JMIR Publications

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 subjective and prone to variability. The advent of electronic stethoscopes and artificial intelligence offers new opportunities for improving diagnostic accuracy and consistency in lung sound analysis.

Objective:

This systematic review and meta-analysis aim to evaluate the performance of machine learning (ML) models in pediatric lung sound analysis, focusing on the tasks, methodologies, and databases used in existing studies. The review also aims to provide insights into the limitations of the current research and identify future directions.

Methods:

A comprehensive search was conducted in Medline via PubMed, Embase, and IEEE Xplore databases. Studies were included if they developed ML models to classify pediatric lung sounds or respiratory pathologies. Data were extracted on study design, sample size, feature extraction methods, ML models, and performance metrics such as accuracy, sensitivity, and specificity. Meta-analyses were performed for binary classification tasks: wheeze and abnormal lung sound detection.

Results:

A total of 35 studies met the inclusion criteria. The most common task was binary classification of abnormal lung sounds, including wheeze detection. Wheeze detection models demonstrated pooled sensitivity of 0.909 (95% CI: 0.749–0.971) and specificity of 0.959 (95% CI: 0.795–0.993). For abnormal lung sound detection, the pooled sensitivity was 0.923 (95% CI: 0.770–0.977) and specificity was 0.846 (95% CI: 0.760–0.904). Significant heterogeneity was observed among studies, especially in model performance metrics and dataset size. The most common feature extraction techniques were Mel-spectrogram and Mel-Frequency Cepstral Coefficients, with Convolutional Neural Networks and Support Vector Machines being the most frequently used ML models. Despite promising results, most studies relied on small, single-center datasets, limiting the generalizability of findings.

Conclusions:

While ML models show high accuracy in pediatric lung sound analysis, the lack of standard guidelines and reliance on small datasets limits their clinical applicability. Future research should focus on standardized protocols, external validation, and the development of large-scale, multi-center datasets to improve model robustness and clinical implementation.