Review
Abstract
Background: Pediatric respiratory diseases, including asthma and pneumonia, are major causes of morbidity and mortality in children. Auscultation of lung sounds is a key diagnostic tool but is prone to subjective variability. The integration of artificial intelligence (AI) and machine learning (ML) with electronic stethoscopes offers a promising approach for automated and objective lung sound.
Objective: This systematic review and meta-analysis assess the performance of ML models in pediatric lung sound analysis. The study evaluates the methodologies, model performance, and database characteristics while identifying limitations and future directions for clinical implementation.
Methods: A systematic search was conducted in Medline via PubMed, Embase, Web of Science, OVID, and IEEE Xplore for studies published between January 1, 1990, and December 16, 2024. Inclusion criteria are as follows: studies developing ML models for pediatric lung sound classification with a defined database, physician-labeled reference standard, and reported performance metrics. Exclusion criteria are as follows: studies focusing on adults, cardiac auscultation, validation of existing models, or lacking performance metrics. Risk of bias was assessed using a modified Quality Assessment of Diagnostic Accuracy Studies (version 2) framework. Data were extracted on study design, dataset, ML methods, feature extraction, and classification tasks. Bivariate meta-analysis was performed for binary classification tasks, including wheezing and abnormal lung sound detection.
Results: A total of 41 studies met the inclusion criteria. The most common classification task was binary detection of abnormal lung sounds, particularly wheezing. Pooled sensitivity and specificity for wheeze detection were 0.902 (95% CI 0.726-0.970) and 0.955 (95% CI 0.762-0.993), respectively. For abnormal lung sound detection, pooled sensitivity was 0.907 (95% CI 0.816-0.956) and specificity 0.877 (95% CI 0.813-0.921). The most frequently used feature extraction methods were Mel-spectrogram, Mel-frequency cepstral coefficients, and short-time Fourier transform. Convolutional neural networks were the predominant ML model, often combined with recurrent neural networks or residual network architectures. However, high heterogeneity in dataset size, annotation methods, and evaluation criteria were observed. Most studies relied on small, single-center datasets, limiting generalizability.
Conclusions: ML models show high accuracy in pediatric lung sound analysis, but face limitations due to dataset heterogeneity, lack of standard guidelines, and limited external validation. Future research should focus on standardized protocols and the development of large-scale, multicenter datasets to improve model robustness and clinical implementation.
doi:10.2196/66491
Keywords
Introduction
Accurate and timely diagnosis is essential for the treatment of pediatric respiratory illnesses, which remain a leading cause of morbidity and mortality among children worldwide [,]. Auscultation of lung sounds is the most widely used method of respiratory diagnosis due to its simplicity, cost-effectiveness, and safety. However, conventional auscultation requires an in-person encounter, is prone to subjective interpretation, and cannot be shared or reviewed among clinicians, leading to high interobserver variability [].
In recent years, the development of electronic stethoscopes has enabled the digital storage and computational analysis of lung sounds, leading to the creation of large-scale databases [,]. Artificial intelligence (AI)–driven lung sound analysis based on these databases has opened new opportunities to enhance the accuracy and reliability of respiratory disease diagnosis []. Automated AI models for lung sound analysis can facilitate prompt diagnosis and monitor disease progression is particularly useful in remote areas lacking experienced pediatricians or during public health crises, such as the COVID-19 pandemic, when large-scale respiratory screenings are needed [,].
The typical process for developing an AI model to assess pediatric lung sounds includes the following steps: (1) patient recruitment, (2) recording of lung sounds, (3) physician labeling of lung sounds, (4) database creation with separate training and testing sets, (5) feature extraction, (6) machine learning (ML) model development and training, and (7) evaluation of the model with appropriate performance metrics (
Figure 1). These steps could vary between studies, as there are no standardized protocols or guidelines in this area. The current body of research demonstrates high variability in study design, dataset sizes, tasks implemented, and reported outcomes, making it difficult to draw definitive conclusions about the effectiveness of these technologies. Therefore, a comprehensive review of the existing evidence for the application of ML models in pediatric lung sound analysis is necessary.
This systematic review and meta-analysis aim to systematically assess the accuracy of reported ML models on pediatric lung sound analysis, by examining the tasks, methodologies, databases, and evaluation metrics used in original studies. Ultimately, we seek to establish the robustness of current ML models and provide insights for further research.
RQ1: Is it feasible to reliably classify pediatric lung sounds using ML models?
RQ2: How accurately can ML models classify lung sounds into different types of lung sounds or lung pathologies?
Methods
Search Strategy
The literature search was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis for Diagnostic Test Accuracy Studies (PRISMA-DTA) statement []. We followed the PRIMSA-DTA checklist ( []). The main research question of this systematic literature review was to find studies that developed ML models for classifying abnormal lung sounds or pediatric lung pathologies using pediatric lung sound data (
Textbox 1).
Inclusion criteria
Population:
- Pediatric population (age≤18 years).
- Lung sound database specified.
- Manuscript in English.
Intervention:
- Machine learning model classifying at least one type of abnormal lung sounds (wheeze, crackle, stridor, or rhonchi) or lung pathologies (pneumonia, asthma, bronchiolitis, etc).
Comparator:
- Labeling provided by the database.
Outcomes:
- At least one performance metric reported: specificity, sensitivity, accuracy, F1-score, or other specified scoring system.
Study design:
- Original studies on machine learning model development.
Exclusion criteria
Population:
- Adult only or adult majority.
- Database not mentioned.
- Studies focused on cardiac auscultation.
Intervention:
- No machine learning model used.
Comparator:
- No labeling provided.
Outcomes:
- No performance metric provided.
Study design:
- Review articles and validation studies of existing models.
Database Search
We searched Medline via PubMed, Embase, Ovid, Web of Science, and IEEE Xplore using search queries including the following keywords and synonyms such as “infant,” “child, preschool,” “child,” “adolescent,” “pediatrics,” “respiratory sounds,” “auscultation,” “adventitious sound,” “wheeze,” “crackle,” “rale,” “stridor,” “rhonchus,” “machine learning,” “neural network,” and “artificial intelligence.” Exact queries are detailed in . The search covered articles from January 1, 1990, to December 16, 2024.
Eligibility Criteria
Studies were included if they were based on a pediatric population (age ≤18 years), or if more than half of the population were of pediatric age. Lung sound databases—whether public or private—needed to be specified. Only manuscripts in English were considered. Eligible studies applied original ML models for classifying at least one type of abnormal lung sound (wheeze, crackle, stridor, or rhonchi) or lung pathology (eg, pneumonia, asthma, bronchiolitis, etc). Studies without ML algorithms and studies that only performed validation of previously developed models were excluded. Labeling methods needed to be specified for comparators, and at least one performance metric was needed. Abstracts, conference proceedings, and journal papers were included.
