Invertible Conditional Generative Adversarial Networks to Effectively Generate Myocardial Infarction from Normal ECG.

Area of Science:

• Biomedical Signal Processing
• Deep Learning
• Cardiovascular Research

Background:

• Deep learning-based style transfer shows promise for generating synthetic physiological signals.
• Existing methods like CycleGAN require multiple models for different transformations.
• There is a need for more efficient and flexible style transfer techniques in biomedical applications.

Purpose of the Study:

• To explore the use of Invertible Conditional Generative Adversarial Networks (IcGANs) for style transfer in 12-lead ECG.
• To transform ECG heartbeats from normal sinus rhythm to myocardial infarction (inferior and antero-septal).
• To compare the efficiency and performance of IcGAN against CycleGAN for ECG style transfer.

Main Methods:

• Training an IcGAN and an encoder for ECG style transfer.
• Utilizing the PTB-XL dataset from Physionet for training.
• Assessing generated ECG signal quality through visual inspection, GAN scores, and ST-segment amplitude analysis.

Main Results:

• IcGAN effectively captured myocardial infarction features while preserving original ECG characteristics.
• Generated ECG signals exhibited clinically meaningful variations.
• IcGAN demonstrated superior efficiency and performance compared to CycleGAN in similar architectures.

Conclusions:

• IcGAN offers a more efficient and flexible framework for biomedical signal style transfer compared to CycleGAN.
• The technique shows potential for domain adaptation and synthetic data generation for rare conditions.
• Controlled ECG feature modification using IcGAN can enhance model generalization for personalized medicine.