A0782
Title: Federated factor analysis for collaborative learning of multi-site imaging data
Authors: Yezhi Pan - University of Maryland (United States)
Shuo Chen - University of Maryland (United States)
Qiong Wu - University of Pittsburgh (United States) [presenting]
Abstract: Multi-site collaboration is increasingly common in biomedical science, providing valuable opportunities to analyze complex datasets, such as neuroimaging data, that require extensive time and resources for collection. Factor analysis is a widely used statistical technique to uncover multivariate relationships by identifying latent factors. However, re-identification risks and privacy policies restrict the sharing of sensitive individual-level data across sites, challenging multi-site factor analysis. Traditional per-site factor analysis can yield inconsistent definitions of latent factors, complicating interpretability when aggregating results across sites. Another primary challenge is the inherent data heterogeneity arising from variations in patient populations and technical factors across sites. Hence, a structure-guided confirmatory factor analysis (SCFA) is proposed to identify a unified set of latent factors across sites while allowing site-specific heterogeneity in factor covariances. A federated learning algorithm, Fed-Factor, is introduced to solve the SCFA model securely with only a single round of summary statistic communication across sites, producing results identical to those obtained from pooling individual-level data. Its effectiveness and utility are demonstrated through simulations and imaging data from Adolescent Brain Cognitive Development (ABCD) study.
