Extraction of low-dimensional features for single-channel common lung sound classification

Area of Science:

• Biomedical Engineering
• Signal Processing
• Medical Informatics

Background:

• Accurate lung sound classification is crucial for diagnosing respiratory diseases.
• Automated analysis of respiratory cycles aids in early disease detection.
• Efficient feature extraction is key for developing compact and effective diagnostic systems.

Purpose of the Study:

• To investigate feature extraction methods for classifying single-channel lung sounds.
• To identify distinctive features for autonomous lung disease detection systems.
• To optimize feature sets for improved classification accuracy and reduced data size.

Main Methods:

• Extracted features using Mel frequency cepstrum coefficients (MFCC), time-domain, frequency-domain, and linear predictive coding (LPC).
• Classified 400 respiratory cycles from 94 individuals using algorithms like k-nearest neighbors (k-NN) and support vector machines (SVM).
• Employed Leave One Out Cross-Validation (LOOCV) and Sequential Forward Selection (SFS) for model validation and feature optimization.

Main Results:

• The k-NN algorithm achieved the highest accuracy (90.14% training, 90.63% test set).
• Optimal performance was obtained using a triple feature combination: standard deviation of LPC, and mean and standard deviation of MFCC.
• Feature selection enhanced classification performance, demonstrating the effectiveness of combined feature sets.

Conclusions:

• A combination of MFCC and LPC features, particularly their statistical measures, provides highly accurate lung sound classification.
• The developed feature extraction and classification approach supports the development of autonomous systems for respiratory disease detection.
• This research contributes to more efficient and accurate non-invasive diagnostic tools for pulmonary conditions.