CovNet: A Transfer Learning Framework for Automatic COVID-19 Detection From Crowd-Sourced Cough Sounds

Area of Science:

• Digital Health
• Machine Learning
• Bioacoustics

Background:

• The COVID-19 pandemic highlighted the need for rapid, low-cost diagnostic tools.
• Existing research on detecting COVID-19 from respiratory sounds often relies on limited datasets.
• Computer audition offers a promising avenue for swift, eco-friendly disease detection.

Purpose of the Study:

• To develop a robust deep learning model for COVID-19 detection using cough sounds.
• To address the challenge of limited data in COVID-19 specific cough sound datasets.
• To leverage large, diverse sound datasets for improved transfer learning in respiratory sound analysis.

Main Methods:

• A transfer learning framework, CovNet, was developed, incorporating parameter transferring and embedding incorporation strategies.
• CovNet utilized the large-scale FluSense dataset to train models for COVID-19 detection.
• The effectiveness of CovNet was validated by transferring knowledge to the COUGHVID dataset and subsequently to smaller datasets like ComParE CCS and DiCOVA Track-1.

Main Results:

• The CovNet framework demonstrated improved performance on COVID-19 detection tasks.
• An absolute improvement of 3.57% in ROC AUC was achieved on the DiCOVA Track-1 validation set.
• An absolute improvement of 1.73% in unweighted average recall (UAR) was observed on the ComParE CCS test set.

Conclusions:

• Transfer learning, as implemented in CovNet, is effective for building robust COVID-19 detection models from cough sounds.
• Leveraging large, general sound datasets can significantly enhance the performance of models trained on smaller, specific disease datasets.
• CovNet provides a valuable approach for developing digital diagnostic tools for infectious respiratory diseases.