Creating, evaluating, and sharing synthetic data for multinational HIV cohorts

Vanderbilt University Medical Center, Nashville TN

Investigators

Abstract

Project Summary/Abstract Principles of open science often clash with requirements of data privacy laws and organizational policies, especially with respect to sensitive health data. This conflict has become particularly evident in international HIV research, which often relies on observational cohort data. Currently, researchers cannot publicly share HIV cohort data (e.g., post on a publicly available website) and therefore many of the benefits of open science cannot be realized, such as enabling the reproducibility of published findings and creating opportunities for a larger number of people to study the data and contribute scientific discoveries. Data privacy methodologies developed for data sharing often lower the fidelity of the data to reduce the risk of re-identification, but they still preserve the linkage between records and real people. Synthetic data generated by computer simulation, by contrast, offer a more desirable solution as they can resemble the characteristics of the original data while severing the linkage between records and real people. Though synthetic data can never fully replace the original data, they can benefit science. However, based on preliminary work, existing methods for generating synthetic data are insufficient to generate synthetic HIV cohort data that closely resemble the original data. In this project, we will adapt and apply state-of-the-art artificial intelligence methods for generating synthetic data that are tailored to longitudinal observational HIV cohorts. We will apply these methods to create synthetic data intended to mimic data from two multinational HIV cohorts: the Caribbean, Central and South American network for HIV epidemiology (CCASAnet) and the East African region of the International epidemiology Databases to Evaluate AIDS (IeDEA-EA). These cohorts are composed of hundreds of thousands of people living with HIV. We will evaluate the quality of our synthetic data using both intrinsic (e.g., proportions, distributions, correlations between variables) and extrinsic comparisons to the original data. Extrinsic comparisons will be performed by a research advisory board of HIV investigators who will select comparative HIV studies independent of the data generation process. We hypothesize that the synthetic data will be useful for hypothesis generation, and these extrinsic comparisons will provide evidence supporting or refuting this hypothesis. We will also study potential ethical, legal, and social (e.g., perception, discomfort) barriers to sharing synthetic HIV cohort data through qualitative studies among people involved in HIV cohort studies (e.g., investigators, administrators, and patients). Finally, we will develop software toolkits so that others can generate synthetic HIV cohort data.