Multi-Stage Respiratory Sound Analysis: Confidence-Driven Wheeze and Crackle Detection
Clinical Snapshot
PICO Framework
| P — Population | Pediatric patients aged 1-59 months from seven countries (n=742) requiring respiratory sound analysis |
| I — Intervention | Multi-stage confidence-driven framework using contrastive variational recurrent neural network (CVRNN) for automated respiratory sound classification |
| C — Comparator | Traditional aggregation methods and existing state-of-the-art respiratory sound classification systems |
| O — Outcomes | Three-way classification accuracy for normal, wheeze, and crackle sounds; sensitivity and specificity for adventitious sound detection |
Bottom Line
This study presents a novel multi-stage approach for automated detection of wheeze and crackle sounds in pediatric patients using machine learning techniques. While the methodology appears sophisticated and shows promise with balanced accuracies around 70% for individual sound types, the clinical utility remains uncertain due to significant methodological limitations. The absence of sensitivity and specificity data, confidence intervals, and detailed reference standard validation severely limits the ability to assess the diagnostic value of this tool. The 62.12% overall classification accuracy, while representing an improvement over existing methods, may not be sufficient for reliable clinical decision-making. The large international dataset is a strength, but without proper validation studies demonstrating clinical impact and cost-effectiveness, implementation in routine practice would be premature. Further research with rigorous diagnostic accuracy methodology and clinical validation is needed before this technology can be recommended for clinical use.
Key Findings
P Value: Not reported
Effect Size: Final classification accuracy: 62.12%; Balanced accuracies: 72.15% (wheeze), 68.1% (crackle)
Primary Outcome: Three-way classification accuracy for normal, wheeze, and crackle respiratory sounds
Nnt Or Sensitivity: Anomaly detection recall: 98.47%; Individual sensitivity and specificity values not provided
Confidence Interval: Not reported
Clinical Application
Requires specialized audio equipment and computational resources; may be challenging to implement in routine clinical practice without appropriate infrastructure Could support RACGP guidelines for early childhood respiratory assessment and complement existing pediatric screening programs, particularly valuable in remote areas with limited specialist access Pediatric patients aged 1-59 months requiring respiratory assessment, particularly in primary care and screening settings
Abstract
OBJECTIVE: Accurate detection of adventitious respiratory sounds, such as wheezes and crackles, is essential for diagnosing and managing respiratory conditions. This study introduces a multi-stage, confidence-driven framework for automated pediatric auscultation analysis, performing a three-way classification of normal, wheeze, and crackle sounds to improve diagnostic accuracy. METHODS: We develop a comprehensive pipeline integrating anomaly-specific segment selection, segment-level classification, and confidence-based fusion. Our contrastive variational recurrent neural network (CVRNN) enhances feature extraction, while a confidence-weighted aggregation strategy refines final predictions. The system is validated using a diverse pediatric dataset from 742 subjects (aged 1-59 months) from seven countries. RESULTS: The multi-level framework is evaluated across three stages. The anomaly-specific segment selection achieves 98.47 $\%$ recall, identifying adventitious regions. Next, segment-level classifiers improve sensitivity, achieving balanced accuracies of 72.15 $\%$ (wheeze) and 68.1 $\%$ (crackle). This performance surpasses state of the art systems on the same dataset and demonstrates enhanced balanced performance in detecting both crackle and wheeze sounds, which present different challenges to automated systems given their markedly different acoustic profiles. Finally, the confidence-driven fusion outperforms traditional aggregation methods, yielding a final three-way classification of 62.12 $\%$. CONCLUSION: Our confidence-based multi-stage approach enhances automated respiratory sound classification by prioritizing high-certainty segment predictions, aligning with expert physician annotations. SIGNIFICANCE: This framework advances computer-aided respiratory diagnostics, improving early detection and monitoring of pediatric respiratory conditions. By integrating expert-inspired segmentation with machine learning-driven confidence estimation, it has potential to enhance clinical workflows and screening for pulmonary diseases, particularly in resource-limited settings.
Related Research
Sleep medicine
Concordance of wearable device sleep metrics with patient-reported sleep quality: A systematic review
3 Aug 2026
Bioelectrochemistry (Amsterdam, Netherlands)
From data to diagnosis: A comprehensive review of machine learning-driven wearable sensors in healthcare.
2 Aug 2026
JMIR mHealth and uHealth
Wearable Devices for Monitoring and Management of Comorbid Obstructive Sleep Apnea and Hypertension: Scoping Review
1 Aug 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service