ABSTRACT
Infectious and non-infectious respiratory diseases are among the primary reasons for deaths, financial and social crises around the world. In this study, we present a comparative analysis of various deep learning techniques for respiratory disease and COVID-19 identification methods from respiratory and cough sound recordings. Our experiments demonstrate that artificial intelligence can help tackle the global crisis by providing an alternative disease diagnosis method. We conduct numerous experiments using deep learning models and model training techniques to find the most efficient disease detection and classification system. We first propose procedures to extract image representations of audio features such as Mel-Spectrograms and Mel-frequency Cepstral Coefficients (MFCC) from each sound recording. Afterward, we compare the performance of the audio features and ten different convolutional neural network (CNN) models on disease classification. We also compare and analyze the performance of various model training methodologies, such as the 1cycle policy, transfer learning, and balanced mini-batch training, to determine the most effective way to train the models. In our experiment, we classify respiratory diseases with 94.57% accuracy and Area under the Receiver Operating Characteristic Curve (AUC) value of 0.93 and COVID-19 infected and healthy patients' cough recordings with 85.62% accuracy and 0.84 AUC value. © 2022 IEEE.