Decentralized Spatially Constrained Source-Based Morphometry

Abstract

There is growing interest in extracting multivariate patterns (covarying networks) from structural magnetic resonance imaging (sMRI) data to analyze brain morphometry. Constrained source-based morphometry (constrained SBM) is a hybrid approach which provides a fully automated strategy for extracting subject-specific parameters characterizing gray matter networks. In constrained SBM, constrained independent component analysis (ICA) is used to compute maximally independent sources and statistical analysis is used to identify sources significantly associated with variables of interest. However, constrained SBM is built on the assumption that the data are locally accessible. As such, it cannot take advantage of decentralized (i.e., federated) data. While open data repositories have grown in recent years, there are various reasons (e.g., privacy concerns for rare disease data, institutional or IRB policies, etc.) that restrict a large amount of existing data to local access only. To overcome this limitation, we introduce a novel approach: decentralized constrained source-based morphometry (dcSBM). In our approach, data samples are located at different sites and each site operates the constrained ICA in a distributed manner. Finally, a master node simply aggregates result estimates from each local site and runs the statistical analysis centrally. We apply our method to UK Biobank sMRI data and validate our results by comparing to centralized constrained SBM results.