Identification of COVID-19 from Cough Sounds Using Non-Linear Analysis and Machine Learning

Abstract

Automatic diagnosis of COVID-19 has an active role in reducing the spread of the disease by minimizing interaction with people. Machine learning models using various signals and images form the basis of automatic diagnosis. This study presents the machine learning based models for detecting COVID-19 infection using 'Virufy' dataset containing cough sound signals labeled as COVID-19 and Non-COVID-19. Since the number of COVID positive coughs in the set is less than those of COVID negative, firstly, data balancing was performed with the ADASYN oversampling technique in the study. Then, features were extracted by non-linear analysis of cough sounds using Multifractal Detrended Fluctuation Analysis (MDFA), Lempel-Ziv Complexity (LZC) and entropy measures. Later, the most effective features were selected by ReliefF method. Finally, five machine learning algorithms, namely Support Vector Machine with Radial Basis Function (SVM-RBF), Random Forest (RF), Adaboost, Artificial Neural Network (ANN), k Nearest Neighbor (kNN) were used to identify cough sounds as COVID-19 or Non-COVID19. As a result of the study, the cough sounds of COVID-19 patients and Non-COVID19 subjects were identified with 95.8% classification accuracy thanks to the RBF kernel function of SVM and the selected effective features. With this classifier, 93.1% sensitivity, 98.6% specificity, 98.6% precision, 0.92 kappa statistical values and 93.2% area under the ROC curve were obtained.