Machine Learning for Automated Classification of Abnormal Lung Sounds Obtained from Public Databases: A Systematic...

Area of Science:

• Pulmonary medicine
• Biomedical engineering
• Artificial intelligence in healthcare

Background:

• Pulmonary auscultation is crucial for diagnosing lung conditions but is operator-dependent.
• Machine learning (ML) models offer automated lung sound classification, requiring large datasets.
• Publicly available databases aim to provide the necessary data for ML model development.

Purpose of the Study:

• To systematically review and compare ML models for lung sound classification.
• To assess the characteristics, diagnostic accuracy, and data sources of these models.
• To identify limitations and concerns within existing research and public databases.

Main Methods:

• A systematic literature review of papers published between 1990 and 2022 from five major databases.
• Quality assessment of included studies using a modified QUADAS-2 tool.
• Analysis of 62 studies employing ML models and public lung sound databases.

Main Results:

• Artificial neural networks (ANN) and support vector machines (SVM) were common ML classifiers.
• Diagnostic accuracy varied: 49.43%–100% for abnormal sound types and 69.40%–99.62% for disease classification.
• Seventeen public databases were identified; ICBHI 2017 was the most utilized (66%).
• Most studies showed high risk of bias, particularly in patient selection and reference standards.

Conclusions:

• ML models show promise for classifying abnormal lung sounds using public data.
• Inconsistent reporting and methodologies hinder field advancement.
• Standardized recording and labeling procedures for public databases are essential for future progress.