A Generative AI-Based Framework for COVID-19 Screening from Cough Audio Signals

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Computational Health

Background:

• Automated COVID-19 screening using cough audio is promising but challenged by noise, data imbalance, and variability.
• Existing methods often lack reproducibility due to these limitations.

Purpose of the Study:

• To develop a structured, generative artificial intelligence (AI)-driven methodology for COVID-19 detection from cough sounds.
• To enhance the robustness and reproducibility of AI-based cough analysis for disease screening.

Main Methods:

• Utilized two public datasets (COUGHVID, Virufy) with labeled cough samples.
• Implemented sequential preprocessing: cough segmentation, denoising, and normalization.
• Extracted acoustic features (MFCCs, chroma, spectral contrast).
• Employed a hybrid generative framework (Variational Autoencoders, Generative Adversarial Networks) for feature synthesis to address data imbalance.
• Performed classification using Deep Convolutional Neural Networks (DCNNs) and attention-based DCNN models.

Main Results:

• Generative augmentation significantly improved performance over non-generative baselines.
• Achieved a peak classification accuracy of 97.2%.
• Reached an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.953 across datasets.

Conclusions:

• Generative modeling is effective for enhancing cough-based COVID-19 detection.
• The proposed methodology provides a reproducible pipeline for acoustic health monitoring research.
• This AI-driven approach offers a practical pathway for improved automated disease screening.