Abstract
This study proposes a novel and robust methodology for detecting COVID-19 through the analysis of cough audio signals by integrating advanced feature engineering with deep learning techniques. We utilize Phase Space Reconstruction (PSR) to convert raw cough sound signals into a multidimensional feature space that captures the complex dynamic behavior inherent in the data. These features are further encoded into a three-dimensional tensor representation, which serves as input to a custom-designed deep convolutional neural network (3D DCNN) comprising five convolutional layers with max-pooling operations. Our model performs multi-class classification to distinguish between COVID-19 positive cases, symptomatic non-COVID individuals, and healthy controls. Leveraging the extensive COUGHVID dataset, which contains over 8,400 cough recordings from diverse populations, the proposed approach achieves a classification accuracy of 98.5%, a recall of 96.5%, and a specificity of 99.7%, outperforming current state-of-the-art methods. The results demonstrate that synthesizing dynamic signal representations with hierarchical 3D deep learning architectures significantly enhances diagnostic accuracy. This non-invasive and scalable framework offers promising potential for rapid and reliable COVID-19 screening, contributing valuable support to public health surveillance and clinical decision-making.