Impact of face covering models on respiratory sound classification applications

Abstract

Respiratory sound evaluation and classification has the potential to provide healthcare professionals with information that would otherwise be unavailable, especially in light of the COVID-19 pandemic. With the adoption of face masks and cough covering best practices, understanding the impact of face coverings on recorded audio measurements is essential. In this paper, system identification has been applied to four face covering states (disposable mask, N95 mask, fabric mask, and elbow covering) leading to four transfer functions that can be applied pre-recorded vocal sounds. As covering a cough with a bent elbow led to the highest level of frequency attenuation, it was used to evaluate three classifiers created using the original uncovered data, the elbow covered modeled data, and a combination of both. Each classifier used YAMNet embeddings to classify between four respiratory sounds. The classifier built using the original uncovered and modeled elbow covered data led to the highest performance when evaluated on either the uncovered or modeled data, with accuracies of 0.72. The application of these models can not only evaluate the robustness of preexisting respiratory classifiers in the presence of face coverings but may also be used as a data augmentation tool for human vocal sounds.