Cough Sound Based Deep Learning Models for Diagnosis of COVID-19 Using Statistical Features and Time-Frequency Spectrum.

Abstract

This paper presents a deep learning model that can classify COVID-19 patients through cough sounds. The cough sound data were selected from the Cambridge data set which is a crowedsourced data set collected from the Cambridge COVID-19 sounds application. Virufy and Coswara data sets were also selected for external testing. For the sound waveform, we extracted Variable frequency complex demodulation (VFCDM) image and applied to Xception, which is selected as the pre-trained model. Then we extracted zero crossing rate (ZCR), and spectral roll-off (SR), spectral centroid (SC), spectral bandwidth (SB), and concatenated them to the output node of the model. Results were evaluated by using area under receiver operating curve. Cambridge data set: 0.9346, Virufy data set: 0.9244, Coswara data set: 0.8250.