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

Amirsina Torfi, Edward A. Fox

TL;DR

This paper introduces CorGAN, a novel Generative Adversarial Network (GAN) architecture that uses Convolutional Neural Networks (CNNs) and Convolutional Autoencoders (CAEs) to generate realistic synthetic electronic health records (EHRs). CorGAN effectively captures correlations between medical features, producing synthetic data that performs comparably to real data in downstream machine learning tasks while offering improved privacy.

What Problem Does This Paper Address?

The increasing adoption of Electronic Health Records (EHRs) has created a wealth of data for healthcare research. However, accessing and utilizing this sensitive data for research is severely hampered by privacy concerns. Traditional de-identification methods are increasingly questioned for their vulnerability to information leakage. This has led to a growing interest in Synthetic Data Generation (SDG) as a privacy-preserving alternative. While GANs have shown promise in generating synthetic data for various domains like image and video, applying them to complex, discrete, and often temporal EHR data presents unique challenges, particularly in capturing the intricate correlations between different medical features.

Methods

The core of CorGAN lies in its architecture, which combines Convolutional GANs (CGANs) with Convolutional Autoencoders (CAEs). The generator part of the GAN is designed to produce continuous synthetic data, which is then transformed into discrete values (mimicking EHR data like diagnosis codes) using the decoder of a pre-trained Convolutional Autoencoder. This indirect method of generating discrete data is chosen because GANs typically struggle with discrete outputs.

Crucially, CorGAN employs 1D Convolutional Neural Networks (CNNs) for both the generator and discriminator, rather than standard Multilayer Perceptrons (MLPs). This choice is motivated by the ability of CNNs to effectively capture local correlations between adjacent features, which is vital for understanding relationships within medical records. The paper also mentions using techniques like minibatch discrimination and batch normalization to improve training stability and prevent mode collapse, a common issue in GANs.

Datasets

The paper evaluates CorGAN on two publicly available datasets:

  • MIMIC-III: This dataset, containing medical records of nearly 46,000 patients, is used for experiments with binary discrete variables (specifically, ICD-9 codes). The representation used is a fixed-size vector of 1071 entries per patient record.
  • UCI Epileptic Seizure Recognition dataset: This dataset, characterized by brain activity (EEG recordings), is used for experiments with continuous variables. It has 179 features and 11,500 samples, with the core task being binary classification to detect seizure activity.

Key Insights and Ideas

  • CNNs for EHR Correlation: The paper strongly advocates for using CNNs over MLPs for capturing inter-feature correlations in EHR data. This is a key insight, suggesting that the spatial (or sequential in 1D) nature of medical features can be better leveraged by convolutional operations.
  • Autoencoder for Discrete Data Transformation: The use of a pre-trained convolutional autoencoder's decoder to transform continuous GAN outputs into discrete EHR-like data is an elegant solution to a common GAN limitation.
  • CorGAN's Superiority: CorGAN demonstrates a significant improvement in capturing correlations compared to other methods, leading to more realistic synthetic data.
  • Privacy-Utility Trade-off: The study provides empirical evidence that CorGAN offers an acceptable level of privacy, with the privacy risk not necessarily increasing with the amount of synthetic data generated, especially when a fixed amount of real data is known to an adversary.

Key Results

  • Dimension-wise Probability: CorGAN shows superior performance in matching the dimension-wise probability distributions of real data compared to baseline models (visualized in Figure 2).
  • Dimension-wise Prediction: CorGAN achieves the smallest F1-score variation (difference between real and synthetic data performance) in dimension-wise prediction tasks, indicating its ability to preserve inter-dimensional relationships. The F1-score difference for CorGAN is 0.021 ± 0.045, significantly lower than other models.
  • Binary Classification: For binary classification tasks, CorGAN achieves the highest AUROC (0.92 ± 0.012) and AUPRC (0.41 ± 0.015) among the generative models tested, outperforming DBM, VAE, and medGAN.
  • Privacy Assessment: The privacy assessment using membership inference attacks suggests that CorGAN provides a good level of privacy. Increasing the number of synthetic records, while keeping the number of known real records fixed, led to a dramatic decrease in precision for the attack, making it harder for an adversary to identify real records.

Limitations and Future Work

  • Temporal Data Ignored (MIMIC-III): For the MIMIC-III dataset experiments, the temporal information within the EHR records was ignored. The authors suggest that their method could be effectively extended to longitudinal records to capture temporal characteristics.
  • Privacy Assessment Scope: While the privacy assessment is promising, it's based on specific attack scenarios. Further research could explore a wider range of privacy attacks.
  • Generalizability: The paper focuses on discrete and continuous EHR data. Future work could explore generating more complex data types or combinations thereof.

Summary updated

Abstract

Deep learning models have demonstrated high-quality performance in areas such as image classification and speech processing. However, creating a deep learning model using electronic health record (EHR) data, requires addressing particular privacy challenges that are unique to researchers in this domain. This matter focuses attention on generating realistic synthetic data while ensuring privacy. In this paper, we propose a novel framework called correlation-capturing Generative Adversarial Network (CorGAN), to generate synthetic healthcare records. In CorGAN we utilize Convolutional Neural Networks to capture the correlations between adjacent medical features in the data representation space by combining Convolutional Generative Adversarial Networks and Convolutional Autoencoders. To demonstrate the model fidelity, we show that CorGAN generates synthetic data with performance similar to that of real data in various Machine Learning settings such as classification and prediction. We also give a privacy assessment and report on statistical analysis regarding realistic characteristics of the synthetic data. The software of this work is open-source and is available at: https://github.com/astorfi/cor-gan.
BibTeX
@inproceedings{torfi2020corgan,
  author = {Amirsina Torfi and Edward A. Fox},
  title = {{CorGAN: Correlation-Capturing Convolutional Generative Adversarial Networks for Generating Synthetic Healthcare Records}},
  booktitle = {FLAIRS},
  year = {2020},
  url = {https://openreview.net/forum?id=czulZemXl3},
}
Figure