CorGAN: Correlation-Capturing Convolutional Generative Adversarial\n Networks for Generating Synthetic Healthcare Records

Abstract

Deep learning models have demonstrated high-quality performance in areas such\nas image classification and speech processing. However, creating a deep\nlearning model using electronic health record (EHR) data, requires addressing\nparticular privacy challenges that are unique to researchers in this domain.\nThis matter focuses attention on generating realistic synthetic data while\nensuring privacy. In this paper, we propose a novel framework called\ncorrelation-capturing Generative Adversarial Network (CorGAN), to generate\nsynthetic healthcare records. In CorGAN we utilize Convolutional Neural\nNetworks to capture the correlations between adjacent medical features in the\ndata representation space by combining Convolutional Generative Adversarial\nNetworks and Convolutional Autoencoders. To demonstrate the model fidelity, we\nshow that CorGAN generates synthetic data with performance similar to that of\nreal data in various Machine Learning settings such as classification and\nprediction. We also give a privacy assessment and report on statistical\nanalysis regarding realistic characteristics of the synthetic data. The\nsoftware of this work is open-source and is available at:\nhttps://github.com/astorfi/cor-gan.\n

Ähnliche Arbeiten