A Novel Approach for Arrhythmia Detection from Electrocardiogram Signal using cGAN-Markov Chain and Hybrid Recurrent Neural Networks

Deep learning methods have shown early progress in analyzing complicated electrocardiogram signals, particularly in tasks like heartbeat classification and arrhythmia detection. Analyzing health-related data has always been a complex challenge. In this study, a novel method for arrhythmia classification that combines a hybrid unidirectional and bidirectional Recurrent Neural Network (RNN) with a multilayered dilated convolutional neural network (CNN) is implemented. The model is trained and validated on the MIT-BIH arrhythmia dataset, showing significant improvements in both performance and interpretability by integrating features with the dilated CNN framework. For the reduction of noise present in the dataset, Discrete Wavelet Transform (DWT) is applied using the symlet function. To counteract the class imbalance issue, Generative Adversarial Network (GAN) is utilized in conjunction with a Markov chain to synthesize new data, recreating underrepresented signal patterns. The hybrid architecture proposed here capitalizes on the strengths of both multilayered dilated CNNs and different types of RNN units, specifically bidirectional units such as BiGRU (Bidirectional Gated Recurrent Units) and BiLSTM (Bidirectional Long Short-Term Memory), as well as unidirectional units like GRU and LSTM. These layers work together to extract and combine desirable features. The experimental results indicate that the model achieves impressive metrics, including an overall accuracy of 98.35%. A key outcome of this research is the significant reduction in computation time and effective mitigation of class imbalance through integrating diverse and realistic synthetic data to enhance the training process.