A Comparative Study of Hybrid Deep Learning Techniques for COVID-19 Detection based on Cough Sound Analysis

The COVID-19 pandemic, produced by the SARS-CoV-2 virus, has severe global consequences, resulting in substantial loss of life and posing a serious threat globally. Cough is a common sign and consequence of COVID-19. Cough sound analysis has the ability to help determine an individual's COVID-19 status. Using deep learning models, this study aims to improve the accuracy of COVID-19 identification based on cough sounds. The work employs twelve separate deep-learning models that were extensively trained on the COUGHVID dataset such as CNN, LSTM, BiLSTM, CNN-LSTM, CNN-BiLSTM with SGD, and Adamax optimizer, Attention-based CNN-LSTM with Adamax, SGD and RMSProp optimizer, Attention-based CNN-BiLSTM Adamax, SGD and RMSProp optimizer. To overcome class imbalance, procedures such as pitch shifting and time-frequency masking are used to increase the positive class. Among these variants, the integration of an attention mechanism model with a convolutional neural network (CNN), a bidirectional long short-term memory (Bi-LSTM) with Adamax optimizer achieved the highest validation accuracy, reaching 95.34%, precision 94.40 %, recall 95.50%, Fl-score 94.44%.