Interpretability analysis of deep models for COVID-19 detection

Abstract

During the coronavirus disease 2019 (COVID-19) pandemic, various research disciplines collaborated to address the impacts of severe acute respiratory syndrome coronavirus-2 infections. This paper presents an interpretability analysis of a convolutional neural network-based model designed for COVID-19 detection using audio data. We explore the input features that play a crucial role in the model’s decision-making process, including spectrograms, fundamental frequency (F0), F0 standard deviation, sex, and age. Subsequently, we examine the model’s decision patterns by generating heat maps to visualize its focus during the decision-making process. Emphasizing an explainable artificial intelligence approach, our findings demonstrate that the examined models can make unbiased decisions even in the presence of noise in training set audios, provided appropriate preprocessing steps are undertaken. Our top-performing model achieves a detection accuracy of 94.44%. Our analysis indicates that the analyzed models prioritize high-energy areas in spectrograms during the decision process, particularly focusing on high-energy regions associated with prosodic domains, while also effectively utilizing F0 for COVID-19 detection.