Multi-Scale Feature Extraction for ECG Beat Classification Using a CNN-Transformer Network with Imbalance Mitigation

Deep learning models have become popular for classifying electrocardiogram (ECG) heartbeats, yet traditional Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have limitations. CNNs struggle with long-range dependencies, while RNNs, though effective for temporal patterns, often suffer from slow convergence and high computational costs. Additionally, many methods inadequately address class imbalance, resulting in biased outcomes. We propose a hybrid model combining CNNs and Transformer Encoders to classify ECGs according to AAMI standards. This model leverages CNNs for local feature extraction and Transformer Encoders with self-attention to capture global dependencies, overcoming CNN and RNN drawbacks. We also apply the Synthetic Minority Over-sampling Technique (SMOTE) and class weighting to manage class imbalance effectively. Experimental results show improved classification accuracy and balanced performance over traditional methods.