A hybrid deep feature extraction and ensemble learning framework for robust multi-class ECG image classification

Abstract ECG signals, the electrical activity of the heart, play a vital role in early identification of cardiovascular diseases (CVDs), which are currently the leading cause of death across the world. However, manual detection of ECG images can be time-consuming, prone to human error, and difficult to scale, particularly in resource-constrained environments. To address these challenges, this work proposes a hybrid classification framework to better analyze ECG images, improving reliability and efficiency. The method integrates deep learning with feature optimization and ensemble learning. An EfficientNet-B0 model is used to extract significant features from ECG images, initialized with weights trained on ImageNet data. High-dimensional features are compressed using a lightweight autoencoder to remove data redundancy while retaining significant feature pat-terns. Features obtained after compression are then applied to classify using two different machine learning algorithms: (i) Support Vector Machine with RBF kernel and (ii) Random Forest classifier. Finally, a majority-voting ensemble integrates the predictions of the CNN, SVM, and RF classifiers. Our system is tested on the Mendeley ECG Images Dataset for four classes (Normal, Myocardial Infarction (MI), Abnormal Heartbeat, and History of MI, achieving an accuracy of 92.09%, precision of 91.3%, recall of 91.1%, and F1 of 91.15%. Therefore, the proposed framework demonstrates promising classification performance with favorable computational characteristics on the evaluated dataset, highlighting its potential for computer-aided ECG image classification; however, external and prospective clinical validation is required before its clinical applicability can be established.