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

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Medical Imaging

Background:

• Deep learning for cardiovascular diagnosis requires large, diverse electrocardiogram (ECG) datasets.
• Patient data confidentiality and scarcity impede the creation of comprehensive ECG datasets.
• Synthetic data generation can enhance deep learning model performance.

Purpose of the Study:

• To introduce a novel Convolutional Autoencoder-WaveGAN (CAE-WaveGAN) for generating synthetic 12-lead ECG images.
• To address the challenge of ECG data scarcity in clinical diagnostics.
• To improve the development of deep learning models for cardiovascular disease detection.

Main Methods:

• A Convolutional Autoencoder (CAE) was used for efficient feature extraction from ECG signals.
• A WaveGAN generator synthesized high-fidelity ECG images using extracted features.
• Ablation studies were performed on the CODE-15% dataset to analyze CAE-WaveGAN configurations.

Main Results:

• CAE-WaveGAN demonstrated superior performance across all evaluation metrics.
• Significant improvements were observed: 19.8% in Peak Signal-to-Noise Ratio (PSNR) and 59.3% in Structural Similarity Index Measure (SSIM).
• The optimal CAE-WaveGAN architecture showed improved stability and loss metrics compared to traditional WaveGAN.

Conclusions:

• CAE-WaveGAN offers a practical solution for expanding ECG datasets while preserving patient privacy.
• The method effectively generates realistic ECG data for clinical machine learning applications.
• This approach facilitates the development of more robust deep learning models for cardiovascular diagnosis.