Generative adversarial networks in electrocardiogram synthesis: Recent developments and challenges.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Machine Learning

Background:

• Deep neural network classifiers for electrocardiograms (ECGs) require substantial data.
• Imbalanced datasets hinder classifier training, necessitating data augmentation techniques.
• Generative Adversarial Networks (GANs) offer a method for generating synthetic ECG data.

Purpose of the Study:

• To review existing literature on synthetic ECG signal generation using GANs.
• To provide a comprehensive overview of GAN architectures, quality evaluation metrics, and classification performance.
• To identify challenges and future directions in GAN-based ECG data augmentation.

Main Methods:

• Systematic literature review of 30 publications from 2019-2022 across three databases.
• Analysis of studies based on their use of quality evaluation metrics and classification performance.
• Categorization of employed quality evaluation metrics for synthetic ECG signals.

Main Results:

• A wide variety of 20 quality evaluation metrics were identified.
• Classification performance improved significantly, with accuracy increasing by 7%-98% and sensitivity by 6%-97% when using GAN-augmented data.
• Discrepancies were observed, with some studies focusing solely on quality metrics, others on classification, and some on both.

Conclusions:

• Synthetic ECG generation using GANs is a promising approach for augmenting imbalanced datasets.
• Consistent and standardized quality evaluation of GAN-generated ECG signals remains a significant challenge.
• Future research should focus on establishing a gold standard for quality evaluation metrics for GANs in ECG applications.