C19-MLE: A Multi-Layer Ensemble Deep Learning Approach for COVID-19 Detection Using Cough Sounds and X-Ray Imaging

Abstract

The COVID-19 pandemic highlighted the urgent need for rapid and efficient screening methods, leading to a growing demand for alternatives to resource-intensive RT-PCR tests. Among these, intelligent, contact-free automated systems emerged as a promising solution for quick preliminary COVID-19 detection. This study introduces the COVID-19 Multi-Layer Ensemble framework (C19-MLE), designed to enhance the accuracy of COVID-19 detection. The approach begins with a 2D convolutional neural network (CNN) combined with a variation autoencoder for precise classification of cough sounds. Additionally, a UNet-based encoder-decoder architecture is used for segmenting chest X-ray images. These segmented images are then classified using two models, ResNet-50 and Inception V3, and their results are combined using an ensemble learning technique. This first-layer ensemble achieves an impressive accuracy of 98.5% in classifying chest X-rays. Meanwhile, the proposed 2D CNN model for cough classification achieves an accuracy of 97.79%. The second-layer ensemble, which fuses the results of both chest X-ray and cough classifications using a meta-classifier with a hard prediction and weighted sum-rule technique, achieves a remarkable overall accuracy of 99.89%. The C19-MLE framework demonstrates the powerful synergy between cough audio signals and chest X-ray images, providing a highly accurate method for preliminary and post-screening COVID-19 diagnosis. The high accuracy of this model highlights its potential as a crucial tool for early disease detection and prevention, especially in settings where resources are limited.