Electronic health records (EHRs) chart patients' interactions with the health system and contain critical information for improving services and supporting research. Data from these systems are routinely incorporated into machine learning and statistical models for clinical decision support on diagnostic and prognostic predictions, as well as for monitoring health and evaluating treatment response [1]. However, access to large-scale EHR datasets is challenging and governed by strict regulations on privacy and security (e.g. HIPAA and GDPR), meaning that many models are based on unicentric data with a high risk of poor generalizability [2]. Traditional approaches for anonymization can be complex and costly, often compromising the data's statistical integrity and failing to provide robust privacy guarantees [3, 4]. The use of synthetic data is thus emerging as a promising solution for optimizing the trade-off between privacy and statistical utility [5, 6]. Generative models, particularly Generative Adversarial Networks (GANs) [7], have shown great potential in producing distribution-preserving synthetic EHR data.

Time series data generation has drawn increasing attention in recent years. Several generative adversarial network (GAN) based methods have been proposed to tackle the problem usually with the assumption that the targeted time series data are well-formatted and complete. However, real-world time series (RTS) data are far away from this utopia, e.g., long sequences with variable lengths and informative missing data raise intractable challenges for designing powerful generation algorithms. In this paper, we propose a novel generative framework for RTS data - RTSGAN to tackle the aforementioned challenges. RTSGAN first learns an encoder-decoder module which provides a mapping between a time series instance and a fixed-dimension latent vector and then learns a generation module to generate vectors in the same latent space. By combining the generator and the decoder, RTSGAN is able to generate RTS which respect the original feature distributions and the temporal dynamics. To generate time series with missing values, we further equip RTSGAN with an observation embedding layer and a decide-and-generate decoder to better utilize the informative missing patterns. Experiments on the four RTS datasets show that the proposed framework outperforms the previous generation methods in terms of synthetic data utility for downstream classification and prediction tasks.