Study Selection
Two researchers (JSP and SYP) independently performed abstract and full-text screening. Disagreements were resolved by discussion and mediation by a third researcher (DIS). Articles in languages other than English, duplicate articles from multiple databases, and studies that did not meet the eligibility criteria for population and intervention were excluded during abstract screening. In the full-text review, duplicate studies (eg, conference abstracts later published as full journal papers) were excluded and studies not meeting eligibility criteria due to study design were excluded. EndNote 21 (Clarivate Analytics) was used during this process.
Data Extraction
Data extraction was performed by 2 researchers (JSP and SYP) using a predefined data extraction form in Excel. Article type (conference proceeding or journal paper), first author, year of publication, and journal or conference name were extracted from Endnote. Country of data collection, age of study population, database characteristics, recording device, sample size, train-test split, classification task (eg, wheeze detection, multiclass classification of lung sounds, and asthma classification), feature extraction methods, summary of ML models, and performance metrics were collected. For binary classification, confusion matrix data—true positive (TP), true negative (TN), false positive (FP), and false negative (FN)—were extracted, and if not provided, they were calculated from the sample size and performance metrics.
Meta-Analysis
We conducted meta-analyses for binary classification tasks: wheeze detection and abnormal lung sound detection. Sensitivity and specificity were calculated based on the TP, TN, FP, and FN extracted from the included studies. Sensitivity was defined as the proportion of correctly identified abnormal lung sounds (TP/[TP + FN]), and specificity as the proportion of correctly identified normal lung sounds (TN/[TN + FP]). A pooled analysis of sensitivity and specificity and their CI was conducted with a bivariate metanalysis using a random-effects model []. The heterogeneity among the studies was established by the Zhou and Dendukuri approach, with inconsistency values (I2) greater than 50% being considered as moderate heterogeneity and values above 75% indicating high heterogeneity. Meta-analysis was conducted using the mada package in R software (version 4.3.2; R Foundation for Statistical Computing).
Quality Assessment
Researchers assessed the quality of the study based on the revised tool for the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) []. Study quality was assessed based on 2 factors: risk of bias and applicability. Risk of bias was high if the systematic limitation in the study design or conduct was likely to influence the results. Applicability referred to the extent to which the study population, index test, or reference standard was representative of the review question. Risk of bias was evaluated in 4 domains: patient selection, index test, reference standard, and flow and timing. Applicability was assessed in 3 domains: patient selection, index test, and reference standard. The QUADAS-2 was modified to suit ML-based diagnostic studies [,] (
Table 1).
| Domains | Signaling question | Review question | |
| Risk of bias domains | |||
| Patient selection |
| —a | |
| Index test |
| — | |
| Reference standard |
| — | |
| Flow and timing |
| — | |
| Applicability domains | |||
| Patient selection | — |
| |
| Index test | — |
| |
| Reference standard | — |
| |
aNot applicable.
Results
Systematic search from the 3 databases yielded a total of 2191 articles including conference abstracts, conference proceedings, and journal papers. After removing duplicates and conducting abstract screening, 126 full-text articles were assessed for eligibility. A total of 55 studies were excluded based on criteria such as inappropriate population, incorrect input data for classifiers, missing results, or lack of original ML models. A total of 41 studies were included in the final analysis [-] (
Figure 2).
The included studies spanned from the 1990s (2 studies), through the 2010s (12 studies), to the 2020s (27 studies). A total of 29 studies focused on classifying abnormal lung sounds; of these, 15 aimed for binary classification and 14 pursued multiple classification. In binary classification, wheeze detection was the most researched topic of the studies. Twelve studies targeted classifying the diagnosis, prognosis, or severity of specific lung pathologies, such as pneumonia and asthma. Among these, 5 studies used binary classification, and 7 studies used multiple classification approaches. Pneumonia was the most frequently studied condition, followed by asthma. One study aimed to classify the severity of cystic fibrosis (CF) based on lung sounds (Table 2).
The performance metrics reported in these studies included accuracy, sensitivity, specificity, F1-score, area under the receiver operating characteristic curve, and other unique metrics calculated using a combination of these metrics. Detailed information on the included studies—such as the databases used, sample sizes, tasks, training and test set sizes, feature extraction methods, and ML models—are presented in [-].
| Data type and binary/multiple classes | Task | Included studies | |||
| Lung sound(n=10) | |||||
| Binary (n=15) |
| [,-,,,,,,-,,,] | |||
| Multiple (n=4) |
| [,,,] | |||
| Binary and multiple (n=10) |
| [,-,,,,,] | |||
| Lung pathology (n=12) | |||||
| Binary (n=5) |
| [,,,,] | |||
| Multiple (n=7) |
| [,,,,,,] | |||
aCAS: continuous adventitious sound.
bCF: cystic fibrosis.
cCAP: community-acquired pneumonia.
We extracted 2×2 confusion matrices (TP, FP, TN, and FN) for binary classification studies. Of the 7 studies that examined wheeze detection, 6 provided confusion matrix data (Table 3). For abnormal lung sound detection, 7 of the 15 studies reported confusion matrix data (Table 4).
The pooled sensitivity and specificity of wheeze detection models were 0.902 (95% CI 0.726-0.970) and 0.955 (95% CI 0.762-0.993), respectively. There was moderate between-study heterogeneity (
Figure 3). The pooled sensitivity of abnormal lung sound detection models was 0.907 (95% CI 0.816-0.956) and the specificity was 0.877 (95% CI 0.813-0.921). Low heterogeneity was found (
Figure 4).
| Author (year) | True positive (TP) | False positive (FP) | True negative (TN) | False negative (FN) |
| Forkheim et al (1995) [] | 64 | 38 | 102 | 31 |
| Mazic et al (2015) [] | 268 | 2 | 2572 | 1 |
| Milicevic et al (2016) [] | N/Aa | N/A | N/A | N/A |
| Kuo et al (2021) [] | 30 | 1 | 62 | 2 |
| Kim et al (2022) [] | 83 | 5 | 179 | 20 |
| Nguyen et al (2022) [] | 401 | 142 | 502 | 50 |
| Park et al (2023) [] | 38 | 8 | 46 | 9 |
aNot available.
| Author (year) | True positive (TP) | False positive (FP) | True negative (TN) | False negative (FN) |
| Emmanouilidou et al (2012) [] | N/Aa | N/A | N/A | N/A |
| Khan et al (2017) [] | 71 | 06 | 71 | 06 |
| Emmanouilidou et al (2018) [] | N/A | N/A | N/A | N/A |
| Liu et al (2019) [] | N/A | N/A | N/A | N/A |
| Zhang et al (2022) [] | 295 | 178 | 822 | 94 |
| Li et al (2022) [] | 310 | 103 | 1295 | 22 |
| Zhang et al (2022) [] | N/A | N/A | N/A | N/A |
| Babu et al (2022) [] | N/A | N/A | N/A | N/A |
| Hu et al (2023) [] | 385 | 281 | 759 | 4 |
| Ngo et al (2023) [] | N/A | N/A | N/A | N/A |
| Ngo et al (2023) [] | 328 | 151 | 889 | 61 |
| Pessoa et al (2023) [] | N/A | N/A | N/A | N/A |
| TaghiBeyglou et al (2024) [] | N/A | N/A | N/A | N/A |
| Chowdhury et al (2024) [] | 18 | 3 | 20 | 3 |
| Crisdayanti et al (2024) [] | 4859 | 366 | 5239 | 690 |
aNot available.
