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

Summary

This study proposes a generative AI-based framework that integrates GANs, VAEs, and attention-based deep convolutional networks to detect COVID-19 from cough audio signals, improving robustness, data balance, and cross-dataset generalization for scalable, non-invasive screening.

Abstract

Cough audio analysis provides a practical pathway for developing automated tools to support COVID-19 disease screening. Despite growing interest, many existing approaches remain sensitive to noise, class imbalance, and dataset variability, limiting their reproducibility. This work presents a structured generative artificial intelligence-driven methodology for COVID-19 detection using cough sound recordings. Experiments are conducted using two publicly accessible datasets, COUGHVID and Virufy, both of which comprise labeled cough samples. The proposed protocol consists of sequential preprocessing stages, including cough segmentation, denoising, and signal normalization. Acoustic characteristics are then captured through Mel-Frequency Cepstral Coefficients, chroma descriptors, and spectral contrast features. To improve representation learning and mitigate data imbalance, a hybrid generative framework integrating Variational Autoencoders and Generative Adversarial Networks is employed to synthesize feature-level samples. Classification is subsequently performed using Deep Convolutional Neural Networks and Attention-based DCNN models. Performance evaluation indicates that incorporating generative augmentation consistently improves over non-generative baselines, achieving a peak classification accuracy of 97.2% and an AUROC of 0.953 across the evaluated datasets. These results demonstrate the effectiveness of generative modeling for enhancing cough-based COVID-19 detection and establish a reproducible analytical pipeline for future research in acoustic health monitoring.

Introduction

Respiratory acoustics have increasingly been explored as a non-invasive source of information for health assessment, particularly in the context of infectious and pulmonary diseases. Advances in signal processing and artificial intelligence (AI) have enabled automated analysis of cough and breathing sounds, supporting their use as digital biomarkers. Recent studies demonstrate that respiratory sounds captured using microphones embedded in smartphones, wearable devices, or clinical sensors can be effectively analyzed using deep learning models to detect COVID-19 infection^1,5. These developments highlight the po....

Protocol

Proposed Methodology

This study proposes a generative AI–based framework for detecting COVID-19 from cough signals. The framework integrates generative models with discriminative classifiers, with training and evaluation data kept strictly separate. Figure 1 illustrates the detailed architecture of the proposed GAN–VAE–ADCNN framework, showing the flow from audio preprocessing and feature extraction through latent representation learning via a Variational Autoencoder (VAE), synthetic feature generation using a Generative Adversarial Network (GAN), and final classification using Deep Convolution....

Results

Dataset Composition and Class Distribution Analysis

Following preprocessing and segmentation, the COUGHVID and Virufy datasets showed substantial differences in both dataset scale and segment density. COUGHVID contributed 428 out of 444 recordings (96.4%) and 1,719 out of 1,953 extracted cough segments (88.0%), confirming its role as the primary dataset for model training and generative augmentation (Figure 3). In contrast, Virufy contributed only.......

Discussion

The results demonstrate that the proposed generative AI–based framework can detect COVID-19 from cough signals across heterogeneous datasets. As shown in Table 4 and Table 6, the GAN–VAE hybrid model achieved the best performance, with 97.2% accuracy and an AUROC of 0.953, along with high precision, recall, F1-score, and specificity, indicating balanced classification performance.

Performance improved consistently when generative models were combin.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
COUGHVID Dataset	École Polytechnique Fédérale de Lausanne (EPFL)	Public release (2021)	Crowdsourced cough audio dataset with self-reported metadata and partial physician annotation; used primarily for training and data augmentation.
CUDA Toolkit	NVIDIA	11.3	GPU acceleration framework used to speed up model training and inference.
LibROSA	Open Source	0.9	Audio analysis library used for resampling, segmentation, MFCC extraction, and spectral feature computation.
Linux Operating System	Canonical	Ubuntu 20.04 LTS	Operating system used for all experiments and model training.
Matplotlib	Open Source	3.5	Visualization library used for plotting dataset distributions and evaluation results.
NumPy	Open Source	1.21	Numerical computing library used for array manipulation and signal processing operations.
NVIDIA GPU	NVIDIA	Tesla / RTX class	Graphics processing unit used for training deep and generative models.
Python Programming Language	Python Software Foundation	3.8	Core programming environment used for data preprocessing, feature extraction, model development, and evaluation.
PyTorch	Meta (Facebook AI Research)	1.12	Deep learning framework used to implement GAN, VAE, DCNN, and attention-based DCNN models.
Scikit-learn	Open Source	1	Machine learning library used for data splitting, class weighting, and performance metric computation.
SciPy	Open Source	1.7	Scientific computing library used for audio preprocessing and signal transformations.
Virufy Dataset	Virufy	Public release (2021)	Clinically collected cough recordings with RT-PCR–verified COVID-19 labels; used exclusively for external validation and cross-dataset evaluation.