Automated Arrhythmia Detection from Raw DualChannel ECG: Integrating ۱D- CNN and Autoencoder Architectures with Machine Learning Algorithms

Cardiovascular diseases require accurate Electrocardiogram (ECG) analysis, but manual interpretation is subjective and traditional automated methods struggle with dataset imbalances. To address these limitations, this study proposes a comprehensive two-stage hybrid framework for automated arrhythmia detection using raw dual-channel ECG signals. To resolve the severe class skewness in the MIT-BIH Arrhythmia Database, a hybrid resampling strategy combining the Synthetic Minority Oversampling Technique (SMOTE) and Random Undersampling (RUS) is implemented to create a perfectly balanced training environment. In the first stage, One-Dimensional Convolutional Neural Networks (۱D-CNNs) and Autoencoders (AEs) of varying depths are evaluated to automatically extract robust spatial representations. In the second stage, these latent features are categorized using machine learning algorithms including SVM, KNN, Gradient Boosting (GB), and Random Forest. Experimental evaluations reveal a critical architectural divergence: while increased depth in supervised ۱D-CNNs significantly enhances feature discriminability, deeper unsupervised AEs suffer from information loss. Ultimately, pairing the deep ۱D-CNN feature extractor with the GB classifier achieved an outstanding peak accuracy, precision, recall, and macro F۱-score of ۹۹.۳۳%. This framework provides a highly interpretable and reliable diagnostic pipeline for clinical monitoring.

کلیدواژه ها:

Automated Arrhythmia Detection ، Dual-Channel ECG ، ۱D-CNN ، Machine Learning Classifiers ، Hybrid Resampling