Bayesian longitudinal tensor response regression for modeling neuroplasticity

Area of Science:

• Neuroimaging
• Neuroscience
• Biostatistics

Background:

• Longitudinal neuroimaging studies aim to detect voxel-level neuroplasticity across visits.
• Traditional voxel-wise methods have limitations compromising accuracy in detecting subtle brain changes.
• Accurate assessment of neuroplasticity is crucial for understanding treatment effects and disease progression.

Purpose of the Study:

• To propose a novel Bayesian tensor response regression for analyzing longitudinal neuroimaging data.
• To develop a method that improves the accuracy and power of detecting neuroplasticity compared to traditional voxel-wise approaches.
• To enable both group-level and individual-level inference of neuroplasticity for personalized treatment assessment.

Main Methods:

• Bayesian tensor response regression implemented using Markov chain Monte Carlo (MCMC) sampling.
• Utilizes low-rank decomposition for dimensionality reduction and preservation of spatial voxel configurations.
• Employs joint credible regions for feature selection, respecting posterior distribution shapes for accurate inference.

Main Results:

• The proposed method demonstrated superior prediction and feature selection capabilities over voxel-wise regression in simulations.
• Analysis of a longitudinal Aphasia fMRI dataset revealed distinct neuroplasticity patterns for control and intention treatments.
• Control therapy showed long-term brain activity increases, while intention treatment yielded short-term changes, localized to specific regions.

Conclusions:

• The novel Bayesian tensor regression method effectively detects neuroplasticity in longitudinal neuroimaging data.
• The approach offers advantages in accuracy, feature selection, and individual-level inference over conventional voxel-wise methods.
• Findings highlight differential long-term and short-term effects of interventions on brain activity in Aphasia patients.