A1387
Title: A multivariate harmonization framework for multi-site neuroimaging studies
Authors: Jingsong Zhou - University of Maryland (United States)
Shuo Chen - University of Maryland (United States)
Qiong Wu - University of Pittsburgh - Pittsburgh, PA (United States) [presenting]
Abstract: Multi-site large consortium studies involving neuroimaging and omics data have become common and essential for advancing biomedical research. Statistical analyses of such datasets rely on harmonization methods to remove technical variance and batch effects while preserving biological signals. However, most existing harmonization methods align the means and variances of neuroimaging features separately, without accounting for dependencies among imaging variables. This can result in incoherent covariance structures across study sites and suboptimal performance in downstream cross-site analyses. To address this gap, graph-guided ComBat (GG-ComBat) is proposed, a new harmonization framework that corrects batch effects in multivariate neuroimaging variables while allowing them to borrow strength from each other based on their dependencies. Specifically, GG-ComBat learns a graph structure from shared covariance patterns across study sites and leverages this structure to model parameters within the ComBat empirical-Bayes framework. Computationally efficient algorithms for empirical-Bayes estimation that explicitly account for cross-feature covariance are developed. Simulation studies and application to the adolescent brain cognitive development (ABCD) study demonstrate that GG-ComBat outperforms existing harmonization methods, including ComBat and CovBat, in recovering site-specific effects, aligning covariance across sites, and improving downstream prediction performance.
