COVID-19 Detection from Cough Sound Signal using Random Forest Learning of Deep Spectral Features

The COVID-19 epidemic and its subsequent rise have brought attention to the need for quick, precise, and reasonably priced testing techniques. Existing assays such as RT-PCR are slow and conditionally sensitive, which presents difficulties, particularly in low-resource settings. Researchers have shown that COVID-19 can be accurately detected by audio-based AI algorithms using cough sounds. In this work, a cough sound-based COVID-19 detection framework is presented. It utilizes a 1D Convolutional Neural Network (CNN) deep feature extractor from handcrafted spectral features and Random Forest classifier for classifying the healthy and Covid-19 subjects from cough sounds. The research considered the cough sound from two datasets named Coswara and Virufy emphasizing the generalization capability of the proposed framework. With the help of feature engineering, spectral features like Short-Term Fourier Transform (STFT), Mel Spectrogram, Mel Frequency Cepstrum Coefficient (MFCC), RMS Energy, Spectral Bandwidth, Zero Crossing Rate (ZCR), Spectral Centroid, and Spectral roll-off have been extracted which are further fed into a CNN based algorithm for extracting deep features. The deep features are more crucial for complex pattern analysis compared to the simple frequency domain analysis. Finally, the Random Forest model is used for classification tasks, which shows a predictive accuracy of 93% with robustness against data diversity. The proposed framework might be useful for smartphone-based COVID-19 detection applications.