Classification of Respiratory Sounds into Crackles and Noncrackles Categories via Convolutional Neural Networks

When evaluating the efficacy of expectoration, it is essential not only to consider the subjective perception of patients but also to rely on objective assessment indicators. Crackles, as a crucial indicator for evaluating expectoration effectiveness, have rarely been independently classified in existing research. Additionally, approximately half of pulmonary sounds are incorrectly identified by hospital interns and medical students. To address these challenges, an automatic identification model for crackles is developed. A Convolutional Neural Network (CNN) is employed to classify 1864 instances of crackles and 2291 instances of noncrackles (including rhonchi, wheezes, and normal lung sounds). The model combines pre-trained image feature extractors, respiratory sound data, and a CNN classifier. After training and evaluation, a predictive accuracy of 0.875 is achieved, with an area under the receiver operating characteristic (ROC) curve of 0.89. This research aims to provide objective assessment indicators for expectoration treatment for patients through our research findings.