Synthetic ECG signal generation using generative neural networks.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiogram (ECG) datasets are often imbalanced due to rare abnormal cases.
• Privacy regulations restrict the use of real patient ECG data, limiting dataset size.
• Machine learning models for automatic ECG diagnosis require balanced datasets for optimal performance.

Purpose of the Study:

• To evaluate the synthetic ECG generation capabilities of five Generative Adversarial Network (GAN) models.
• To compare the performance of these GAN models in generating normal cardiac cycles.
• To assess the utility of synthetic ECG data for augmenting imbalanced datasets and improving classification accuracy.

Main Methods:

• Generated synthetic normal cardiac cycles using five different GAN models.
• Quantitatively measured performance using Dynamic Time Warping (DTW), Fréchet, and Euclidean distances.
• Proposed and applied five novel evaluation methods, including threshold, accepted beat, and productivity rate, for systematic comparison.

Main Results:

• All tested GAN models demonstrated the ability to generate acceptable heartbeats with high morphological similarity.
• BiLSTM-DC GAN and WGAN visually produced more statistically acceptable beats.
• Classic GAN achieved the highest productivity rate at 72%.
• Augmenting an imbalanced dataset with synthetic ECGs significantly improved the performance of the ECGResNet34 classifier.

Conclusions:

• GANs show potential for generating synthetic ECG data to augment imbalanced datasets.
• Specific GAN architectures (BiLSTM-DC GAN, WGAN) excel in generating visually acceptable beats.
• Synthetic ECG data augmentation can significantly enhance the performance of automated cardiac diagnostic models.