A performance analysis of convolutional autoencoder modified WaveGAN architectures for realistic 12 lead electrocardiogram synthesis

Abstract

The burgeoning necessity for copious and diverse electrocardiogram (ECG) datasets for deep learning applications in clinical diagnostics has been impeded by the confidential nature of patient data. Related works have shown the effectiveness of additional data generation in enhancing the deep learning models’ performance. This research study introduces a novel Convolutional Autoencoder-WaveGAN (CAE-WaveGAN) technique for generating synthetic but realistic 12-lead ECG images to address data scarcity. The proposed model leverages a convolutional autoencoder for efficient feature extraction from ECG signals, which is then utilized by a WaveGAN generator to synthesize high-fidelity ECG images. The method provides a practical solution for expanding ECG training datasets where patient privacy constraints and data scarcity limit the development of robust deep learning models for cardiovascular diagnosis. We conducted a comprehensive performance analysis of various CAE-WaveGAN configurations through an ablation study on the CODE-15% dataset. Experimental results demonstrate that CAE-WaveGAN achieves superior performance across all evaluation metrics, with 19.8% improvement in PSNR and 59.3% enhancement in SSIM compared to baseline methods. Our findings from the ablation study reveal that the optimal CAE-WaveGAN architecture significantly surpasses the traditional WaveGAN in terms of stability and loss metrics, offering a promising solution for generating realistic ECG data in clinical machine learning applications.