MSCCov19Net: multi-branch deep learning model for COVID-19 detection from cough sounds

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Bioacoustics

Background:

• Coronavirus (COVID-19) poses a significant global health threat, necessitating rapid and accessible diagnostic tools.
• Current COVID-19 detection methods like RT-PCR, chest X-ray, and CT scans have limitations including cost, radiation exposure, and detection rates.
• Respiratory transmission of COVID-19 underscores the need for non-invasive and remote diagnostic solutions.

Purpose of the Study:

• To develop and evaluate a deep neural network-based model for detecting COVID-19 using only cough sounds.
• To create a fast, remotely operable, and non-invasive diagnostic system for COVID-19.
• To assess the model's performance against existing deep learning architectures and traditional diagnostic methods.

Main Methods:

• A multi-branch deep neural network, MSCCov19Net, was designed to analyze cough sounds.
• The model utilizes Mel Frequency Cepstral Coefficients (MFCC), Spectrogram, and Chromagram as input features.
• Training was performed on publicly available crowdsourced datasets, with testing on unseen clinical and non-clinical datasets.

Main Results:

• MSCCov19Net demonstrated superior performance compared to six popular deep learning architectures across four datasets.
• The model achieved an accuracy of 61.5% on the Virufy dataset and 90.4% on the NoCoCoDa dataset for unseen test data.
• The proposed system exhibited strong generalization capabilities on diverse datasets.

Conclusions:

• Deep neural network analysis of cough sounds presents a viable and effective method for COVID-19 detection.
• MSCCov19Net offers a promising, low-cost, and accessible alternative for remote COVID-19 screening.
• Further research and validation are warranted to integrate cough-based AI diagnostics into public health strategies.