COVID-19 and flu diagnosis from short electrocardiogram images using a residual neural network

Abstract

Background and Objective The concurrent prevalence of respiratory illnesses such as influenza during the COVID-19 pandemic has increased the healthcare burden of accurate diagnosis through clinical and laboratory methods. To alleviate this burden, we propose a deep learning model capable of diagnosing multiple respiratory diseases using electrocardiogram (ECG) images.

Methods We developed a shallow residual network (ResNet) composed of four residual blocks with skip connections, convolutional layers, ReLU functions, and dropout, to classify COVID-19, influenza, cardiovascular disease (CVD), and normal conditions using 5-second ECG data. The model’s performance was compared against machine learning models (Linear Regression, XGBoost, SVM) and pre-trained models (VGG16, Inception ResNet-v2, DenseNet121).

Results The model achieved average accuracies of 87.8 % for normal vs. COVID-19, 77.2 % for normal vs. influenza, and 95.4 % for COVID-19 vs. influenza. It also achieved average accuracies of 79.4 % for ternary classification (normal vs. COVID-19 vs. influenza) and 86.2 % for quaternary classification (normal, COVID-19, influenza, CVD). These results outperformed traditional machine learning models and matched or exceeded transfer learning models with lower computational costs.

Conclusions This study demonstrates the potential of ECG images for noninvasive, rapid, and cost-effective respiratory disease diagnosis, offering promising applications as ECG monitoring becomes increasingly accessible.