Page 120 - AIH-1-3

Artificial Intelligence in Health

ORIGINAL RESEARCH ARTICLE
Interpretability analysis of deep models for

COVID-19 detection

Daniel Peixoto Pinto da Silva 1 , Edresson Casanova 2 ,
Lucas Rafael Stefanel Gris 3 , Marcelo Matheus Gauy * , Arnaldo Candido Junior 5 ,
4
Marcelo Finger 4 , Flaviane Romani Fernandes Svartman 6 ,
Beatriz Raposo de Medeiros 7 , Marcus Vinícius Moreira Martins 8 ,
Sandra Maria Aluísio 2 , Larissa Cristina Berti 9 , and João Paulo Teixeira 10
1 Academic Department of Computing, Federal University of Technology – Paraná, Medianeira,
Paraná, Brazil
2 Department of Computer Science, Institute of Mathematical and Computer Sciences, University of
São Paulo, São Carlos, São Paulo, Brazil
3 Institute of Informatics, Federal University of Goiás, Goiania, Goiás, Brazil
4 Department of Computer Science, Institute of Mathematics and Statistics, University of São Paulo,
São Paulo, São Paulo, Brazil
5 Department of Computing and Statistics, Institute of Biosciences, Humanities and Exact Sciences,
São Paulo State University, São José do Rio Preto, São Paulo, Brazil
6 Department of Classical and Vernacular Literature, Faculty of Philosophy, Language, Literature and
Human Sciences, University of São Paulo, São Paulo, São Paulo, Brazil
7 Department of Linguistics, Faculty of Philosophy, Language, Literature and Human Sciences,
University of São Paulo, São Paulo, São Paulo, Brazil
8 Department of Literature and Linguistics, University of the State of Minas Gerais, Belo Horizonte,
Minas Gerais, Brazil
9 Department of Speech Therapy, Faculty of Philosophy and Sciences, São Paulo State University,
Marília, São Paulo, Brazil
10 Department of Eletronics, Research Centre in Digitalization and Intelligent Robotics (CeDRI),
Instituto Politécnico de Bragança, Bragança, Portugal
*Corresponding author:
Marcelo Matheus Gauy
(marcelo.gauy@usp.br) Abstract
Citation: da Silva DPP, During the coronavirus disease 2019 (COVID-19) pandemic, various research
Casanova E, Gris LRS, et al. disciplines collaborated to address the impacts of severe acute respiratory syndrome
Interpretability analysis of deep
models for COVID-19 detection. coronavirus-2 infections. This paper presents an interpretability analysis of a
Artif Intell Health. 2024;1(3):114-126. convolutional neural network-based model designed for COVID-19 detection using
doi: 10.36922/aih.2992 audio data. We explore the input features that play a crucial role in the model’s decision-
Received: February 21, 2024 making process, including spectrograms, fundamental frequency (F0), F0 standard
Accepted: June 17, 2024 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.
Published Online: July 30, 2024 Emphasizing an explainable artificial intelligence approach, our findings demonstrate
Copyright: © 2024 Author(s). that the examined models can make unbiased decisions even in the presence of noise
This is an Open-Access article in training set audios, provided appropriate preprocessing steps are undertaken.
distributed under the terms of the
Creative Commons Attribution Our top-performing model achieves a detection accuracy of 94.44%. Our analysis
License, permitting distribution, indicates that the analyzed models prioritize high-energy areas in spectrograms
and reproduction in any medium, during the decision process, particularly focusing on high-energy regions associated
provided the original work is
properly cited. with prosodic domains, while also effectively utilizing F0 for COVID-19 detection.
Publisher’s Note: AccScience
Publishing remains neutral with Keywords: Coronavirus disease 2019 detection; Voice processing; Gradient-weight class
regard to jurisdictional claims in
published maps and institutional activation mapping
affiliations.

Volume 1 Issue 3 (2024) 114 doi: 10.36922/aih.2992