A Scalable Approach to Independent Vector Analysis by Shared Subspace Separation for Multi-Subject fMRI Analysis.

Area of Science:

• Neuroimaging
• Data Analysis
• Machine Learning

Background:

• Joint blind source separation (JBSS) is crucial for analyzing related datasets but faces computational challenges with high-dimensional data.
• Existing JBSS methods can be inefficient or inaccurate if the data's latent dimensionality is poorly modeled, leading to overparameterization.
• Scalability is limited by the number of datasets that can be practically included in analysis.

Purpose of the Study:

• To develop a computationally scalable JBSS method for high-dimensional and multi-dataset analyses.
• To address limitations in existing JBSS approaches concerning dimensionality and performance.
• To enhance the modeling of latent structures across multiple related datasets.

Main Methods:

• Proposes a scalable JBSS method by separating a 'shared' subspace, defined by low-rank structures across datasets.
• Utilizes efficient initialization of independent vector analysis (IVA) with a Gaussian source prior (IVA-G) for estimating shared sources.
• Applies JBSS separately to shared and non-shared sources after evaluation, reducing problem dimensionality.

Main Results:

• The proposed method demonstrates excellent estimation performance on resting-state fMRI data.
• Achieves significantly reduced computational costs compared to traditional JBSS methods.
• Effectively handles analyses involving a larger number of datasets by reducing problem dimensionality.

Conclusions:

• The novel scalable JBSS method offers improved efficiency and accuracy for multi-dataset analysis.
• This approach effectively models shared latent structures, particularly beneficial for high-dimensional neuroimaging data.
• The method enhances the tractability and performance of JBSS, enabling broader applications.