A tensor based varying-coefficient model for multi-modal neuroimaging data analysis.

Area of Science:

• Neuroscience
• Statistics
• Data Science

Background:

• Neuroimaging research increasingly combines multiple modalities to overcome individual limitations.
• Integrating diverse data types like behavioral, genetic, and neuroimaging data is a growing trend.
• Many complex datasets, including neuroimaging, can be represented as time-varying tensors.

Purpose of the Study:

• To propose a novel time-varying tensor regression model for analyzing neural correlates.
• To handle tensor-valued brain images and tensor-valued covariates collected over time.
• To extend existing regression models for complex, large-scale structural data while preserving inherent data structures.

Main Methods:

• Developed a time-varying tensor regression model with structural composition for responses and covariates.
• Utilized B-spline techniques to express regression coefficients.
• Employed CP-decomposition to estimate basis function coefficients by minimizing a penalized loss function.
• Created a varying-coefficient model accommodating both tensor-valued covariates and responses.

Main Results:

• The proposed model effectively analyzes complex, multidimensional neuroimaging data.
• Demonstrated the method's efficacy through simulated data analysis.
• Validated the approach using real-world data, including functional magnetic resonance imaging (fMRI) and eye-tracking data.

Conclusions:

• The developed tensor regression model offers a powerful approach for integrating multimodal neuroimaging and non-imaging data.
• This method preserves the inherent structure of complex data, advancing the study of neural correlates.
• The approach provides a significant extension for analyzing large-scale, time-varying, tensor-valued datasets in neuroscience.