Acoustic analysis of chronic obstructive pulmonary disorder using transfer learning: a three-class problem

This study involves a comparative analysis of deep learning and transfer learning techniques for diagnosing the respiratory disease chronic obstructive pulmonary disease (COPD). Unlike many previous machine learning approaches to this problem, we consider it as a three-class problem. Clinicians are required to inform the patient whether they have COPD, or are healthy, or have an unknown respiratory disorder, making this a three-class classification challenge. We use the ICBHI 2017 respiratory challenge dataset for training and evaluation, with the 'unhealthy' class encompassing data from all other respiratory diseases. To preserve time-series information, we transform lung sounds into Mel spectrograms. We train and test both a convolutional neural network (CNN) and a VGG16 model, and the VGG16 model outperforms the CNN with an impressive 95% accuracy.