Identifying the Relationship Structure Among Multiple Datasets Using Independent Vector Analysis: Application to Multi-Task fMRI Data

Abstract

Identifying relationships among multiple datasets is an effective way to summarize information and has been growing in importance. In this paper, we propose a robust 3-step method for identifying the relationship structure among multiple datasets based on Independent Vector Analysis (IVA) and bootstrap-based hypothesis testing. Unlike previous approaches, our theory-backed method eliminates the need for user-defined thresholds and can effectively handle non-Gaussian data. It achieves this by incorporating higher-order statistics through IVA and employing an eigenvalue decomposition-based feature extraction approach without distributional assumptions. This way, our method estimates more interpretable components and effectively identifies the relationship structure using hierarchical clustering. Simulation results demonstrate the effectiveness of our method, as it achieves perfect Adjusted Mutual Information (AMI) for different values of the correlation between the components. When applied to multi-task fMRI data from patients with schizophrenia and healthy controls, our method successfully reveals activated brain regions associated with the disorder, and identifies the relationship structure of task datasets that matches our prior knowledge of the experiment. Moreover, our proposed method extends beyond task datasets, offering broad applicability in subgroup identification in neuroimaging and other domains.

Keywords

• Blind source separation
• bootstrap
• data-driven
• fMRI
• independent vector analysis
• relationship structure