Comparative Analysis of ECG Image Classification for Cardiac Disease Detection Using Deep Learning Approaches

Background: Electrocardiogram (ECG) interpretation is essential for cardiac disease diagnosis. While signal-based methods have been dominant, recent advances in deep learning introduced image-based classification, offering new directions for automated detection. Objectives: This study investigates the effectiveness of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN–LSTM model for ECG image classification, focusing on whether combining spatial and temporal feature learning improves accuracy compared to individual models. Methods: ECG images were collected from public repositories; CNN and LSTM models were first applied directly to raw data but faced limitations due to noise and class imbalance. To address these, preprocessing was introduced: Pix2Pix GAN for denoising and augmentation, and resampling with class weighting to balance classes. The models were retrained on the processed data, followed by the development of a hybrid CNN–LSTM framework combining convolutional feature extraction with sequential learning. Results: The hybrid model achieved the best performance, with 97.89% accuracy and a macro F1-score of 0.89. Preprocessing significantly improved recognition of minority classes, highlighting the role of GAN-based denoising and balanced training. Conclusion: A hybrid CNN–LSTM with GAN preprocessing provides a reliable and robust framework for ECG image classification, offering improved generalization and the potential to complement or enhance traditional signal-based ECG analysis in clinical practice.