The DC-ren partner Mosaiques diagnostics, Hannover, Germany, headed by Harald Mischak, has just published a manuscript on the creation of synthetic patients on the medrxiv preprint server.
Abstract
Generating high quality, real-world clinical and molecular datasets is challenging, costly and time intensive. Consequently, such data should be shared with the scientific community, which however carries the risk of privacy breaches. The latter limitation hinders the scientific community’s ability to freely share and access high resolution and high quality data, which are essential especially in the context of personalised medicine. In this study, we present an algorithm based on Gaussian copulas to generate synthetic data that retain associations within high dimensional (peptidomics) datasets. For this purpose, 3,881 datasets from 10 cohorts were employed, containing clinical, demographic, molecular (> 21,500 peptide) variables, and outcome data for individuals with a kidney or a heart failure event. High dimensional copulas were developed to portray the distribution matrix between the clinical and peptidomics data in the dataset, and based on these distributions, a data matrix of 2,000 synthetic patients was developed. Synthetic data maintained the capacity to reproducibly correlate the peptidomics data with the clinical variables. Consequently, correlation of the rho-values of individual peptides with eGFR between the synthetic and the real-patient datasets was highly similar, both at the single peptide level (rho = 0.885, p < 2.2e-308) and after classification with machine learning models (rho synthetic = -0.394, p = 5.21e-127; rho real = -0.396, p = 4.64e-67). External validation was performed, using independent multi-centric datasets (n = 2,964) of individuals with chronic kidney disease (CKD defined as eGFR < 60 mL/min/1.73m2) or those with normal kidney function (eGFR > 90 mL/min/1.73m2). Similarly, the association of the rho-values of single peptides with eGFR between the synthetic and the external validation datasets was significantly reproduced (rho = 0.569, p = 1.8e-218). Subsequent development of classifiers by using the synthetic data matrices, resulted in highly predictive values in external real-patient datasets (AUC values of 0.803 and 0.867 for HF and CKD, respectively), demonstrating robustness of the developed method in the generation of synthetic patient data. The proposed pipeline represents a solution for high-dimensional sharing while maintaining patient confidentiality.
