A convolutional attention transformer network for ECG beat classification

Abstract

This work aims to improve automated ECG beat classification in accordance with the Association for the Advancement of Medical Instrumentation (AAMI)-recommended classes by learning both local morphological patterns and longer-range temporal dependencies from ECG segments. We develop a complete analysis pipeline using the Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) Arrhythmia Database, including beat segmentation around detected R-peaks, standard pre-processing, and evaluation under a consistent train/validation/test protocol. The proposed Convolutional–Attention–Transformer (CAT) Network integrates convolutional feature extraction for local waveform morphology, an attention mechanism to emphasize diagnostically salient regions, and transformer-based sequence modeling to capture contextual dependencies. It is benchmarked against representative machine learning and ensemble baselines such as k-nearest neighbor (k-NN), support vector machine (SVM), Random Forest (RF), XGBoost, and CatBoost. Experimental results demonstrate that the CAT Network achieves superior or competitive performance across standard metrics, particularly improved macro-averaged performance and minority-class sensitivity, indicating better balance under class imbalance. We also provide a confusion matrix and class-wise analyses to highlight error patterns across AAMI classes and explain the practical impact of the proposed design choices. These findings suggest that the CAT Network can serve as an effective and reproducible framework for AAMI-compliant ECG beat classification, supporting downstream decision support and large-scale screening, while motivating future work on cross-database generalization, computational optimization for edge deployment, and enhanced interpretability for clinical adoption.