A total of 29 distinct databases were identified across the included studies (Table 5). Eight databases labeled lung pathologies including 6 studies that labeled a single lung pathology (pneumonia in 2 studies, asthma in 2 studies, bronchitis in 1 study, and CF in 1 study) and 2 studies that labeled multiple lung pathologies. Seventeen studies labeled abnormal lung sounds as follows: 6 studies with binary labels of wheeze/normal or wheeze/nonwheeze, 3 studies with binary labels of abnormal (or adventitious)/normal, and 8 studies with multiple abnormal lung sound labels (eg, wheeze, crackle, and stridor) labels. Labeling methods or personnel were specified in 19 databases and not specified in 6. Details of each database are available in [-,-, -,-,-].
The most used database was Shanghai Pediatric Respiratory Sound Database, a Chinese database collected from Shanghai University Hospital. The database was used for a challenge in the IEEE BIOCAS 2022 and again in 2023 with an additional test set []. This is the only open-source pediatric lung sound database available. Another public database, the International Conference on Biomedical and Health Informatics 2017 includes only a small number of pediatric participants []. The International Conference on Biomedical and Health Informatics 2017 database was used for pretraining of models or for external validation in some studies.
| Database | Size | Label type | Availability | Studies |
| SPRSounda | 292 subjects | Lung sound | Public | [,,,,,,,,,,] |
| ICBHIb | 49 subjects | Lung sound | Public | [,,,] |
| Liu et al (2019) | 12 recordings | Disorder | Private | [,] |
| CCAP-LSDc | 198 subjects | Disorder | Private | [,] |
| Forkheim et al (1995) | 710 patterns | Lung sound | Private | [] |
| Rietveld et al (1999) | 60 subjects | Disorder | Private | [] |
| Emmanouilidou et al (2012) | 28 recordings | Lung sound | Private | [] |
| Khan et al (2012) | 40 subjects | Disorder | Private | [] |
| Jin et al (2014) | 21 subjects | Lung sound | Private | [] |
| Mazic et al (2015) | 16 subjects | Lung sound | Private | [] |
| Milicevic et al (2016) | 863 samples | Lung sound | Private | [] |
| PERCH | 1157 subjects | Lung sound | Private | [] |
| Khan et al (2017) | 254 sounds | Lung sound | Private | [] |
| Mohamed et al (2018) | 48 patients | Disorder | Private | [] |
| Gouda et al (2019) | 446 sounds | Lung sound | Private | [] |
| Grzywalski et al (2019) | 50 subjects (test set) | Lung sound | Private | [] |
| Liu et al (2019) | 222 subjects | Lung sound | Private | [] |
| Kotb et al (2020) | 116 children | Lung sound | Private | [] |
| Karimizadeh et al (2021) | 209 subjects | Disorder | Private | [] |
| Kuo et al (2021) | 94 subjects | Lung sound | Private | [] |
| Cheng et al (2021) | 73 samples | Lung sound | Private | [] |
| Gelman et al (2022) | 1118 records | Disorder | Private | [] |
| Kim et al (2022) | 76 subjects | Lung sound | Private | [] |
| Nguyen et al (2022) | 1095 recordings | Lung sound | Private | [] |
| DeepBreath | 572 subjects | Disorder | Private | [] |
| Park et al (2023) | 1112 clips | Lung sound | Private | [] |
| R.A.L.E. | >50 recordings | Lung sound | Public | [] |
| Chowdhury et al (2024) | 19 toddlers | Disorder | Private | [] |
| Crisdayanti et al (2024) | 675 patients | Lung sound | Private | [] |
aSPRSound: Shanghai Pediatric Respiratory Sound Database.
bICBHI: International Conference on Biomedical and Health Informatics.
cCCAP-LSD: Cystic Fibrosis Center for Advanced Pediatric Learning and Study Data.
Various feature extraction methods were used (Table 6). The most frequently used method was Mel-frequency cepstral coefficients (MFCC), which was used in 11 studies. MFCC captures the overall shape of the spectral envelope in a compressed form across the Mel scale, a perceptual frequency scale that reflects human auditory sensitivity. The second most widely used method was the (log) Mel-spectrogram, which was applied in 9 studies. This technique provides a visual representation of an audio signal’s frequency content over time, with frequencies converted to the Mel scale. Other Fourier transform-based methods, including short-time Fourier transform (STFT) and spectral features derived from the Fourier spectrum, were used to analyze the frequency content of the signals. Statistical features such as kurtosis, sample entropy, and time-frequency domain characteristics were also used in some studies. Other techniques such as wavelet transformations (continuous and discrete), cochleogram, and time-varying linear predictive coding were also used to extract features from the lung sound recordings.
| Feature extraction | Features | Studies |
| Short-time Fourier transform (STFT) or spectrogram | FTa decomposes a signal into its frequency components, assuming the frequency is stationary throughout the entire time sequence. STFT analyzes how frequencies change over time by dividing the signal into overlapping short frames and applying FT to each frame. This provides a time-frequency representation, assuming stationarity within each frame. A spectrogram is a visual representation of the STFT that displays the magnitude or power as a function of time and frequency, providing a time-frequency representation of the signal. | [,,,,,,,,,] |
| Fourier transform-based spectral feature | Derived from the Fourier spectrum of respiratory sounds, the process involves converting the audio into a time series, applying the FT, and deriving features like spectral bandwidth, centroid, roll-off, and chroma. | [] |
| (log) Mel-spectrogram | A visual representation of an audio signal’s frequency content over time, where the frequencies are converted to the Mel scale, which is based on human perception of pitch. It shows how the energy of different frequency bands evolves over time. The log Mel-spectrogram is computed by taking the logarithm of the Mel-spectrogram values, which helps to compress the dynamic range and emphasize lower energies. | [,,,,,,,,] |
| Mel-frequency cepstral coefficients (MFCC) | MFCC are features derived by applying a Discrete Cosine Transform to the log Mel-spectrogram, capturing key spectral characteristics of an audio signal, often used in speech and audio recognition. | [,,,,,,,,,,] |
| Statistics | A set of features that capture time-domain, frequency-domain, and complexity aspects of breath sounds, including kurtosis, sample entropy, lung sound power ratio, respiratory rate, breathing cycle metrics, peak frequency, and wheezing characteristics, providing key insights for diagnosing respiratory diseases. | [-,,,,] |
| Continuous wavelet transformation (CWT) | Technique to analyze the frequency content of a signal over time using wavelets—small, localized functions that can stretch or compress to capture details at various scales. The CWT provides good time resolution for high-frequency components and good frequency resolution for low-frequency components, making it suitable for analyzing nonstationary signals. | [,] |
| Discrete wavelet transformation | A sampled version of the CWT that selects scales and positions based on powers of 2. The signal is passed through a series of high-pass and low-pass filters, down-sampled by a factor of 2 at each level, and decomposed into wavelet coefficients representing different frequency bands and time scales. | [,] |
| Cochleogram | A time-frequency representation of an audio signal that mimics human auditory processing. Sound signal is passed through bandpass filters based on the Equivalent Rectangular Bandwidth scale, simulating the cochlear frequency resolution, and smoothed and compressed, providing a biologically-inspired representation of the signal’s spectral content. | [] |
| Time-varying linear predictive coding (TVLPC) | A feature extraction technique for nonstationary breath sounds, where the signal is modeled as a linear combination of past samples with time-varying filter coefficients. Unlike traditional LPC, TVLPC allows these coefficients to change over time by expressing them as a linear combination of basis functions, capturing the dynamic spectral characteristics essential for accurately classifying respiratory conditions. | [] |
aFT: Fourier Transform.
