Hierarchical deep learning with Generative Adversarial Network for automatic cardiac diagnosis from ECG signals

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Cardiac disease remains a leading cause of mortality in the US.
• Electrocardiogram (ECG) analysis is crucial for detecting cardiac abnormalities.
• Effective analytical models are needed to fully utilize ECG data for reliable heart disease detection.

Purpose of the Study:

• To propose a novel two-level hierarchical deep learning framework for ECG signal analysis.
• To enhance the accuracy of heart disease detection and arrhythmia identification.
• To address challenges of data scarcity and imbalance in ECG datasets.

Main Methods:

• Developed a two-level hierarchical deep learning framework incorporating Generative Adversarial Network (GAN).
• First level: Memory-Augmented Deep AutoEncoder with GAN (MadeGAN) for anomaly detection.
• Second level: Transfer learning and multi-branching architecture for robust multi-class arrhythmia classification.

Main Results:

• The proposed framework demonstrated superior performance in ECG signal analysis.
• Achieved high accuracy in differentiating normal from abnormal ECG signals.
• Outperformed existing methods in identifying various types of cardiac arrhythmias using the MIT-BIH database.

Conclusions:

• The proposed deep learning framework offers a significant advancement in ECG analysis for cardiac disease detection.
• The model effectively handles data limitations and imbalance issues.
• This approach holds promise for improving the timely diagnosis and treatment of heart conditions.