Abstract
The performance of data-hungry Deep Learning (DL) models in the domain of healthcare is severely limited due to the lack of large quantities of good-quality data. The challenge of data availability leads to class imbalance problems which are exacerbated for rare medical conditions. This also leads to problems of model drift and lack of generalization in the development of deployable healthcare models. Synthetic data generators have been proposed to mitigate these challenges as a data augmentation strategy. However, most of the generative algorithms are resource intensive, and their performance is restricted, especially for structured signals. The training of popular generative algorithms is challenging to converge for a single task-specific generation, often leading to mode collapse and requiring large data to generate from random noise. In this work, we propose a convolutional Conditional Variational Autoencoder (CVAE) architecture to generate Electrocardiogram signals of a variable number of beats for 2 classes of arrhythmias as images. We present experimental results using two arrhythmia classes, namely, Normal Sinus Rhythm (NSR) and Atrial Fibrillation (AFIB) for individual beats (200 data points) and rhythm signals (3600 data points, 6-10 beats), based on the available dataset restrictions for training and testing our proposed generative model. Our proposed generative model is trained with a mix of rhythms and individual beats across different classes of ECG arrhythmias. Our approach is independent of the number of classes and can be extended to generate single ECG beats or ECG rhythms (variable lengths) for multiple classes without training individually for each type of arrhythmia or for specific lengths. Our generated synthetic data improves the performance of rhythm classifiers by 32.74% testing accuracy respectively on real-world test data.