From Mathematics to Machine Learning: The Evolution of ECG Preprocessing and Feature Engineering

The Electrocardiogram (ECG) is an essential non-invasive diagnostic modality that captures the bioelectric potentials generated by the heart to detect critical cardiovascular conditions such as arrhythmias, myocardial infarction, and ischemia. However, raw cardiac signals acquired from the skin surface are inherently weak and heavily corrupted by physiological and environmental artifacts, making direct clinical evaluation unreliable. To achieve an accurate and definitive medical diagnosis, implementing a robust signal processing pipeline is an absolute prerequisite before any analytical decision-making can occur. This comprehensive review systematically analyzes the complete processing continuum, encompassing data acquisition, noise removal, feature extraction, and post-processing optimization. We examine classical mathematical approaches, including digital bandpass filtering, the Pan-Tompkins algorithm, and the Discrete Wavelet Transform (DWT) for precise waveform delineation. Furthermore, the paradigm shift toward data-driven techniques is thoroughly evaluated, highlighting Machine Learning (ML) algorithms like Principal Component Analysis (PCA) and the Genetic Algorithm (GA) for feature selection, alongside advanced Deep Learning (DL) models such as Convolutional Neural Network (CNN) architectures and Signal-to-Image (S۲I) transformations for automated feature learning. Ultimately, this review demonstrates how structured signal processing provides pristine data and optimized feature representations required for reliable computer-aided cardiac diagnosis.

کلیدواژه ها:

Electrocardiogram ، Signal Processing ، Feature Extraction ، Deep Learning ، Noise Removal ، Machine Learning ، Dimensionality Reduction