Assessment of an Automated Wheezing Respiratory Cycle Identification Based on CNN and Classical and High-Resolution Time-Frequency Representations

Respiratory sounds (RS) provide relevant information that helps diagnose pathologies. Some common continuous adventitious RS are wheezes, present in diseases such as asthma and chronic obstructive pulmonary disease (COPD). In recent decades, efforts have been made to automatically detect wheezes in RS, e.g., by classical time-frequency (TF) representations such as the spectrogram, which allows the visualization of instantaneous changes in frequency over time. However, due to the characteristics and limitations of the previously used TF representations, the detection of wheezes with these techniques still offers areas of opportunity. In this work, we propose to perform the classification of normal RS cycles and RS cycles containing wheezes using two TF methods: (1) Spectrogram (SP) and (2) Hilbert-Huang Spectrum (HHS), as input to a pre-trained convolutional neural network (CNN), the GoogleNet. The proposed method was evaluated on the ICBHI Challenge 2017 database, using 575 wheezing respiratory cycles from 45 patients and 575 normal cycles from 44 subjects. The best experimental result achieved a specificity of 0.883, precision of 0.880, and accuracy of 0.872 for the validation set using SP method. Comparing this performance with those reported in the literature, it was found that the classification results obtained are at least comparable to those reported in most state-of-the-art works.