Automatic identification of abnormal lung sounds by machine learning methods

Abstract

This study introduces an automated approach for identifying abnormal lung sounds from audio recordings utilizing time-frequency analysis and convolutional neural networks. Acoustic signals captured via a stethoscope are subjected to noise removal using a bandpass filter. Subsequently, distinctive features are extracted via a short-time Fourier transform to represent frequency components in the form of a spectrogram. The spectrogram facilitates the segmentation of breathing cycles by identifying the highest and lowest peaks, thereby quantifying the number of breathing cycles within the audio signal. Following this segmentation, the breathing cycle is partitioned into training and test datasets, with the convolutional neural networks trained on the former to optimize model performance. Experimental findings demonstrate that the proposed method effectively achieves the accuracies of 85.34 percent, 68.20 percent, and 60.64 percent for wheezing sounds, crackle sounds, and normal sounds, respectively.