GAEA-Net: Generating Activity-Enriched Abnormal ECGs via Adversarial Network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiovascular Monitoring

Background:

• Wearable devices necessitate non-prescription ECG diagnosis during physical activity.
• Existing ECG abnormality models perform poorly during exercise due to lack of relevant data.
• Abnormal ECGs are typically recorded at rest, limiting model generalizability to active states.

Purpose of the Study:

• To develop a novel method for generating synthetic, activity-enriched ECG data.
• To improve the performance of ECG abnormality classification models during physical activity.
• To address the scarcity of abnormal ECG data captured during exercise.

Main Methods:

• Proposed GAEA-Net model for ECG synthesis.
• Utilized healthy individuals' exercise ECGs and resting-state abnormal ECGs.
• Generated synthetic activity-enriched ECGs for model training and testing.

Main Results:

• Achieved average performance improvements across five datasets: 1.3% Accuracy, 1.3% F1-score, 0.9% AUROC, 1.6% MCC, and 1.4% Cohen's Kappa.
• Clinical Turing test confirmed high fidelity of synthesized ECGs.
• Cardiologists showed comparable diagnostic accuracy on synthetic vs. real ECGs (55.7% vs. 54.9%).

Conclusions:

• GAEA-Net effectively synthesizes realistic ECG data for physical activity scenarios.
• The approach enhances the reliability of wearable ECG monitoring during exercise.
• Synthetic data generation is a viable strategy to overcome data limitations in active ECG analysis.