Multi-Stage Respiratory Sound Analysis: Confidence-Driven Wheeze and Crackle Detection

Area of Science:

• Medical Informatics
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Accurate detection of adventitious respiratory sounds (wheezes, crackles) is crucial for diagnosing pediatric respiratory conditions.
• Automated auscultation analysis offers potential for improved diagnostic accuracy and clinical workflows.

Purpose of the Study:

• To introduce a multi-stage, confidence-driven framework for automated pediatric auscultation analysis.
• To perform a three-way classification (normal, wheeze, crackle) to enhance diagnostic accuracy.

Main Methods:

• Development of a pipeline integrating anomaly-specific segment selection, segment-level classification, and confidence-based fusion.
• Utilizing a contrastive variational recurrent neural network (CVRNN) for enhanced feature extraction.
• Validation on a diverse pediatric dataset from 742 subjects across seven countries.

Main Results:

• Anomaly-specific segment selection achieved 98.47% recall for identifying adventitious respiratory regions.
• Segment-level classifiers demonstrated balanced accuracies of 72.15% for wheeze and 68.1% for crackle.
• The confidence-driven fusion yielded a final three-way classification accuracy of 62.12%, outperforming traditional aggregation methods.

Conclusions:

• The confidence-based multi-stage approach enhances automated respiratory sound classification by prioritizing high-certainty predictions.
• This framework advances computer-aided respiratory diagnostics for early detection and monitoring of pediatric respiratory conditions.
• Potential to improve clinical workflows and screening, especially in resource-limited settings.