Classification of respiratory diseases and COVID-19 from respiratory and cough sound using deep learning techniques

Abstract

Infectious and non-infectious respiratory diseases are among the major reasons for deaths, financial and social crises around the world. However, medical personnel still find it very difficult to detect the diseases using conventional methods to combat this global crisis. We propose a respiratory disease identification method from respiratory auscultation sounds and COVID-19 infected and healthy patients from cough sound recordings. Our experiments demonstrate that artificial intelligence can be utilized as an alternative method to detect respiratory illnesses. We extract image representations of audio features such as Mel-frequency Cepstral Coefficients (MFCCs) and Mel-Spectrogram from each audio recording and use convolutional neural network models for our experiments. Also, we compare the two audio features and ten different convolutional neural network architecture’s performance on disease classification. We conduct experiments with various model training procedures’ such as transfer learning and 1cycle policy, and balanced mini-batch training. In our experiment, we classified respiratory diseases with 94.57 percent accuracy and 0.93 ROC-AUC scores and COVID-19 affected and healthy patients’ cough recordings with 85.96 percent accuracy and 0.84 ROC-AUC scores.