A Novel Approach for Long ECG Synthesis Utilize Diffusion Probabilistic Model

Deep neural networks (DNNs) have gained popularity and out-performed in ECG signal classification and detection challenges. However, the paucity of data for abnormal rhythms such as the atrioventricular block, ventricular tachycardia, or supraventricular tachycardia has restricted DNN performance. ECG synthesis has lately been regarded as a new efficient approach to supplement imbalanced training data to overcome DNN issues and replace an ECG simulator. Most previous ECG-creation studies have focused on constructing a simple QRS complex, which cannot depict the characteristics of an ECG rhythm composed of numerous QRS complexes. This study addresses lengthy ECG synthesis using the diffusion probabilistic model, a class of generative models that has attracted a lot of attention lately. Our proposal is efficient, flexible, and lightweight by using powerful DiffWave architecture as a baseline model. This approach has revealed the capability to generate long ECG signals similar to those obtained from patients. The Physionet dataset, including MIT-BIH Arrhythmia and MIT-BIH Atrial Fibrillation Databases, is applied to train and test our models. Consequently, we produced 10-second ECGs with several rhythms that were not visible in earlier investigations. The proposal outperforms the most satisfactory research in ECGs when comparing our high-quality synthesized data with actual signals on DNN models. Furthermore, our method also offers a novel technique to investigate electrocardiogram data without using an ECG simulator.