Deep Learning Model for the Automatic Analysis of Cough in the Early Diagnosis of COVID-19

  • Nahum Flores, José La Rosa, Ciro Rodriguez, Freddy Kaseng


The purpose of the research is to identify people infected with COVID-19 through a diagnostic
model based on Deep Learning, which will be implemented in a multiplatform application that will support
the available clinical methods that require specialized and exhaustive supervision also to protect the
doctors. The methodology considers re-training the Deep Learning model trained with the SickSounds
dataset with cough sounds of sick and not sick people with COVID-19, a fine-tuning to the ResNet50
architecture model trained for the diagnostic module will be used. The accuracy, sensitivity, and
specificity results are 81.06%, 85.4%, and 74.96%, respectively, using 10-fold cross-validation.