Generation of Synthetic Dataset for Electronic Health Record (EHR) and Medical Images

Machine learning models require large quantities of high-quality data for training. In applications such as rare disease diagnosis, remote sensing, and agriculture, data unavailability is a cutting-edge constraint leading to sub-optimal model performance. In this paper, we explore the application of data augmentation by deep learning-based Generative Adversarial Networks (GANs) and Diffusion models for generating synthetic data sets for Electronic Health Records (EHRs) and medical images. Due to medical data privacy and limited access for high-resolution medical data, synthetic data generation is one potential direction towards training effective machine learning models. Our aim is to compare the synthetic medical data generation performance of the GAN and Diffusion model and choose the most effective solution. Utilizing synthetic data, our aim is to enhance model performance without compromising data privacy and enhancing training data availability for medical research. The proposed approach is extensible to broad ranges of applications in non-medical domains such as security, remote sensing, and anomaly detection. Experiments demonstrate that the use of synthetic datasets generated on the basis of GAN enhances the accuracy of the model and thus is an efficient approach for machine learning-based data augmentation.