Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffECG: A Versatile Probabilistic Diffusion Model for ECG Signals Synthesis

Nour Neifar, Achraf Ben-Hamadou, Afef Mdhaffar, Mohamed Jmaiel

TL;DR

This work tackles data scarcity, missing data, and forecasting in ECG-based cardiovascular diagnostics by introducing DiffECG, a versatile conditional diffusion model. It leverages a spectrogram-based conditioning scheme and a task-switching encoding to unify heartbeat generation, imputation, and forecasting within a single DDPM framework, operating over $T$ diffusion steps. Empirical results on the MIT-BIH dataset show competitive generation quality relative to GAN baselines and clear improvements in imputation and forecasting, while also enhancing downstream classifier performance. The approach demonstrates practical impact for ECG data augmentation and restoration, with potential extensions to multi-lead ECG and integration with adversarial training.

Abstract

Within cardiovascular disease detection using deep learning applied to ECG signals, the complexities of handling physiological signals have sparked growing interest in leveraging deep generative models for effective data augmentation. In this paper, we introduce a novel versatile approach based on denoising diffusion probabilistic models for ECG synthesis, addressing three scenarios: (i) heartbeat generation, (ii) partial signal imputation, and (iii) full heartbeat forecasting. Our approach presents the first generalized conditional approach for ECG synthesis, and our experimental results demonstrate its effectiveness for various ECG-related tasks. Moreover, we show that our approach outperforms other state-of-the-art ECG generative models and can enhance the performance of state-of-the-art classifiers.

DiffECG: A Versatile Probabilistic Diffusion Model for ECG Signals Synthesis

TL;DR

This work tackles data scarcity, missing data, and forecasting in ECG-based cardiovascular diagnostics by introducing DiffECG, a versatile conditional diffusion model. It leverages a spectrogram-based conditioning scheme and a task-switching encoding to unify heartbeat generation, imputation, and forecasting within a single DDPM framework, operating over diffusion steps. Empirical results on the MIT-BIH dataset show competitive generation quality relative to GAN baselines and clear improvements in imputation and forecasting, while also enhancing downstream classifier performance. The approach demonstrates practical impact for ECG data augmentation and restoration, with potential extensions to multi-lead ECG and integration with adversarial training.

Abstract

Within cardiovascular disease detection using deep learning applied to ECG signals, the complexities of handling physiological signals have sparked growing interest in leveraging deep generative models for effective data augmentation. In this paper, we introduce a novel versatile approach based on denoising diffusion probabilistic models for ECG synthesis, addressing three scenarios: (i) heartbeat generation, (ii) partial signal imputation, and (iii) full heartbeat forecasting. Our approach presents the first generalized conditional approach for ECG synthesis, and our experimental results demonstrate its effectiveness for various ECG-related tasks. Moreover, we show that our approach outperforms other state-of-the-art ECG generative models and can enhance the performance of state-of-the-art classifiers.
Paper Structure (14 sections, 4 equations, 4 figures, 5 tables)

This paper contains 14 sections, 4 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Deep generative models applied to ECG
  4. Principle of diffusion models
  5. Proposed Approach
  6. Experimental evaluation
  7. Dataset
  8. Training settings
  9. Experiments and results
  10. Quantitative evaluation
  11. Classification evaluation
  12. Evaluation metrics
  13. Qualitative evaluation
  14. Conclusion

Figures (4)

  • Figure 1: The general principle of the proposed approach.
  • Figure 2: Examples of synthetic heartbeats for classes N, V, and F classes obtained from our generation approach (a), nour022Leveraging (b), nour2021Disentangling (c), goodfellow2014generative (d) and delaney2019synthesis (e). The gray background represents the distribution of the real dataset, while the red signals depict the generated heartbeats.
  • Figure 3: Examples of synthetic heartbeats for classes N, V, and F classes obtained from our generation approach (a), LSTM (b), VAE (c), and xu2022pulseimpute (d). The gray background represents the distribution of the real dataset, while the red portions depict the completed portions of the heartbeats.
  • Figure 4: Examples of synthetic heartbeats for classes N, V, and F classes obtained from our generation approach (a), LSTM (b), VAE (c), and xu2022pulseimpute (d). The gray background represents the distribution of the real dataset, while the red and green portions depict the completed and ground-truth the heartbeats.