Efficient Bayesian multivariate fMRI analysis using a sparsifying spatio-temporal prior.

Area of Science:

• Neuroimaging analysis
• Statistical modeling
• Machine learning

Background:

• Neuroimaging data analysis often requires sophisticated statistical models to handle high dimensionality and complex spatial-temporal dependencies.
• Existing methods may struggle with interpretability and computational efficiency for large datasets.

Purpose of the Study:

• To introduce a novel multivariate Bayesian logistic regression approach for neuroimaging data analysis.
• To incorporate spatio-temporal constraints for improved interpretability of results.
• To demonstrate the method's performance and feasibility on functional Magnetic Resonance Imaging (fMRI) data.

Main Methods:

• Development of Bayesian logistic regression utilizing a multivariate Laplace prior.
• Rewriting the multivariate Laplace distribution as a scale mixture to enable spatio-temporal constraints.
• Computation of the posterior distribution using expectation propagation and efficient inversion of a sparse precision matrix.
• Application to a functional Magnetic Resonance Imaging (fMRI) dataset involving visual stimuli (handwritten digits).

Main Results:

• The proposed method yields smooth and interpretable importance maps, facilitating the understanding of neuroimaging data.
• The models demonstrate competitive predictive performance compared to existing approaches.
• The computational approach is feasible even for very large models with thousands of variables.

Conclusions:

• The Bayesian logistic regression with a multivariate Laplace prior offers a powerful and interpretable tool for neuroimaging data analysis.
• Incorporating spatio-temporal constraints enhances the clinical and research utility of the derived importance maps.
• The method is computationally efficient and scalable to large-scale neuroimaging datasets.