Multi-class electrocardiogram signal classification using deep learning and two-dimensional time-frequency codification

Abstract

The automated classification of cardiac arrhythmias in electrocardiogram (ECG) signals is critical for cardiovascular diagnosis. Several studies focus on the classification of ECG rhythms. However this task remains a challenge, as is often affected by significant class imbalances, long-term temporal dependencies and fine morphological features. Supervised deep learning methods have shown potential for signal classification due to high performance and scalability. This work presents a novel deep learning framework for multi-class arrhythmia classification based on a dual ResNet-18 architecture. The proposed method transforms ECG segments into two-dimensional time-frequency representations using Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) processed to extract spatial features and sequential temporal self-attention mechanism. Furthermore, to mitigate class imbalance Focal Loss objective function and weighted random sampler were integrated during optimization. Experimental validation using PhysioNet MIT-BIH Arrhythmia Database demostrates effectiveness, achieving an overall classification accuracy of 97.39 % and Macro-F1 score of 0.954 across the four classes. These results confirm that the dual time-frequency codification with temporal attention works as a robust and reliable alternative for automated ECG signal classification.