An automated Covid-19 respiratory sound classification method based on novel local symmetric Euclidean distance pattern and ReliefF iterative MRMR feature selector

Abstract

Covid-19 is a new variety of coronavirus that affects millions of people around the world. This virus infected millions of people and hundreds of thousands of people have passed away. Due to the panic caused by Covid-19, recently several researchers have tried to understand and to propose a solution to Covid-19 problem. Especially, researches in machine learning (ML) have been proposed to detect Covid-19 by using X-ray images. In this study, 10 classes of respiratory sounds, including respiratory sounds diagnosed with Covid-19 disease, were collected and ML methods were used to tackle this problem. The proposed respiratory sound classification method has been proposed in this study from feature generation network through hybrid and iterative feature selection to classification phases. A novel multileveled feature generating network is presented by gathering multilevel one-dimensional wavelet transform and a novel local symmetric Euclidean distance pattern (LSEDP). An automated hybrid feature selection method is proposed using ReliefF and ReliefF Iterative Maximum Relevancy Minimum Redundancy (RIMRMR) to select the optimal number of features. Four known classifiers were used to test the capability of our approach for lung disease detection in respiratory sounds. K nearest neighbors (kNN) method has achieved an accuracy of 91.02%.