A Machine Learning Strategy to Detect COVID-19 by Inspecting the Statistical Significance of Standard Deviation of Sound Samples of Healthy, Covid and Post Covid Conditions

In the battle against COVID-19, the implementation of an automated non-invasive system for COVID-19 detection offers significant advantages. By reducing exposure risks, such a system prioritizes the safety of both healthcare workers and patients. Furthermore, it improves efficiency by streamlining the testing process, allowing for rapid and accurate results. The system provides standardized and objective assessments, eliminating the potential for human error and ensuring consistent outcomes. Additionally, it enhances patient comfort by eliminating the need for invasive procedures. This paper explores the potential of utilizing time domain features, specifically the standard deviation, extracted from different audio samples like Deep Breath, Dry Cough, Count Normal, and Vowels (a, e, and o), to detect COVID-19 cases by measuring their statistical significance. The study investigates the feature separability of different conditions using Box-Whisker plots and finds that the ‘Count normal’ audio record exhibits superior discriminative ability with ‘H’ and ‘P’ values of 32.46 and 8.943xl0-8, respectively. To evaluate the efficacy of standard deviation of ‘Count normal’ audio records in detecting COVID-19, various Support Vector Machines (SVMs) such as Linear, Quadratic, Cubic, Fine Gaussian, Medium Gaussian, and Coarse Gaussian are employed. Results indicate that the Fine Gaussian SVM outperforms other types of SVMs, achieving training and testing accuracies of 70% and 76.70%, respectively.