Robust and Interpretable Temporal Convolution Network for Event Detection in Lung Sound Recordings.

Area of Science:

• Medical acoustics
• Respiratory diagnostics
• Machine learning in healthcare

Background:

• Lung sound auscultation is crucial for diagnosing respiratory diseases.
• Manual segmentation of lung sound recordings is time-consuming and prone to errors.
• Existing automated methods often struggle with noisy and irregular recordings.

Purpose of the Study:

• To develop a novel framework for accurate and efficient lung sound event detection.
• To segment continuous lung sound recordings into discrete respiratory events.
• To recognize specific auscultation events such as inhalations, crackles, and rhonchi.

Main Methods:

• Utilized a multi-branch Temporal Convolutional Network (TCN) architecture.
• Implemented a novel feature fusion strategy for combining branch outputs.
• Processed recordings of arbitrary length, retaining salient information and disregarding irrelevant features.

Main Results:

• Evaluated the framework on diverse public and in-house datasets with noisy recordings.
• Successfully identified key auscultation events including inhalation, crackles, and rhonchi.
• Demonstrated superior performance in feature fusion, leading to a robust, lightweight network.

Conclusions:

• The proposed feature concatenation method effectively suppresses non-informative features, reducing classifier overhead.
• The framework offers a cost-effective and efficient alternative to manual segmentation.
• End-to-end model interpretability enhances trust for clinical application.