Classification of lung sounds using higher-order statistics: A divide-and-conquer approach

Area of Science:

• Medical Informatics
• Biomedical Signal Processing
• Machine Learning in Healthcare

Background:

• Lung sound auscultation is a common diagnostic method for respiratory diseases.
• Effectiveness relies heavily on physician expertise, leading to potential misdiagnosis.
• Objective classification of lung sounds is needed to improve diagnostic accuracy.

Purpose of the Study:

• To implement a pattern recognition system for automated lung sound classification.
• To evaluate the effectiveness of higher-order statistics (HOS) for feature extraction.
• To develop a robust classifier for distinguishing various lung sound types.

Main Methods:

• Utilized a dataset of five lung sound types: normal, coarse crackle, fine crackle, monophonic wheeze, and polyphonic wheeze.
• Extracted features using higher-order statistics (HOS) cumulants.
• Employed Genetic Algorithms (GA) and Fisher's Discriminant Ratio (FDR) for dimensionality reduction.
• Classified lung sounds using k-Nearest Neighbors and Naive Bayes within a tree-based system.

Main Results:

• Genetic Algorithms demonstrated superior performance over Fisher's Discriminant Ratio for feature selection.
• Higher-order statistics revealed distinct signature patterns for each lung sound class.
• The proposed tree-based classifier achieved 98.1% training and 94.6% validation accuracy.
• The divide-and-conquer approach accurately classified different lung sound types.

Conclusions:

• The developed pattern recognition system shows high accuracy in classifying lung sounds.
• Higher-order statistics proved to be a promising tool for lung sound feature extraction.
• The proposed classifier is feasible for implementation in embedded systems for real-time diagnostics.