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

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Pediatric Pulmonology

Background:

• Pediatric respiratory diseases are significant causes of childhood illness and death.
• Traditional lung sound auscultation is subjective and varies between clinicians.
• AI and machine learning (ML) offer objective, automated analysis of lung sounds via electronic stethoscopes.

Purpose of the Study:

• To systematically review and meta-analyze the performance of ML models in pediatric lung sound analysis.
• To evaluate methodologies, model performance, and database characteristics.
• To identify limitations and future directions for clinical implementation of ML in pediatric respiratory diagnostics.

Main Methods:

• Systematic literature search across major databases (PubMed, Embase, Web of Science, etc.) from 1990 to 2024.
• Inclusion of studies developing ML models for pediatric lung sound classification with defined databases and physician-labeled standards.
• Risk of bias assessment using a modified QUADAS-2 framework and bivariate meta-analysis for binary classification tasks.

Main Results:

• 41 studies met inclusion criteria; most focused on wheeze and abnormal lung sound detection.
• Pooled sensitivity and specificity for wheeze detection were 0.902 and 0.955, respectively.
• Convolutional Neural Networks were common ML models, but high heterogeneity in datasets and methods limited generalizability.

Conclusions:

• ML models show promise for accurate pediatric lung sound analysis.
• Limitations include dataset heterogeneity, lack of standardized guidelines, and insufficient external validation.
• Future research should prioritize standardized protocols and large, multicenter datasets for improved clinical utility.