ECG-DPM, a multi-channel spectrogram-based diffusion framework, outperformed the best baseline by 12.5% to 62.5% when generating 30-second ECGs.
A novel multi-channel spectrogram-based diffusion framework significantly improves the generation of synthetic ECGs compared to existing baselines, addressing data scarcity in deep learning models.
An electrocardiogram (ECG) records the electrical signals from the heart to assess various cardiovascular conditions. Deep learning methods have been proposed to model ECGs, but the insufficient availability of ECG data and annotations often hinders their performance. To address this challenge, this paper explores the latest data synthesis technique, i. e. , diffusion probabilistic models (DPMs), to enable the generation of an unlimited number of ECGs representing various cardiovascular conditions. In contrast to previous approaches that treat ECGs as time series data or convert them into power spectrograms, we introduce a novel multi-channel spectrogram-based diffusion framework. In our method, the diffusion model enhances generation diversity, while the multi-channel spectrogram preserves both magnitude and phase information, ensuring high fidelity in the reconstructed ECGs. Extensive experiments conducted on real-world ECG data demonstrate the superiority of our approach. Notably, our ECG-DPM outperforms the best baseline by a margin ranging from 1 2. 5 \% to 6 2. 5 \% when generating ECGs for 30 seconds.
Li et al. (Mon,) conducted a other in Cardiovascular conditions. ECG-DPM (multi-channel spectrogram-based diffusion framework) vs. Baseline approaches was evaluated on ECG generation performance. ECG-DPM, a multi-channel spectrogram-based diffusion framework, outperformed the best baseline by 12.5% to 62.5% when generating 30-second ECGs.