Classification of COVID-19 Using Temporal and Spectral Features of Cough Sounds

• Published in: Bioautomation Vol. 29; no. 1; pp. 5 - 18
• Main Author: Abera Tessema, Biruk
• Format: Journal Article
• Language: English
• Published: Sophia Bulgarska Akademiya na Naukite / Bulgarian Academy of Sciences 01.03.2025; Bulgarian Academy of Sciences
• Subjects: coronavirus; Coronaviruses; COVID-19; Disease; feature extraction; hyperparameter optimization; Sound
• ISSN: 1314-1902; 1314-2321
• DOI: 10.7546/ijba.2025.29.1.000931

Chest X-ray and computed tomography scan play a major role in the diagnosis of lung diseases, including coronavirus disease (COVID-19). However, their cost, the obstacles to their implementation in health facilities in small settlements of developing countries, and the limitations of their use for daily assessment due to the risk of repeated radiation dose, greatly limit their application. In response to the search for safe, simple, rapid, non-invasive, and cost-effective promising alternatives for the diagnosis of COVID-19, researchers in the field are increasingly turning to the analysis of human respiratory sound signals, including cough, breathing, and voice sounds. This is due to the direct connection of the respiratory sound signals with the lungs. Despite the detection efficiency obtained in earlier related works, further studies are still needed on the ability of breath sounds to provide meaningful information about COVID-19. This study used 2660 samples of cough sounds (1 330 recordings from healthy subjects and 1 330 recordings from subjects infected with COVID-19) from the CoughVid dataset, to train models for the classification of the COVID-19 disease. An attempt has been made to classify COVID-19 using different machine-learning models. Temporal and spectral features were extracted from the amplitude spectrum of cough sound signals, and evaluated using a periodogram, and those with higher discriminative power were selected. 1862 cough sound recordings were used for training and 798 cough sound recordings were used to test the model. On the test set, the final optimized model achieved classification accuracy, sensitivity, and specificity of 97.87%, 97.90%, and 97.85%, respectively. The experimental results of the study showed that the proposed method provides significant accuracy for classifying the COVID-19 disease, making it a reliable decision-support tool in healthcare settings where reverse transcription polymerase chain reaction is not available and test kits are scarce.