The automatic classification of arrhythmias from ECG signals transformed into images represents a major advance in computer vision for cardiology, enabling faster and more accurate diagnoses. This study presents a comparative analysis of the deep learning frameworks PyTorch and TensorFlow for arrhythmia classification, using the publicly available PTB Diagnostic ECG Database. ECG signals were segmented and labeled into two categories (Abnormal and Normal), generating 14,552 vectors with 187 positions each at 125 Hz, divided into training and testing sets (80/20). Different signal representations were evaluated, including one-dimensional (1D) and two-dimensional (2D) formats, with Recurrence Plot used for image transformation. Implemented models included multilayer Perceptron’s and convolutional neural networks (ResNet-18 and AlexNet). Performance was evaluated using various metrics, while computational efficiency was measured by training time and hardware utilization. The results showed that AlexNet in PyTorch achieved the best performance (99.55% accuracy, 99.49% precision, 99.38% recall, 99.44% F1 score, 0.9889 MCC), outperforming TensorFlow (99.17%, 98.97%, 98.97%, and 98.97%, 0.9784). The comparative analysis showed that PyTorch proved more efficient and accurate under identical conditions.
