Gradient Boosting Machine and Efficient Combination of Features for Speech-Based Detection of COVID-19

In recent times, speech-based automatic disease detection systems have shown several promising results in biomedical and life science applications, especially in the case of respiratory diseases. It provides a quick, cost-effective, reliable, and non-invasive potential alternative detection option for COVID-19 in the ongoing pandemic scenario since the subject's voice can be remotely recorded and sent for further analysis. The existing COVID-19 detection methods including RT-PCR, and chest X-ray tests are not only costlier but also require the involvement of a trained technician. The present paper proposes a novel speech-based respiratory disease detection scheme for COVID-19 and Asthma using the Gradient Boosting Machine-based classifier. From the recorded speech samples, the spectral, cepstral, and periodicity features, as well as spectral descriptors, are computed and then homogeneously fused to obtain relevant statistical features. These features are subsequently used as inputs to the Gradient Boosting Machine. The various performance matrices of the proposed model have been obtained using thirteen sound categories' speech data collected from more than 50 countries using five standard datasets for accurate diagnosis of respiratory diseases including COVID-19. The overall average accuracy achieved by the proposed model using the stratified k-fold cross-validation test is above 97%. The analysis of various performance matrices demonstrates that under the current pandemic scenario, the proposed COVID-19 detection scheme can be gainfully employed by physicians.