Dry and Wet Cough Detection using Fusion of Cepstral base Statistical Features

Abstract

Nowadays with technological advancements, ma-chine learning is widely used in healthcare sector to help patients and doctors. Machine learning offers various tools for healthcare to diagnose various diseases in effective manner. In clinical diagnosis machine learning is used to analyse audio recording of coughs in order to detect respiratory illness. To clear lung and throat from any foreign substance, human body’s inundate mechanism create a substance called Cough. Audio recordings of coughs consists of patterns and depending on the pattern, cough can be classified as wet cough and dry cough. The COUGHVID dataset consists of more than 20,000 audio recordings of cough which includes wide range of subject such as gender, ages, geographic locations, from which more than 2000 recording are labelled by medical experts to diagnose abnormalities present in cough. In this paper, fusion of different cepstral based statistical features and classification using machine learning algorithm is presented. After analysis, it is observed that through ADASYN oversampling highest accuracy of 85.84%, f1 score of 86.80% and the area under the curve as 0.857 is achieved for MLP model.