Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Conditional Electrocardiogram Generation Using Hierarchical Variational Autoencoders

Ivan Sviridov, Konstantin Egorov

TL;DR

This study introduces cNVAE-ECG, a conditional Nouveau VAE tailored to generate high-quality 12-lead ECG signals of standard 10-second duration. By leveraging a hierarchical VAE framework and physiological lead relationships, it enables conditional synthesis across multiple pathologies and integrates a pathology embedding into the generative hierarchy. Empirical results show that 1D cNVAE-ECG data augmentation improves AUROC on binary pathology classification and transfer learning tasks, outperforming GAN-based competitors under the same training budget. The work suggests VAEs can rival GANs for ECG synthesis, offering stable training and effective data augmentation, with practical potential in mitigating data scarcity and class imbalance in clinical settings. Future directions include clinical validation and federated learning for real-world deployment.

Abstract

Cardiovascular diseases (CVDs) are disorders impacting the heart and circulatory system. These disorders are the foremost and continuously escalating cause of mortality worldwide. One of the main tasks when working with CVDs is analyzing and identifying pathologies on a 12-lead electrocardiogram (ECG) with a standard 10-second duration. Using machine learning (ML) in automatic ECG analysis increases CVD diagnostics' availability, speed, and accuracy. However, the most significant difficulty in developing ML models is obtaining a sufficient training dataset. Due to the limitations of medical data usage, such as expensiveness, errors, the ambiguity of labels, imbalance of classes, and privacy issues, utilizing synthetic samples depending on specific pathologies bypasses these restrictions and improves algorithm quality. Existing solutions for the conditional generation of ECG signals are mainly built on Generative Adversarial Networks (GANs), and only a few papers consider the architectures based on Variational Autoencoders (VAEs), showing comparable results in recent works. This paper proposes the publicly available conditional Nouveau VAE model for ECG signal generation (cNVAE-ECG), which produces high-resolution ECGs with multiple pathologies. We provide an extensive comparison of the proposed model on various practical downstream tasks, including transfer learning scenarios showing an area under the receiver operating characteristic (AUROC) increase up to 2% surpassing GAN-like competitors.

Conditional Electrocardiogram Generation Using Hierarchical Variational Autoencoders

TL;DR

This study introduces cNVAE-ECG, a conditional Nouveau VAE tailored to generate high-quality 12-lead ECG signals of standard 10-second duration. By leveraging a hierarchical VAE framework and physiological lead relationships, it enables conditional synthesis across multiple pathologies and integrates a pathology embedding into the generative hierarchy. Empirical results show that 1D cNVAE-ECG data augmentation improves AUROC on binary pathology classification and transfer learning tasks, outperforming GAN-based competitors under the same training budget. The work suggests VAEs can rival GANs for ECG synthesis, offering stable training and effective data augmentation, with practical potential in mitigating data scarcity and class imbalance in clinical settings. Future directions include clinical validation and federated learning for real-world deployment.

Abstract

Cardiovascular diseases (CVDs) are disorders impacting the heart and circulatory system. These disorders are the foremost and continuously escalating cause of mortality worldwide. One of the main tasks when working with CVDs is analyzing and identifying pathologies on a 12-lead electrocardiogram (ECG) with a standard 10-second duration. Using machine learning (ML) in automatic ECG analysis increases CVD diagnostics' availability, speed, and accuracy. However, the most significant difficulty in developing ML models is obtaining a sufficient training dataset. Due to the limitations of medical data usage, such as expensiveness, errors, the ambiguity of labels, imbalance of classes, and privacy issues, utilizing synthetic samples depending on specific pathologies bypasses these restrictions and improves algorithm quality. Existing solutions for the conditional generation of ECG signals are mainly built on Generative Adversarial Networks (GANs), and only a few papers consider the architectures based on Variational Autoencoders (VAEs), showing comparable results in recent works. This paper proposes the publicly available conditional Nouveau VAE model for ECG signal generation (cNVAE-ECG), which produces high-resolution ECGs with multiple pathologies. We provide an extensive comparison of the proposed model on various practical downstream tasks, including transfer learning scenarios showing an area under the receiver operating characteristic (AUROC) increase up to 2% surpassing GAN-like competitors.

Paper Structure

This paper contains 25 sections, 3 equations, 6 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. ECG Generation
  4. NVAE Background
  5. Variational Autoencoders
  6. NVAE Model
  7. Proposed Approach
  8. ECG generation
  9. Conditional generation
  10. Experimental Setup
  11. Estimating the quality of generated signals
  12. Enrichment of the training dataset for the binary classification task
  13. Enrichment of the pre-train dataset for the transfer learning
  14. Comparison with existing ECG generation methods for the same training time
  15. Evaluating how the quantity of initial data impacts the quality of generated data
  16. ...and 10 more sections

Figures (6)

  • Figure 1: Proposed cNVAE-ECG architecture.
  • Figure 2: Values of AUROC on the PTB-XL test set for each proportion using four different enrichment methods of both classes in the training dataset.
  • Figure 3: Values of AUROC on the PTB-XL test set for each proportion using four different enrichment methods of the Myocardial Infarction class in the training dataset.
  • Figure 4: Real (a) and generated by cNVAE-ECG (b) ECG signal for the Sinus rhythm class.
  • Figure 5: Real (a) and generated by cNVAE-ECG (b) ECG signal for Myocardial infarction.
  • ...and 1 more figures