Real-valued (Medical) Time Series Generation with Recurrent Conditional\n GANs

Abstract

Generative Adversarial Networks (GANs) have shown remarkable success as a\nframework for training models to produce realistic-looking data. In this work,\nwe propose a Recurrent GAN (RGAN) and Recurrent Conditional GAN (RCGAN) to\nproduce realistic real-valued multi-dimensional time series, with an emphasis\non their application to medical data. RGANs make use of recurrent neural\nnetworks in the generator and the discriminator. In the case of RCGANs, both of\nthese RNNs are conditioned on auxiliary information. We demonstrate our models\nin a set of toy datasets, where we show visually and quantitatively (using\nsample likelihood and maximum mean discrepancy) that they can successfully\ngenerate realistic time-series. We also describe novel evaluation methods for\nGANs, where we generate a synthetic labelled training dataset, and evaluate on\na real test set the performance of a model trained on the synthetic data, and\nvice-versa. We illustrate with these metrics that RCGANs can generate\ntime-series data useful for supervised training, with only minor degradation in\nperformance on real test data. This is demonstrated on digit classification\nfrom 'serialised' MNIST and by training an early warning system on a medical\ndataset of 17,000 patients from an intensive care unit. We further discuss and\nanalyse the privacy concerns that may arise when using RCGANs to generate\nrealistic synthetic medical time series data.\n

Ähnliche Arbeiten