Synthesis of Realistic ECG Waveforms Using a Composite Generative Adversarial Network for Classification of Atrial Fibrillation

In recent days, computer-aided diagnosis systems powered by artificial intelligence and machine learning have become an important part of medicine for assisting the doctors in critical decision making. They are popularly deployed in cardiology for early and automatic detection of various life-threatening diseases. However, a machine learning algorithm requires a large volume of training data to create the learning model which is an empirical problem in medical domain. Generating synthetic patient data has emerged as an important area of research to solve the issue. In this paper, we propose a novel Generative Adversarial Network (GAN) architecture using medical domain knowledge to create realistic Electrocardiogram (ECG) waveforms containing the signature of Atrial Fibrillation (AF), a common type of arrhythmia. Our composite architecture consists of a pair of GANs to simulate the disease-specific Heart Rate Variability (HRV) pattern and the unique signal morphology in the generated waveforms. The proposed architecture is applied on two public datasets for synthesis of AF-specific ECG to mitigate the class imbalance issue. Results show that the performance of existing AF classifiers significantly improves on both datasets by adding the synthetic data to the training set.