Detection of crackle events using a multi-feature approach

Area of Science:

• Pulmonary Medicine
• Biomedical Signal Processing
• Machine Learning in Healthcare

Background:

• Adventitious lung sounds, such as crackles, are key indicators of respiratory diseases like chronic obstructive pulmonary disease (COPD).
• Accurate and automated detection of these sounds is crucial for effective patient monitoring and management.
• Crackles specifically are explosive respiratory sounds linked to inflammation or infection in the lower airways.

Purpose of the Study:

• To propose and evaluate a multi-feature approach for the automatic detection of crackles in respiratory sounds.
• To identify optimal features and parameters for enhancing the accuracy of crackle detection.
• To assess the performance of the proposed method using a dataset comprising normal and abnormal lung sounds.

Main Methods:

• A multi-feature approach was employed, testing thirty-five distinct features including Music Information Retrieval (MIR) features, a wavelet-based feature, Teager energy, and entropy.
• A logistic regression classifier was utilized for the classification of lung sound events.
• Optimal detection parameters, including frame size (128 ms) and number of features (twenty-seven), were determined via grid search.

Main Results:

• The proposed multi-feature approach demonstrated effective detection of crackles within the tested dataset.
• The optimal configuration involved a frame size of 128 ms and the selection of twenty-seven features.
• Performance was evaluated based on sensitivity and positive predictive value, indicating the method's potential for clinical application.

Conclusions:

• The developed multi-feature detection method shows promise for the automated identification of crackles.
• This technique can serve as a valuable tool for monitoring respiratory conditions, aiding in early diagnosis and management.
• Further validation with larger and diverse datasets is recommended to solidify its clinical utility.