Robust ECG Paper-Based COVID-19 Classification: Addressing Class Imbalance with Resampling Techniques and Lightweight MobileNetV2

Abstract

The COVID-19 pandemic has highlighted the critical need for rapid, accessible, and complementary diagnostic methods. This study investigates the practicality of utilizing deep learning on electrocardiogram (ECG) signals for the binary classification of COVID-19 vs. normal cases. The proposed approach employs a lightweight pre-trained model, MobileNetV2, and an imbalanced ECG paper dataset, Khan’s dataset. To address the key issue of class imbalance in the dataset, a thorough evaluation of sampling strategies—under-sampling, over-sampling, and hybrid sampling—was carried out. A 60/10/30 split was employed for the training, validation, and test sets, respectively. The lightweight MobileNetV2 model, along with a hybrid sampling approach (SMOTE-Tomek) and the RMSprop optimizer, demonstrated exceptional performance. The proposed model obtained a max accuracy of 99.4% and an F1-measure of 99% on the test set in 17.35 seconds. These findings suggest that ECG-based classification, when paired with strong imbalance-handling approaches and efficient neural networks, can be a highly accurate and computationally economical tool for screening COVID-19