A wide range of ML models were used in the included studies for classifying lung sounds and diagnosing respiratory conditions (Table 7). Convolutional neural networks (CNNs) were the most frequently used, featured in 12 studies, in combination with other architectures like recurrent neural networks in 2 studies. Artificial neural networks were also popular, and were used in 10 studies. Residual network (ResNet), a deep CNN architecture with residual connections, was used in 6 studies. Support vector machines (SVMs), another widely used model, was used in 8 studies for optimizing decision boundaries. Other models explored included transformer, hidden Markov models, k-nearest neighbors, ensemble models, and probabilistic classifiers like Naïve Bayes. The choice of model depended on factors such as the nature of the data, the complexity of the classification task, and desired performance metrics.
| Model name | Features | Studies |
| Artificial neural network | Comprised of interconnected nodes arranged in layers. Data flows from the input layer, through hidden layers, to the output layer. Nodes are connected by weighted links that are tuned during training. | [-,-,,,,] |
| Convolutional neural network (CNN) | A type of neural network optimized for processing grid-like data such as images. It uses convolutional layers to extract spatial features and pooling layers to reduce dimensionality. | [,,-,,,,,,,] |
| Residual Network (ResNet) | A deep CNN architecture that introduces residual connections, which skip one or more layers and alleviate the vanishing gradient problem. This enables the training of very deep networks. | [,,,,,] |
| Support vector machine (SVM) | A discriminative classifier that constructs a hyperplane or set of hyperplanes in a high-dimensional space to maximize the margin between classes. It aims to find the optimal decision boundary. | [,-,,] |
| CNN + recurrent neural networks (RNNs) | A hybrid architecture that combines the spatial feature extraction capabilities of CNNs with the temporal modeling abilities of RNNs. This allows for capturing both spatial and sequential patterns in data. | [,] |
| Hidden Markov model | A probabilistic model that assumes the system being modeled is a Markov process with hidden states. It consists of a sequence of state variables and observed variables, with transitions between states governed by probability distributions. | [,] |
| K-nearest neighbors (KNN) | A nonparametric method that classifies data points based on the majority class among their K nearest neighbors in the feature space. It assigns unlabeled examples to the class most common among the K closest labeled instances. | [] |
| Two-level ensemble model | A hierarchical ensemble architecture with 2 levels of model combination. Level 1 separates models by gender to capture data heterogeneity. It trains specialized models for each gender group. Level 2 trains multiple diverse models within each gender group using AutoGluon, an AutoML framework. These multiple models are strategically generated and combined to solve a computational intelligence problem. | [] |
| Naïve Bayes | A probabilistic classifier based on applying Bayes’ theorem with strong independence assumptions between features. It calculates the probability of each class given the feature values and predicts the class with the highest probability. | [] |
| Transformer | A deep learning architecture that uses self-attention mechanisms to process sequential data. It replaces traditional recurrent connections with attention layers that can directly model relationships between all positions in a sequence, allowing parallel processing and better handling of long-range dependencies. | [] |
Quality assessment of the included studies identified several areas of potential bias. Seventeen studies showed a high risk of bias in patient selection, mainly due to insufficient description of the consecutive or random sampling process for the test set. Index test domain was marked as high risk in 11 studies, mainly due to the lack of an independent test set. Studies that used k-fold cross-validation or leave-one-out method, ensuring no overlap of samples from the same subject in both the training and validation sets, were considered low risk. One study included data beyond lung sounds in the classification model, making the index test results dependent on patient characteristics. Studies that did not specify the test set were marked as unknown. Regarding reference standards, labels assigned by at least 2 independent personnel or based on objective measures were considered reliable. A total of 8 studies were marked as high risk, while 7 studies that did not describe their labeling methods were marked as unknown. Most studies did not have issues with flow and timing (
Figure 5 and [-]).
Discussion
Principal Findings
Overview
This systematic literature review of studies of lung sound analysis highlighted the feasibility of AI models in classifying pediatric lung sounds with moderate to high accuracy. There has been a substantial increase in the number of studies in recent years, especially in the 2020s, reflecting advancements in computational power, availability of large datasets, and improvements in ML techniques. This review included 35 studies that varied in classification tasks, sample sizes, age ranges, lung sound databases used, and model evaluation metrics.
Classification Tasks and Challenges
Classification tasks in respiratory sound analysis generally involve distinguishing between different lung pathologies and lung sounds. While the ultimate goal of respiratory physical examination, including auscultation, is to reach a clinical diagnosis, labeling lung sounds based on respiratory diagnoses presents significant challenges, particularly in children. Many respiratory conditions overlap; for example, bronchiolitis and pneumonia can coexist in a child, or an asthma exacerbation may be triggered by a respiratory infection such as viral pneumonia.
Since no single diagnostic test can definitively differentiate between all overlapping conditions, objective criteria are often used to standardize labeling. For instance, pneumonia diagnosis may rely on chest x-ray findings as part of an operational definition. Similarly, when assessing disease severity, physicians often use clinical scoring systems such as the community-acquired pneumonia score for pneumonia []. In some cases, more objective measures are available, such as using forced expiratory volume in one second % predicted to evaluate the severity of CF []. These standardized approaches help ensure more consistent and objective labeling in classification tasks.
