Classification of ECG signals using 1D-2D transformation and convolutional neural networks (CNN)

Abstract

This thesis investigates automatic ECG signal classification using CNNs by transforming 1D signals into 2D spectrogram images. It addresses the need for accurate, scalable arrhythmia detection with deep learning approaches. A complete pipeline, including preprocessing, transformation, CNN training, and evaluation, was developed and tested. Experiments examined architecture choices, training epochs, and input resolution impacts on performance. Results show CNNs achieve high accuracy, particularly on normal beats. Challenges like class imbalance and overfitting remain and limit generalization. Future work includes advanced architectures, data augmentation, and larger dataset validation.