ECGNet: High-Precision ECG Classification Using Deep Learning and Advanced Activation Functions

The electrocardiogram (ECG) is an important tool in diagnosing heart diseases. In this study, we introduce ECGNet a customized deep learning model that utilizes advanced activation functions and modified classifiers to enhance ECG classification. ECGNet’s performance is compared with three benchmark deep learning models. Image transformation and optimization are performed using the LabelIMG tool. ECGNet unifies an embedding layer for feature extraction, followed by two dense layers for disease detection. The proposed AELU (Advanced Exponential Linear Unit) activation, with a learning rate of 0.001, addresses the vanishing gradient issues in ReLU and Sigmoid function, allowing the absolute values of the small negative gradient to be preserved, thereby introducing higher nonlinearity in the output and higher model flexibility. Additionally, the Modified SparseMax classifier is employed to enhance the classification by mapping logits to sparse probability distributions. The ECGNet model achieves 99% accuracy, outperforming the existing methods. These matrices reflected their ability to efficiently classify the five different classes. The proposed model also compared with most trending pretrained models such as LSTM, GRU and TabNet to validate the effectiveness.