Crackle and wheeze detection in lung sound signals using convolutional neural networks

Area of Science:

• Medical Technology
• Artificial Intelligence in Healthcare
• Respiratory Medicine

Background:

• Respiratory diseases are a major global health concern, necessitating effective diagnostic tools.
• Pulmonary auscultation is a vital, non-invasive screening method, but human limitations exist in detecting subtle lung sound abnormalities.
• Computer-assisted decision systems offer potential for enhanced detection of abnormal respiratory sounds like crackles and wheezes.

Purpose of the Study:

• To develop and evaluate a novel system for detecting and classifying abnormal lung sound events.
• To leverage artificial intelligence for improved accuracy in respiratory sound analysis.
• To provide a framework for comparable results using a standard dataset and metrics.

Main Methods:

• A convolutional neural network (CNN) was employed for lung sound analysis.
• Mel spectrograms were utilized as input features for the CNN model.
• The system was trained and validated on the ICBHI 2017 challenge dataset.

Main Results:

• The proposed system demonstrated the ability to both detect and classify abnormal lung sound events.
• Performance metrics achieved were 43% accuracy and 51% sensitivity.
• Results are comparable to the current state-of-the-art in respiratory sound analysis.

Conclusions:

• The developed system shows promise in assisting clinicians with the diagnosis of respiratory conditions.
• AI-powered analysis of lung sounds can overcome human auditory limitations.
• Further research and development can enhance the accuracy and clinical utility of such systems.