[Paper Review] Synthetic ECG Signal Generation using Probabilistic Diffusion Models

Edmond Adib, Amanda Fernandez+2 more|arXiv (Cornell University)|Mar 4, 2023

ECG Monitoring and Analysis4 citations

TL;DR

This paper proposes a novel pipeline to generate synthetic 1D ECG signals using 2D Improved Denoising Diffusion Probabilistic Models (DDPM), by embedding 1D ECG time series into 3-channel 2D space via GASF/GADF and MTF. Despite DDPM's superior image generation performance, results show WGAN-GP consistently outperforms DDPM in generating realistic, distributionally accurate, and classifiable synthetic ECG beats across all metrics.

ABSTRACT

Deep learning image processing models have had remarkable success in recent years in generating high quality images. Particularly, the Improved Denoising Diffusion Probabilistic Models (DDPM) have shown superiority in image quality to the state-of-the-art generative models, which motivated us to investigate their capability in the generation of the synthetic electrocardiogram (ECG) signals. In this work, synthetic ECG signals are generated by the Improved DDPM and by the Wasserstein GAN with Gradient Penalty (WGAN-GP) models and then compared. To this end, we devise a pipeline to utilize DDPM in its original $2D$ form. First, the $1D$ ECG time series data are embedded into the $2D$ space, for which we employed the Gramian Angular Summation/Difference Fields (GASF/GADF) as well as Markov Transition Fields (MTF) to generate three $2D$ matrices from each ECG time series, which when put together, form a $3$-channel $2D$ datum. Then $2D$ DDPM is used to generate $2D$ $3$-channel synthetic ECG images. The $1D$ ECG signals are created by de-embedding the $2D$ generated image files back into the $1D$ space. This work focuses on unconditional models and the generation of \emph{Normal Sinus Beat} ECG signals exclusively, where the Normal Sinus Beat class from the MIT-BIH Arrhythmia dataset is used in the training phase. The \emph{quality}, \emph{distribution}, and the \emph{authenticity} of the generated ECG signals by each model are quantitatively evaluated and compared. Our results show that in the proposed pipeline and in the particular setting of this paper, the WGAN-GP model is consistently superior to DDPM in all the considered metrics.

Motivation & Objective

• To investigate the feasibility of using 2D probabilistic diffusion models (DDPM) for generating synthetic 1D ECG signals.
• To compare the quality, distribution, and authenticity of synthetic ECG beats generated by DDPM versus WGAN-GP.
• To evaluate whether synthetic ECG signals can effectively augment real data for downstream classification tasks.
• To establish a novel data embedding and de-embedding pipeline for adapting 2D diffusion models to 1D time series ECG data.

Proposed method

• 1D ECG time series are transformed into polar coordinates and embedded into 2D space using Gramian Angular Summation/Difference Fields (GASF/GADF) and Markov Transition Fields (MTF).
• Three 2D matrices (GASF, GADF, MTF) are combined into a single 3-channel 2D image-like representation for each ECG beat.
• The Improved DDPM model is trained and sampled on the 2D embedded data to generate synthetic 2D ECG images.
• Generated 2D data are de-embedded back into 1D ECG signals using inverse transformations of the embedding process.
• Three hyperparameter settings of DDPM are evaluated, and results are compared against WGAN-GP-generated beats and real ECG data.
• Evaluation uses DTW, Fréchet Distance, MMD for distributional similarity, and classification metrics (AUC, precision) with a ResNet34 classifier.

Experimental results

Research questions

• RQ1 Can 2D diffusion models like DDPM be effectively adapted to generate realistic 1D ECG signals through 2D embedding?
• RQ2 How does the performance of DDPM in synthetic ECG generation compare to that of WGAN-GP in terms of signal quality and distributional fidelity?
• RQ3 To what extent can DDPM-generated ECG signals replace real data in downstream classification tasks?
• RQ4 Does the 2D embedding and de-embedding pipeline preserve the morphological and temporal characteristics of real ECG signals?

Key findings

• WGAN-GP consistently outperformed all three DDPM hyperparameter settings in generating ECG beats that are closer to real beats in terms of signal quality and distribution.
• The average DTW and Fréchet distances between generated and real beats were significantly lower for WGAN-GP than for DDPM, indicating superior structural similarity.
• The Maximum Mean Discrepancy (MMD) values showed that WGAN-GP-generated beats had a distribution much closer to real data than DDPM-generated beats.
• In classification tasks using a ResNet34 model, WGAN-GP-generated beats achieved higher average precision, AUC-PR, and AUC-ROC scores than DDPM-generated beats.
• Despite DDPM's superior image generation performance, the 12,288-parameter generation process in 2D space led to a less optimized loss landscape compared to the 64-parameter 1D WGAN-GP, resulting in poorer ECG signal fidelity.
• The study demonstrates that direct 1D modeling (as in WGAN-GP) is more effective than 2D embedding for 1D ECG signal generation, even when using state-of-the-art 2D diffusion models.