Enhanced Respiratory Sound Classification Using Deep Learning and Multi-Channel Auscultation.

Area of Science:

• Medical Diagnostics
• Bioacoustics
• Artificial Intelligence

Background:

• Accurate classification of abnormal lung sounds is crucial for diagnosing respiratory disorders.
• Multichannel auscultation signals offer greater diagnostic potential than single-channel recordings.
• Leveraging positional characteristics from multiple auscultation sites enhances diagnostic capabilities.

Purpose of the Study:

• To improve respiratory sound classification accuracy using multichannel signals.
• To capture and analyze positional characteristics from multiple lung sound recording sites.
• To evaluate the impact of channel number and placement on classification performance.

Main Methods:

• A deep learning model combining convolutional neural networks (CNN) and long short-term memory (LSTM) was employed.
• Mel-frequency cepstral coefficients (MFCCs) were used to analyze respiratory sound data.
• Performance was evaluated based on the number and placement of recording channels.

Main Results:

• Four-channel recordings demonstrated significant improvements in accuracy, sensitivity, specificity, precision, and F1-score.
• Improvements ranged from approximately 1.05 to 1.15 times compared to fewer channels.
• Multichannel data captured richer features for respiratory sound classification.

Conclusions:

• Multichannel data acquisition significantly enhances respiratory sound classification performance.
• The proposed deep learning approach shows promise for clinical respiratory diagnostics.
• The method has potential applications beyond medicine, including speech and audio processing.