Comprehensive Analysis System for Automated Respiratory Cycle Segmentation and Crackle Peak Detection

Area of Science:

• Medical signal processing
• Respiratory acoustics
• Computational phoniatrics

Background:

• Digital auscultation for lung sound assessment is subjective and lacks robust real-world application for crackle analysis.
• Existing crackle detection methods are often limited by integration into comprehensive systems and validation on non-ideal data.

Purpose of the Study:

• To develop and validate a complete signal analysis methodology for objective crackle detection in challenging respiratory sound recordings.
• To improve the accuracy and real-world applicability of automated lung sound analysis.

Main Methods:

• A five-block sequential processing methodology including motion artifact detection, deep learning denoising, respiratory cycle segmentation, adventitious sound separation, and crackle peak detection.
• Utilized novel methods and robustness improvements for analyzing respiratory cycles and crackles.
• Validated on a database of 1000 simulated lung sounds with controlled variations and a small set of 20 patient recordings.

Main Results:

• Achieved an average F-score of 91.07% for motion artifact detection and 94.43% for respiratory cycle extraction.
• Demonstrated an overall F-score of 94.08% for detecting individual crackle locations.
• Successfully identified healthy recordings and showed comparable performance on preliminary patient data.

Conclusions:

• The developed methodology provides a robust and quantifiable approach to analyzing respiratory sounds, specifically crackles.
• Enables objective distinction of crackle types, timing, and occurrence, contributing to better understanding of cardiorespiratory disease severity and progression.
• Offers a simple, non-invasive tool for clinicians to enhance diagnostic capabilities in respiratory medicine.