ECG Synthesis with Neural ODE and GAN Models

Abstract

This paper uses Neural ODE (NODE) based models to generate continuous medical time series. We also introduce a new technique to design the generative adversarial network (GAN) with Neural ODE (NODE) based Generator and Discriminator. During the Performace evaluation of the proposed model, we find that data generated by a NODE-based generator is more continuous than traditional GAN. Therefore, the proposed GAN model becomes more robust. On the other hand, the traditional GAN model demonstrates unstable training and unsupervised learning, which often make it difficult to determine the quality of the result. For this work, we design NODE based models as a generator and a discriminator to make the GAN model datadriven. The data-driven approach helps us overcome the unstable training of the traditional GAN model and improve the quality of the result. We used different evaluation metrics to quantitatively assess generated synthetic data for real-world applications and data analysis. We also evaluate the existing GAN model and the proposed models to understand the comparative efficiency of medical data synthesis.