Evaluation Methods and Performance Metrics
Although some studies had a separate external validation set or a prospective validation set, in many cases, a separate test set was not specified, limiting the reliability of their results. Furthermore, in many studies, not all the necessary evaluation metrics were specified. According to the Standards for Reporting of Diagnostic Accuracy Studies (STARD) 2015 guidelines [], cross-tabulation of the index test results (or their distribution) by the results of the reference standard should be provided. STARD for AI-centered diagnostic tests (STARD-AI) is still under development [], but a confusion matrix is needed to give a fair evaluation of an AI-centered diagnostic test, as a single performance metric can be influenced by not only the performance of the index test but also the distribution of the samples with and without the diagnosis. If the study failed to provide a confusion matrix, we reconstituted one from the distribution of the test population of the reference standard, and sensitivity and specificity. However, in our review, we have found that many tests did not provide enough metrics to form a confusion matrix.
Dataset Quality and Availability
Significant variation was observed in the quality and quantity of the datasets used across studies. Most studies used small, institution-specific datasets, limiting the generalizability of the findings. Additionally, there was a lack of standardized data collection and annotation protocols. Additionally, there was a lack of standardized data collection and annotation protocols. Particularly, the stethoscopes used to collect data varied widely, limiting their generalizability. The only publicly available pediatric database, Shanghai Pediatric Respiratory Sound Database, came from a single center []. A few multicenter datasets were used in different studies, but the datasets were not available for public use. To develop a more robust and generalizable pediatric lung sound analysis model, a large-scale multicenter pediatric lung sound database is crucial.
Feature Extraction Methods
The most frequently used feature extraction methods in the reviewed studies were MFCC, Mel-spectrogram, and Fourier transform-based methods. Mel-spectrogram and STFT are both time-frequency representations of an audio signal, providing information about the signal’s frequency content over time []. They involve dividing the signal into short frames and applying the Fourier transform to each frame. However, while STFT provides a linear frequency scale, Mel-spectrogram applies a Mel-scale filter bank based on human perception of pitch. This emphasizes the lower frequencies that humans are more sensitive to, making Mel-spectrogram more perceptually relevant. MFCC is derived from the Mel-spectrogram by applying additional processing steps, such as taking the logarithm and applying the discrete cosine transform, resulting in a compressed feature vector that captures the overall shape of the spectral envelope []. In adult-based lung sound classification studies, new features such as Chromagram, representation of the intensity of different pitch classes (chromas) over time, irrespective of the octave, have been explored [].
ML Models
The reviewed studies used a diverse range of ML models for classifying lung sounds and diagnosing respiratory conditions. CNN and its variants, such as CNN combined with recurrent neural network and ResNet, emerged as the most popular choice, with a total of 17 studies adopting these architectures. The prevalence of CNN-based models can be attributed to their ability to automatically learn and extract relevant features from the input data, making them well-suited for processing complex signals such as lung sounds. ResNet, a deep CNN architecture with residual connections, allows for the training of much deeper networks without the vanishing gradient problem, enabling the network to capture complex hierarchical features and dependencies in lung sound data []. Apart from neural network-based models, SVM was also widely used, being used in 6 studies. SVMs are known for their ability to find optimal decision boundaries in high-dimensional feature spaces, making them effective in classifying lung sounds based on extracted features []. While other models such as Hidden Markov Models, k-nearest neighbors, ensemble models, and probabilistic classifiers were explored in some studies, the dominance of CNN-based architectures, particularly ResNet, and SVMs underscores their effectiveness in accurately classifying lung sounds and diagnosing respiratory conditions.
Comparison With Prior Work
There is a limited number of reviews on lung sound analysis using ML [,-]. Compared with prior work in adult populations, this study provides a unique focus on pediatric datasets, where challenges such as overlapping conditions and smaller dataset sizes are more pronounced. While a recent scoping review on pediatric asthma diagnosis using lung sound analysis was published, this study expands the scope by including various lung pathologies and sound classification tasks [].
Strengths
Our systematic review and meta-analysis provide a comprehensive review of ML used in pediatric lung sound analysis, offering an evidence base for future research in this field. The study synthesizes findings from multiple studies, identifying common trends, challenges, and gaps in research. Additionally, this review emphasizes the need for standardized reporting guidelines and the development of multicenter pediatric lung sound datasets.
Limitations
The study has some limitations. First, we have only searched for articles in English. Second, due to the heterogeneity of the classification tasks of the included studies and the lack of essential results in some studies, meta-analysis was possible for only parts of the data. Third, the lack of standard guidelines in AI-centered diagnostic studies undermines the objectivity of quality assessment in our study. To address this limitation, we have made our best efforts to modify the QUADAS-2 to fit AI-centered diagnostics. These weaknesses may be overcome in the future when STARD-AI is published and upcoming studies conform to this guideline [], and when QUADAS-AI is also published for evaluating these studies []. In the meanwhile, since AI in medical imaging is the leading field of AI-based diagnostics, we can draw from existing guidelines on reporting AI in medical imaging [].
Future Directions
While ML models have shown high accuracy in the classification of lung sounds, several challenges remain in implementing the models into clinical practice. First, the black-box nature of ML models limits clinical interpretability. Explainable AI techniques that provide visual insights into the model would be useful for clinical decision-making and sharing information with patients. Second, for the practical use of these models, stethoscopes equipped with real-time analysis models need to be developed. Technical issues must be resolved to implement lung sound classification models in clinical practice and electronic health record systems. Last, the clinical value of these studies has not been shown in any studies. Clinical trials to validate the efficacy of ML models in clinical settings, such as better diagnostic accuracy or faster decision-making in the clinic, should be implemented according to guidelines []. Before these models are used in clinical and household settings, ethical and privacy issues must be addressed [,].
In conclusion, pediatric lung sound analysis can be performed with high accuracy. However, due to the lack of standard guidelines, there is significant heterogeneity in the reported studies. Standardization in this emerging field is necessary. Future research should prioritize robust designs with external validation, detailed descriptions of model development, and comprehensive performance results. With the development of real-time analysis tools that can be deployed in various clinical settings, pediatric lung sound analysis has the potential to improve respiratory care, providing timely and accurate diagnoses, and ultimately enhancing clinical outcomes for pediatric patients.
Acknowledgments
JSP and SYP contributed equally as co-first authors. KK and DIS contributed equally as co-corresponding authors. We would like to thank Bo-Kyeong Kim, medical librarian at the Medical Library of Seoul National University, for her invaluable assistance in conducting the systematic literature search. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (number RS-2024-00343860). No generative AI tools were used in the writing or data extraction of this manuscript.
Data Availability
The data supporting this systematic review are extracted from publicly available studies included in the review.
Authors' Contributions
JSP, SYP, JWM, KK, and DIS were responsible for the concept and design. Data collection and curation were conducted by JSP and SYP. Statistical analysis was performed by JSP, SYP and JWM, while data interpretation was carried out by JSP, SYP, KK, and DIS. JSP, SYP and JWM drafted the manuscript, and KK and DIS revised it for critical content. Supervision was provided by KK and DIS. All authors reviewed and approved the final version of the manuscript.
