Deep Learning against COVID-19: Respiratory Insufficiency Detection in Brazilian Portuguese Speech

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 625–633 August 1–6, 2021. ©2021 Association for Computational Linguistics 625 Deep Learning against COVID-19: Respiratory Insufficiency Detection in Brazilian Portuguese Speech Edresson Casanova1, Lucas Gris2, Augusto Camargo3, Daniel da Silva2, Murilo Gazzola1, Ester Sabino4, Anna S. Levin4, Arnaldo Candido Jr2, Sandra Aluisio1, Marcelo Finger † 3 1Instituto de Ciˆencias Matem´aticas e de Computac¸˜ao, University of S˜ao Paulo, S˜ao Carlos/SP, Brazil 2Federal University of Technology – Paran´a, Medianeira/PR, Brazil 3Instituto de Matem´atica e Estat´ıtica, University of S˜ao Paulo, S˜ao Paulo/SP, Brazil 4Faculdade de Medicina, University of S˜ao Paulo, S˜ao Paulo/SP, Brazil Abstract Respiratory insufficiency is a symptom that requires hospitalization. This work investi- gates whether it is possible to detect this condi- tion by analyzing patient’s speech samples; the analysis was performed on data collected dur- ing the first wave of the COVID-19 pandemic in 2020, and thus limited to respiratory insuf- ficiency in COVID-19 patients. For that, a dataset was created consisting of speech emis- sions of both COVID-19 patients affected by respiratory insufficiency and a control group. This dataset was used to build a Convolution Neural Network to detect respiratory insuffi- ciency using speech emission MFCC represen- tations. Methodologically, dealing with back- ground noise was a challenge, so we also collected background noise from COVID-19 wards where patients were located. Due to the difficulty in filtering noise without eliminat- ing crucial information, noise samples were in- jected in the control group data to prevent bias. Moreover, we investigated (i) two approaches to address the duration variance of audios, and (ii) the ideal number of noise samples to in- ject in both patients and the control group to prevent bias and overfitting. The techniques developed reached 91.66% accuracy. Thus we validated the project’s Leading Hypothe- sis, namely that it is possible to detect respira- tory insufficiency in speech utterances, under real-life environmental conditions; we believe our results justify further enquiries into the use of automated speech analysis to support health professionals in triage procedures. 1 Introduction This work started as part of the academic initiative to help in the effort to deal with the COVID-19 pandemic in a severely affected region in Brazil. COVID-19 is an infectious disease caused by the †Corresponding author:mfinger@ime.usp.br . virus SARS-CoV-2. This illness is mainly associ- ated to severe acute respiratory syndrome, although it is harmful to other organs, like heart, kidney and brain. About 82% of cases are mild or moderate, while the rest are severe or grave, demanding hos- pitalization or intensive care. The most vulnerable groups are people over the 60’s, and people with specific medical conditions such as diabetes, obe- sity, hypertension and heart disease. According to WHO1, in August 3 2020, more than 19.2 mil- lion people in the world had contracted COVID-19, with a Case Fatality Ratio of CFR=2.8%. Respira- tory Insufficiency (RI) is a symptom that requires hospitalization, which is aggravated due to a fre- quent COVID-19 condition called silent hypoxia, low blood oxygen concentration without breath shortness …

Content truncated to 500 words