ECG Signal Generation Using Variable β-Conditional Variational Autoencoder.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Automated cardiovascular disease diagnosis using machine learning (ML) is promising but hindered by costly ECG signal annotation, insufficient data, and class imbalance.
• Generative models like GANs have been applied to ECG synthesis, but they involve complex training and may exacerbate class imbalance.

Purpose of the Study:

• To develop an efficient and effective method for generating synthetic ECG signals to augment limited clinical datasets.
• To improve the generalization and performance of ML-based arrhythmia classification models.

Main Methods:

• A conditional variational autoencoder (CVAE)-based method was proposed for ECG signal generation.
• The CVAE approach simplifies the generation process and efficiently handles multiple ECG classes.
• A variable beta parameter was used to balance the fidelity and diversity of generated signals by adjusting KL divergence.

Main Results:

• The CVAE model successfully generated high-quality synthetic ECG signals.
• The generated ECG signals improved the accuracy of arrhythmia classification.
• The method demonstrated potential for effective ECG data augmentation, addressing sample insufficiency and class bias.

Conclusions:

• Conditional variational autoencoder-based ECG generation is a viable strategy for augmenting limited clinical datasets.
• This approach can optimize arrhythmia diagnostic performance by mitigating data scarcity and class distribution bias.
• The proposed method offers a simplified and efficient alternative to existing generative models for ECG synthesis.