Denoising/Synthesizing ECG Signal Data Classification (Arrhythmia) using GAN (Generative Adversarial Networks)

Abstract

Deep learning techniques have made importance in the analysis of complex ECG signals, particularly in the classification of heartbeats and the prediction of arrhythmias. Despite this, additional research needs to be carried out on the area of the analysis of health-related data. The difficulties with multilayered convolution neural network (CNN) models are addressed by this research's hybrid an in both directions and organized recurrent neural technique for classifying arrhythmias. In order to address uneven signal class, a synthesis signal is generated using a generative adversarial network (GAN). CNN with Bidirectional LSTM and Attention Mechanism is a proposed architecture that combines multilayered CNN with bidirectional LSTM and Attention units to provide fusion features. Finally, the signals are categorized using the Swish activation function. The proposed model is trained and validated using the dataset from the PhysioNet 2017. The efficiency of classification is significantly improved by the learned model. The experimental results demonstrate that the proposed CNN-BiLSTM-Attention model outperforms existing models with 99.29% accuracy, F1, and precision. In general, our combined CNN and BiLSTM focus model offers a high-performance automatic classification technique to detect arrhythmia and a cost-effective way to reduce ECG signal.