ECG Generation Based on Denoising Diffusion Probabilistic Models

Arrhythmia diseases seriously damage people's life and health, and identifying abnormal points in ECG signals by deep neural networks is an effective method for detecting arrhythmias.However, their accuracy is often limited by the biased data distribution of the training set, and a large number of labeled ECG signals are usually harder to obtain.Therefore, this paper proposes to synthesize virtual heart beat data by denoising diffusion probability model (DDPM) based on the MIT-BIH arrhythmia database to complement the real data.Three different methods for generating heartbeat signals are also used, which are (i) generating heartbeat signals directly, (ii) generating time-frequency maps of heartbeats and transforming them into heartbeat signals, and (iii) generating sub-signals of heartbeats and fusing them into complete heartbeat signals.Regarding the evaluation of the synthesized signals, we compare the advantages and disadvantages of the three heartbeat generation methods by four metrics: DTW, PCC, ED and KLD.The experimental results showed that the optimum values of 4.37, 17.09, 0.972 and 0.0094 were obtained for ED, DTW of method (i) and PCC, KLD of method (iii), respectively.