A COMPREHENSIVE REVIEW OF FEATURE-BASED AND DEEP LEARNING APPROACHES FOR ECG SIGNAL CLASSIFICATION

Cardiovascular diseases remain one of the main causes of death in the world; thus, accurate classification of Electrocardiogram (ECG) signal becomes an important role for CADs detection and treatment. Over the years, several feature-based machine learning approaches as well as models based on deep learning have been investigated to automate ECG interpretation and alleviate the limitations of manual interpretation. It provides a fundamental comparison between the traditional handcrafted feature-based techniques and the deep learning-based models in terms of their advantages and disadvantages and their suitability for clinical usage. Fundamental aspects such as ECG signal structure, benchmark datasets, preprocessing techniques and performance evaluation metrics are discussed. While traditional classifiers based on features such as SVM, Random Forest, and XGBoost are easy to map into a manageable computation, very interpretable, and efficient, they are not as accurate as deep learning models like CNN, LSTM, and hybrids that offer end-to-end learning. A brief comparative review of a few newly published studies is performed together with a previously published XGBoost-based study that evaluated the MIT-BIH data base and showed potential results. In addition, the review discusses important remaining challenges of data imbalance, signal noise, and model interpretability and proposes future research directions to further enhance the clinical practicality and accuracy of ECG classification systems. The objective of this work is to guide researchers and practitioners to more applicable and efficient solutions for cardiac monitoring in real-time.