ECG Signal Generation Using Variable β-Conditional Variational Autoencoder

Machine learning methods offer a shortcut for automated cardiovascular disease diagnosis. However, the high cost of ECG signal annotation, along with insufficient labeled data and class imbalance, severely limits the generalization of these methods in real-world applications. Recently, generative models, especially generative adversarial networks and their variants, have made significant progress in image and natural language generation tasks, and have been applied to electrocardiogram synthesis. However, adversarial networks require complex training and typically generate data for each ECG class separately, which not only complicates the process but also may lead to class imbalance issues. To address these problems, we propose a conditional variational autoencoderbased ECG generation method. This approach simplifies the generation process, handles multiple ECG classes efficiently, and uses a variable β parameter to adjusts the KL divergence weight, balancing the fedility and diversity of the generated signals. Experimental results show that the conditional variational autoencoder model can generate high-quality ECG signals, improving arrhythmia classification accuracy, and demonstrating its potential in ECG data augmentationClinical Relevance- This study highlights the potential of conditional variational autoencoder-based ECG signal generation as an effective strategy for augmenting limited clinical datasets, thereby optimizing arrhythmia diagnostic performance by alleviating sample insufficiency and class distribution bias.