Artificial Intelligence Models for Pediatric Lung Sound Analysis: Systematic Review and Meta-Analysis

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.