Conflicts of Interest
None declared.
Multimedia Appendix 1
Preferred Reporting Items for Systematic Reviews and Meta-Analysis for Diagnostic Test Accuracy Studies (PRISMA-DTA) checklist.
DOCX File , 24 KBMultimedia Appendix 3
Characteristics of studies included in this review on pediatric lung sound analysis.
DOCX File , 32 KBMultimedia Appendix 4
Details of databases used by studies included in this review on pediatric lung sound analysis.
DOCX File , 26 KBMultimedia Appendix 5
Quality assessment results of studies included in this review.
DOCX File , 20 KBReferences
- Liu L, Oza S, Hogan D, Chu Y, Perin J, Zhu J, et al. Global, regional, and national causes of under-5 mortality in 2000-15: an updated systematic analysis with implications for the sustainable development goals. Lancet. 2016;388(10063):3027-3035. [FREE Full text] [CrossRef] [Medline]
- Nair H, Simões EAF, Rudan I, Gessner B, Azziz-Baumgartner E, Zhang J, et al. Severe Acute Lower Respiratory Infections Working Group. Global and regional burden of hospital admissions for severe acute lower respiratory infections in young children in 2010: a systematic analysis. Lancet. 2013;381(9875):1380-1390. [FREE Full text] [CrossRef] [Medline]
- Aviles-Solis JC, Vanbelle S, Halvorsen PA, Francis N, Cals JWL, Andreeva EA, et al. International perception of lung sounds: a comparison of classification across some European borders. BMJ Open Respir Res. 2017;4(1):e000250. [FREE Full text] [CrossRef] [Medline]
- Ramanathan A, Zhou L, Marzbanrad F, Roseby R, Tan K, Kevat A, et al. Digital stethoscopes in paediatric medicine. Acta Paediatr. 2019;108(5):814-822. [CrossRef] [Medline]
- Ellington LE, Emmanouilidou D, Elhilali M, Gilman RH, Tielsch JM, Chavez MA, et al. Developing a reference of normal lung sounds in healthy Peruvian children. Lung. 2014;192(5):765-773. [FREE Full text] [CrossRef] [Medline]
- Kim Y, Hyon Y, Jung SS, Lee S, Yoo G, Chung C, et al. Respiratory sound classification for crackles, wheezes, and rhonchi in the clinical field using deep learning. Sci Rep. 2021;11(1):17186. [FREE Full text] [CrossRef] [Medline]
- Zhu H, Lai J, Liu B, Wen Z, Xiong Y, Li H, et al. Automatic pulmonary auscultation grading diagnosis of Coronavirus disease 2019 in China with artificial intelligence algorithms: a cohort study. Comput Methods Programs Biomed. 2022;213:106500. [FREE Full text] [CrossRef] [Medline]
- Lapteva EA, Kharevich ON, Khatsko VV, Voronova NA, Chamko MV, Bezruchko IV, et al. Automated lung sound analysis using the Lungpass platform: a sensitive and specific tool for identifying lower respiratory tract involvement in COVID-19. Eur Respir J. 2021;58(6):2101907. [FREE Full text] [CrossRef] [Medline]
- McInnes MDF, Moher D, Thombs BD, McGrath TA, Bossuyt PM, The PRISMA-DTA Group, et al. Preferred reporting items for a systematic review and meta-analysis of diagnostic test accuracy studies: the PRISMA-DTA statement. JAMA. 2018;319(4):388-396. [CrossRef] [Medline]
- Reitsma JB, Glas AS, Rutjes AW, Scholten RJ, Bossuyt PM, Zwinderman AH. Bivariate analysis of sensitivity and specificity produces informative summary measures in diagnostic reviews. J Clin Epidemiol. 2005;58(10):982-990. [CrossRef] [Medline]
- Whiting PF, Rutjes AW, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, et al. QUADAS-2 Group. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155(8):529-536. [FREE Full text] [CrossRef] [Medline]
- Jayakumar S, Sounderajah V, Normahani P, Harling L, Markar SR, Ashrafian H, et al. Quality assessment standards in artificial intelligence diagnostic accuracy systematic reviews: a meta-research study. NPJ Digit Med. 2022;5(1):11. [FREE Full text] [CrossRef] [Medline]
- Garcia-Mendez JP, Lal A, Herasevich S, Tekin A, Pinevich Y, Lipatov K, et al. Machine learning for automated classification of abnormal lung sounds obtained from public databases: a systematic review. Bioengineering (Basel). 2023;10(10):1155. [FREE Full text] [CrossRef] [Medline]
- Forkheim KE, Scuse D, Pasterkamp H. A comparison of neural network models for wheeze detection. 1995. Presented at: IEEE WESCANEX 95. Communications, Power, and Computing. Conference Proceedings; 1995 May 15-16; Winnipeg, MB, Canada. [CrossRef]
- Rietveld S, Oud M, Dooijes E. Classification of asthmatic breath sounds: preliminary results of the classifying capacity of human examiners versus artificial neural networks. Comput Biomed Res. 1999;32(5):440-448. [CrossRef] [Medline]
- Emmanouilidou D, Patil K, West J, Elhilali M. A multiresolution analysis for detection of abnormal lung sounds. Annu Int Conf IEEE Eng Med Biol Soc. 2012;2012:3139-3142. [FREE Full text] [CrossRef] [Medline]
- Khan SI, Jawarkar NP, Ahmed V. Cell phone based remote early detection of respiratory disorders for rural children using modified stethoscope. 2012. Presented at: International Conference on Communication Systems and Network Technologies; 2012 May 11-13; Rajkot, Gujarat, India.
- Jin F, Sattar F, Goh D. New approaches for spectro-temporal feature extraction with applications to respiratory sound classification. Neurocomputing. 2014;123:362-371. [CrossRef]
- Mazić I, Bonković M, Džaja B. Two-level coarse-to-fine classification algorithm for asthma wheezing recognition in children's respiratory sounds. Biomed Signal Process Control. 2015;21:105-118. [CrossRef]
- Milicevic M, Mazic I, Bonkovic M. Asthmatic wheezes detection - what contributes the most to the role of MFCC in classifiers accuracy? Int J Biol Biomed Eng. 2016;10:176-182. [FREE Full text]
- Khan SI, Ahmed V, Jawarkar NP. Application of signal processing techniques for preliminary detection of adventitious lung sounds in paediatric population using electronic stethoscope. 2017. Presented at: International Conference on Big Data Analytics and Computational Intelligence (ICBDAC); 2017 March 23-25; Chirala, Andhra Pradesh, India.
