npj Digital Medicine (Nov 2020)

Generating high-fidelity synthetic patient data for assessing machine learning healthcare software

  • Allan Tucker,
  • Zhenchen Wang,
  • Ylenia Rotalinti,
  • Puja Myles

DOI
https://doi.org/10.1038/s41746-020-00353-9
Journal volume & issue
Vol. 3, no. 1
pp. 1 – 13

Abstract

Read online

Abstract There is a growing demand for the uptake of modern artificial intelligence technologies within healthcare systems. Many of these technologies exploit historical patient health data to build powerful predictive models that can be used to improve diagnosis and understanding of disease. However, there are many issues concerning patient privacy that need to be accounted for in order to enable this data to be better harnessed by all sectors. One approach that could offer a method of circumventing privacy issues is the creation of realistic synthetic data sets that capture as many of the complexities of the original data set (distributions, non-linear relationships, and noise) but that does not actually include any real patient data. While previous research has explored models for generating synthetic data sets, here we explore the integration of resampling, probabilistic graphical modelling, latent variable identification, and outlier analysis for producing realistic synthetic data based on UK primary care patient data. In particular, we focus on handling missingness, complex interactions between variables, and the resulting sensitivity analysis statistics from machine learning classifiers, while quantifying the risks of patient re-identification from synthetic datapoints. We show that, through our approach of integrating outlier analysis with graphical modelling and resampling, we can achieve synthetic data sets that are not significantly different from original ground truth data in terms of feature distributions, feature dependencies, and sensitivity analysis statistics when inferring machine learning classifiers. What is more, the risk of generating synthetic data that is identical or very similar to real patients is shown to be low.