A Cough-based deep learning framework for detecting COVID-19.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Machine Learning for Healthcare

Background:

• The COVID-19 pandemic necessitates innovative diagnostic tools.
• Cough sounds contain unique biomarkers for respiratory illnesses.
• Current diagnostic methods can be invasive or time-consuming.

Purpose of the Study:

• To develop and evaluate a deep learning framework for COVID-19 detection using cough sounds.
• To combine pre-trained embeddings with handcrafted audio features for improved detection accuracy.
• To assess the framework's performance against established benchmarks and state-of-the-art systems.

Main Methods:

• A two-step approach involving front-end feature extraction and back-end classification.
• Front-end: Combining embeddings from pre-trained models with handcrafted audio features.
• Back-end: Utilizing various classification models to identify COVID-19 positive subjects.

Main Results:

• Achieved second-place ranking in the Second 2021 DiCOVA Challenge (Track-2).
• Obtained an Area Under the Curve (AUC) score of 81.21 and a top F1 score of 53.21 on a Blind Test set.
• Improved upon the challenge baseline by 8.43% (AUC) and 23.4% (F1 score).

Conclusions:

• The proposed deep learning framework demonstrates deployability, robustness, and competitiveness for COVID-19 detection via cough sounds.
• The feature extraction method effectively captures relevant cough sound characteristics.
• This approach offers a promising non-invasive tool for COVID-19 screening.