- Emmanouilidou D, McCollum ED, Park DE, Elhilali M. Computerized lung sound screening for pediatric auscultation in noisy field environments. IEEE Trans Biomed Eng. 2018;65(7):1564-1574. [FREE Full text] [CrossRef] [Medline]
- Mohamed-Rida B, Starkov P, Manzano S, Hugon F, Solà CJ, Gervaix A. Diagnostic performance of an unsupervised electronic stethoscope for community-acquired childhood pneumonia in a paediatric emergency department: a feasibility case-control study. Swiss Med Wkly. 2018;147:2S. [FREE Full text]
- Gouda A, Shehaby S, Diaa N, Abougabal M. Classification techniques for diagnosing respiratory sounds in infants and children. 2019. Presented at: IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC); 2019 January 07-09; Las Vegas, NV, USA.
- Grzywalski T, Piecuch M, Szajek M, Bręborowicz A, Hafke-Dys H, Kociński J, et al. Practical implementation of artificial intelligence algorithms in pulmonary auscultation examination. Eur J Pediatr. 2019;178(6):883-890. [FREE Full text] [CrossRef] [Medline]
- Liu L, Li W, Jiang C. Breath sounds recognition and classification for respiration system diseases. 2019. Presented at: ICC 2019 - 2019 IEEE International Conference on Communications (ICC); 2019 May 20-24; Shanghai, China.
- Liu R, Cai S, Zhang K, Hu N. Detection of adventitious respiratory sounds based on convolutional neural network. 2019. Presented at: International Conference on Intelligent Informatics and Biomedical Sciences (ICIIBMS); 2019 November 21-24; Shanghai, China.
- Kotb MA, Elmahdy HN, Seif El Dein HM, Mostafa FZ, Refaey MA, Rjoob KWY, et al. The machine learned stethoscope provides accurate operator independent diagnosis of chest disease. Med Devices (Auckl). 2020;13:13-22. [FREE Full text] [CrossRef] [Medline]
- Karimizadeh A, Vali M, Modaresi M. Multichannel lung sound analysis to detect severity of lung disease in cystic fibrosis. Biomed Signal Process Control. 2021;64:102266. [CrossRef]
- Kuo H, Lin B, Wang Y, Lin B. Development of automatic wheeze detection algorithm for children with asthma. IEEE Access. 2021;9:126882-126890. [CrossRef]
- Liu L, Li W, Jiang C. Time-varying respiratory feature recognition and classification for respiration health and disease screening in children. 2021. Presented at: IEEE International Conference on Systems, Man, and Cybernetics (SMC); 2021 October 17-20; Melbourne, Australia.
- Cheng ZR, Zhang H, Thomas B, Tan YH, Teoh OH, Pugalenthi A. Assessing the accuracy of artificial intelligence enabled acoustic analytic technology on breath sounds in children. J Med Eng Technol. 2022;46(1):78-84. [CrossRef] [Medline]
- Gelman A, Furman E, Kalinina N, Malinin S, Furman G, Sheludko V, et al. Computer-aided detection of respiratory sounds in bronchial asthma patients based on machine learning method. Sovrem Tekhnologii Med. 2022;14(5):45-51. [FREE Full text] [CrossRef] [Medline]
- Kim BJ, Kim BS, Mun JH, Lim C, Kim K. An accurate deep learning model for wheezing in children using real world data. Sci Rep. 2022;12(1):22465. [FREE Full text] [CrossRef] [Medline]
- Ma W, Deng X, Yang Y, Fang W. An effective lung sound classification system for respiratory disease diagnosis using DenseNet CNN model with sound pre-processing engine. 2022. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2022 October 13-15; Taipei, Taiwan.
- Nguyen TN, Arjoune Y, Schroeder JC, Pillai D, Teach S, Patel S. Machine learning for automated wheeze detection in children. 2022. Presented at: IEEE International Conference on Big Data (Big Data); 2022 December 17-20; Osaka, Japan.
- Zhang Q, Zhang J, Yuan J, Huang H, Zhang Y, Zhang B, et al. SPRSound: Open-source SJTU paediatric respiratory sound database. IEEE Trans Biomed Circuits Syst. 2022;16(5):867-881. [CrossRef] [Medline]
- Li J, Wang X, Wang X, Qiao S, Zhou Y. Improving the resnet-based respiratory sound classification systems with focal loss. 2022. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2022 October 13-15; Taipei, Taiwan.
- Zhang L, Zhu Y, Tu S, Xu L. A feature polymerized based two-level ensemble model for respiratory sound classification. 2022. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2022 October 13-15; Taipei, Taiwan.
- Babu N, Kumari J, Mathew J, Satija U, Mondal A. Multiclass categorisation of respiratory sound signals using neural network. 2022. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2022 October 13; Taipei, Taiwan.
- Heitmann J, Glangetas A, Doenz J, Dervaux J, Shama DM, Garcia DH, et al. with the Pneumoscope Study Group. DeepBreath-automated detection of respiratory pathology from lung auscultation in 572 pediatric outpatients across 5 countries. NPJ Digit Med. 2023;6(1):104. [FREE Full text] [CrossRef] [Medline]
- Hu J, Leow C, Tao S, Goh W, Gao Y. Supervised contrastive pretrained resnet with mixup to enhance respiratory sound classification on imbalanced and limited dataset. 2023. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2023 October 19; Toronto, ON, Canada.
- Huang D, Wang L, Lu H, Wang W. A contrastive embedding-based domain adaptation method for lung sound recognition in children community-acquired pneumonia. 2023. Presented at: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP); 2023 June 04-10; Rhodes Island, Greece.
- Huang D, Wang L, Wang W. A multi-center clinical trial for wireless stethoscope-based diagnosis and prognosis of children community-acquired pneumonia. IEEE Trans Biomed Eng. 2023;70(7):2215-2226. [CrossRef] [Medline]
- Ngo D, Pham L, Phan H, Tran M, Jarchi D. A deep learning architecture with spatio-temporal focusing for detecting respiratory anomalies. 2023. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2023 October 19-21; Toronto, ON, Canada.
- Ngo D, Pham L, Phan H, Tran M, Jarchi D, Kolozali Ş. An inception-residual-based architecture with multi-objective loss for detecting respiratory anomalies. 2023. Presented at: IEEE 25th International Workshop on Multimedia Signal Processing (MMSP); 2023 September 27-29; Poitiers, France.
- Ntalampiras S. Explainable siamese neural network for classifying pediatric respiratory sounds. IEEE J Biomed Health Inform. 2023;27(10):4728-4735. [CrossRef] [Medline]
- Park JS, Kim K, Kim JH, Choi YJ, Kim K, Suh DI. A machine learning approach to the development and prospective evaluation of a pediatric lung sound classification model. Sci Rep. 2023;13(1):1289. [FREE Full text] [CrossRef] [Medline]
- Pessoa D, Petmezas G, Papageorgiou V, Rocha B, Stefanopoulos L, Kilintzis V. Pediatric respiratory sound classification using a dual input deep learning architecture. 2024. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2023 October 19-21; Toronto, ON, Canada.
