A Prospective Study on Application of Artificial Intelligence in Detection of Wheeze Sounds in Children

Background: Wheeze is a common and clinically important adventitious lung sound in children that often signals airway obstruction and may indicate asthma exacerbations. Manual auscultation is subjective and can miss subtle adventitious sounds, particularly in young or uncooperative children. Recent advances in deep learning offer an opportunity to automate and standardize wheeze detection. We performed a Methods: prospective analytical study at the Department of Paediatrics, A.J. Institute of Medical Sciences, Mangalore (April 2023–October 2024). One hundred eight children (age 1–12 years) presenting with respiratory complaints were enrolled. Four chest auscultation recordings per child (anterior/posterior, left/right) were collected with an electronic stethoscope (432 recordings total). After preprocessing and mel-spectrogram extraction, we trained and compared multiple models (3-layer LSTM, 4-layer CNN, 4-layer CNN + tabular data, and a proposed ResNet34 with Convolutional Block Attention Module [CBAM] combined with tabular data). Model performance was evaluated on held-out test data using accuracy, AUC, precision, recall, and F1-score. Clinical performance of the AI algorithm was compared against general pediatricians' auscultation. Results: The proposed ResNet34+CBAM+tabular model achieved the highest performance: accuracy 91.2%, AUC 0.891, precision 94.4%, recall 81.0%, and F1-score 87.2%. Across 403 analyzable recordings (93.2% of total), the AI algorithm outperformed general pediatricians for wheeze detection (sensitivity 86.4% vs 82.2%; specicity 83.0% vs 72.1%) and markedly for crackles (sensitivity 81.1% vs 47.8%; specicity 94.1% vs 77.1%). Multivariable analysis identied male sex, family history of asthma (OR 2.20), household smoker (OR 1.43), and history of eczema (OR 3.85) as independent risk factors for current wheeze. A ResNet34-based deep learning model with atte Conclusion: ntion and minimal tabular inputs accurately detects wheeze and other adventitious sounds and outperforms general pediatricians in this cohort. AI-assisted auscultation shows promise as an objective adjunct for pediatric respiratory assessment and could aid earlier diagnosis and monitoring of wheeze-associated conditions.