Privacy-Preserving Generative Deep Neural Networks Support Clinical Data Sharing

Abstract

Background: Data sharing accelerates scientific progress but sharing individual-level data while preserving patient privacy presents a barrier.

Methods and results: Using pairs of deep neural networks, we generated simulated, synthetic participants that closely resemble participants of the SPRINT trial (Systolic Blood Pressure Trial). We showed that such paired networks can be trained with differential privacy, a formal privacy framework that limits the likelihood that queries of the synthetic participants' data could identify a real a participant in the trial. Machine learning predictors built on the synthetic population generalize to the original data set. This finding suggests that the synthetic data can be shared with others, enabling them to perform hypothesis-generating analyses as though they had the original trial data.

Conclusions: Deep neural networks that generate synthetic participants facilitate secondary analyses and reproducible investigation of clinical data sets by enhancing data sharing while preserving participant privacy.

Keywords: blood pressure; deep learning; machine learning; privacy; propensity score.

Figure 1.

Median systolic blood pressure trajectories from initial visit to 27 mo.

Figure 2.

Pairwise Pearson correlation between columns. A, Original and real data, (B) nonprivate and auxiliary classifier generative adversarial network (AC-GAN) simulated data, and (C) differentially private and AC-GAN simulated data (RZ, randomization visit; 1M, 1-mo visit; 2M, 2-mo visit; 3M, 3-mo visit; 6M, 6-mo visit; 9M, 9-mo visit; 12M, 12-mo visit; 15M, 15-mo visit; 18M, 18-mo visit; 21M, 21-mo visit; 24M, 24-mo visit; and 27M, 27-mo visit).

Figure 3.

Clinician evaluation of synthetic data. A, Synthetic participant scored a 2 by clinician expert. B, Synthetic participant scored a 4 by clinician expert. C, Synthetic participant scored a 6 by clinician expert. D, Synthetic participant scored an 8 by clinician expert. E, Comparison of scores between real and synthetic participant (dotted red lines indicate means). F, Distribution of scores between real (blue) and synthetic (green) patients. BP indicates blood pressure.

Figure 4.

Accuracy of models trained on synthetic participants vs real data. Line indicates performance on real data, which on average should provide the best possible performance; bar indicates performance of classifier trained on private synthetic participants; bottom of chart indicates random performance.

Figure 5.

The value of delta as a function of epoch for different ε values. An ε value of 3.5 allows for 1000 epochs of training and δ<10⁻⁵.

Figure 6.

Machine learning and statistical evaluation of synthetic data. A–D, Performance on transfer learning task by source of training data for each machine learning method. E, Pairwise Pearson correlation between columns for the original and real data. F, Pairwise Pearson correlation between columns for the private synthetic data. AUROC indicates area under the receiver operator characteristic; LR, logistic regression; RF, random forest.