- TaghiBeyglou B, Assadi A, Elwali A, Yadollahi A. TRespNET: A dual-route exploratory CNN model for pediatric adventitious respiratory sound identification. Biomed Signal Process Control. 2024;93:106170. [CrossRef]
- Chowdhury S, Doulah ABMSU, Rasheduzzaman M, Rafa TS. Pediatric pneumonia diagnosis: integration of a self-assembled digital stethoscope with raspberry Pi and 1D CNN model. 2024. Presented at: International Conference on Advances in Computing, Communication, Electrical, and Smart Systems (iCACCESS); 2024 March 08-09; Dhaka, Bangladesh.
- Crisdayanti IAPA, Nam SW, Jung SK, Kim S. Attention feature fusion network via knowledge propagation for automated respiratory sound classification. IEEE Open J Eng Med Biol. 2024;5:383-392. [FREE Full text] [CrossRef] [Medline]
- Wang F, Yuan X, Bao J, Lam C, Huang G, Chen H. OFGST-Swin: swin transformer utilizing overlap fusion-based generalized S-transform for respiratory cycle classification. IEEE Trans. Instrum. Meas. 2024;73:1-13. [CrossRef]
- Yeh C, Chiu S, Deng X, Fang W. A novel AI-Inspired method and system implementation for detecting and classifying pediatric respiratory sound events. 2024. Presented at: IEEE Biomedical Circuits and Systems Conference (BioCAS); 2024 October 24; Xi'an, China.
- Rocha BM, Filos D, Mendes L, Serbes G, Ulukaya S, Kahya YP, et al. An open access database for the evaluation of respiratory sound classification algorithms. Physiol Meas. 2019;40(3):035001. [CrossRef] [Medline]
- Florin T, Ambroggio L, Lorenz D, Kachelmeyer A, Ruddy R, Kuppermann N, et al. Development and internal validation of a prediction model to risk stratify children with suspected community-acquired pneumonia. Clin Infect Dis. 2021;73(9):e2713-e2721. [FREE Full text] [CrossRef] [Medline]
- Cohen JF, Korevaar DA, Altman DG, Bruns DE, Gatsonis CA, Hooft L, et al. STARD 2015 guidelines for reporting diagnostic accuracy studies: explanation and elaboration. BMJ Open. 2016;6(11):e012799. [FREE Full text] [CrossRef] [Medline]
- Sounderajah V, Ashrafian H, Golub RM, Shetty S, De Fauw J, Hooft L, et al. STARD-AI Steering Committee. Developing a reporting guideline for artificial intelligence-centred diagnostic test accuracy studies: the STARD-AI protocol. BMJ Open. 2021;11(6):e047709. [FREE Full text] [CrossRef] [Medline]
- Boashash B. Time-frequency signal analysis and processing: a comprehensive reference. New York City. Academic Press; 2015.
- Koduru A, Valiveti HB, Budati AK. Feature extraction algorithms to improve the speech emotion recognition rate. Int J Speech Technol. 2020;23(1):45-55. [CrossRef]
- Wanasinghe T, Bandara S, Madusanka S, Meedeniya D, Bandara M, De La Torre Díez I. Lung sound classification with multi-feature integration utilizing lightweight CNN model. IEEE Access. 2024;12:21262-21276. [CrossRef]
- He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. 2016. Presented at: IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2016 June 27:770-778; Las Vegas, NV, USA.
- Grama L, Tuns L, Rusu C. On the Optimization of SVM kernel parameters for improving audio classification accuracy. 2017. Presented at: 14th International Conference on Engineering of Modern Electric Systems (EMES); 2017 June 01; Oradea, Romania.
- Wanasinghe T, Bandara S, Madusanka S, Meedeniya D, Bandara M, De la Torre Díez I. Lung sound classification for respiratory disease identification using deep learning: a survey. Int J Online Eng. 2024;20(10):115-129. [CrossRef]
- Huang D, Huang J, Qiao K, Zhong N, Lu H, Wang W. Deep learning-based lung sound analysis for intelligent stethoscope. Mil Med Res. 2023;10(1):44. [FREE Full text] [CrossRef] [Medline]
- Kim Y, Hyon Y, Lee S, Woo S, Ha T, Chung C. The coming era of a new auscultation system for analyzing respiratory sounds. BMC Pulm Med. 2022;22(1):119. [FREE Full text] [CrossRef] [Medline]
- Ruchonnet-Métrailler I, Siebert JN, Hartley M, Lacroix L. Automated interpretation of lung sounds by deep learning in children with asthma: scoping review and strengths, weaknesses, opportunities, and threats analysis. J Med Internet Res. 2024;26:e53662. [FREE Full text] [CrossRef] [Medline]
- Sounderajah V, Ashrafian H, Rose S, Shah NH, Ghassemi M, Golub R, et al. A quality assessment tool for artificial intelligence-centered diagnostic test accuracy studies: QUADAS-AI. Nat Med. 2021;27(10):1663-1665. [CrossRef] [Medline]
- Mongan J, Moy L, Kahn CE. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers. Radiol Artif Intell. 2020;2(2):e200029. [FREE Full text] [CrossRef] [Medline]
- Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, et al. DECIDE-AI expert group. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28(5):924-933. [CrossRef] [Medline]
- Karimian G, Petelos E, Evers SMAA. The ethical issues of the application of artificial intelligence in healthcare: a systematic scoping review. AI Ethics. 2022;2(4):539-551. [CrossRef]
- Petersson L, Larsson I, Nygren JM, Nilsen P, Neher M, Reed JE, et al. Challenges to implementing artificial intelligence in healthcare: a qualitative interview study with healthcare leaders in Sweden. BMC Health Serv Res. 2022;22(1):850. [FREE Full text] [CrossRef] [Medline]
Abbreviations
| AI: artificial intelligence |
| CF: cystic fibrosis |
| CNN: convolutional neural network |
| FN: false negative |
| FP: false positive |
| HMM: hidden Markov model |
| MFCC: Mel-frequency cepstral coefficients |
| ML: machine learning |
| PRISMA-DTA: Preferred Reporting Items for Systematic Reviews and Meta-Analysis for Diagnostic Test Accuracy Studies |
| QUADAS-2: Quality Assessment of Diagnostic Accuracy Studies (Version 2) |
| ResNet: residual network |
| STARD: Standards for Reporting of Diagnostic Accuracy Studies |
| STFT: short-time Fourier transform |
| SVM: support vector machine |
| TN: true negative |
| TP: true positive |
Edited by A Coristine; submitted 14.09.24; peer-reviewed by I Ruchonnet-Métrailler, D Meedeniya, F Ponzio; comments to author 23.11.24; revised version received 14.02.25; accepted 13.03.25; published 18.04.25.
Copyright©Ji Soo Park, Sa-Yoon Park, Jae Won Moon, Kwangsoo Kim, Dong In Suh. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 18.04.2025.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.