Fully synthetic neuroimaging data for replication and exploration

Scientific transparency, data exploration, and education are advanced through data sharing. However, risk for disclosure of personal information and institutional data sharing regulations can impede human subject/patient data sharing and thus limit open science initiatives. Sharing fully synthetic data is an alternative when it is not possible to share real or observed data. Here we describe a data sharing approach that borrows principles and methods from multiple imputation to replace observed values with synthetic values, thereby creating a fully synthetic neuroimaging dataset that accurately represents the covariance structure of the observed dataset. Predictor tables composed of demographic, site, behavioral and total intracranial volume (ICV) variables from 264 pediatric cases were used to create synthetic predictor tables, which were then used to synthesize gray matter images derived from T1-weighted data. The synthetic predictor tables demonstrated pooled variance and statistical estimates that closely approximated the observed data, as reflected in measures of efficiency and statistical bias. Similarly, the synthetic gray matter data accurately represented the variance and voxel-level associations with predictor variables (age, sex, verbal IQ, and ICV). The magnitude and spatial distribution of gray matter effects in the observed imaging data were replicated in the pooled results from the synthetic datasets. This approach for generating fully synthetic neuroimaging data has widespread potential for data sharing, including replication, new discovery, and education. Fully synthetic neuroimaging datasets can enable data-sharing because it accurately represents patterns of variance in the original data, while diminishing the risk of privacy disclosures that can accompany neuroimaging data sharing.