therefore, further research and refinement of AI algorithms are imperative to elevate their performance and enhance their applicability across various pediatric populations and respiratory conditions. These advances will contribute to achieving more accurate and reliable AI-assisted lung sound analysis in pediatric asthma.

This study has limitations. First, we did not conduct a systematic review. Given the novelty of AI in the field of lung auscultation and the rapid evolution of techniques, a scoping review was preferred to a traditional systematic one for a preliminary assessment of a potentially large and diverse body of literature on this broad topic and its application to the pediatric population [104]. Similarly, a formal meta-analysis was not feasible due to the paucity and heterogeneity of available data. Instead, by synthesizing existing evidence in a charting format, this scoping review provides a comprehensive overview of the research landscape, identifies key themes, and highlights areas to inform the decision-making process for further investigation and potentially guide the design of future systematic reviews and meta-analyses. A second limitation relates to the potential omission of significant research related to the use of AI in children with asthma. This has been minimized by an exhaustive exploration of the main medical literature databases, a meticulous analysis of the references within the selected articles, and a double assessment of the identified studies. However, our inclusion criteria may have led to the unintentional exclusion of information. Third, as AI-powered lung sound analysis is an emerging topic and studies take time to be published, studies or gray literature may have been published after the study inclusion date, potentially leading to incomplete results. Finally, using the ChAMAI tool to assess the quality of the selected studies could be criticized. Nevertheless, to the best of our knowledge, this tool is the most comprehensive and advanced tool to date for such an assessment.

This scoping review and the accompanying SWOT analysis provide a comprehensive assessment of the existing literature on the use of AI-assisted lung auscultation in children aged <18 years with asthma or wheezing disorders. The findings highlight the potential benefits of integrating AI algorithms in objective lung sound evaluation, offering improved diagnostic accuracy, personalized management strategies, and remote monitoring capacities. Addressing the identified limitations, promoting standardization, and addressing ethical considerations will be critical for the successful implementation of AI-based systems in pediatric asthma care, ultimately improving the outcome of affected children. Further research and collaborations are necessary to deploy the full potential of AI in revolutionizing the diagnosis and management of pediatric asthma.