On the replicability of diffusion weighted MRI-based brain-behavior models.

Area of Science:

• Neuroimaging
• Connectomics
• Biostatistics

Background:

• Replicability of brain-wide association studies (BWAS) using MRI is a growing concern.
• Previous studies focused on functional and anatomical MRI, leaving structural connectomes under-evaluated.

Purpose of the Study:

• To comprehensively evaluate the replicability of BWAS using various diffusion MRI-derived structural connectome metrics.
• To investigate factors influencing BWAS replicability, including phenotype type and effect size.

Main Methods:

• Utilized diffusion MRI data from Human Connectome Project (HCP) and Autism Brain Imaging Data Exchange (ABIDE) datasets.
• Assessed replicability of brain-phenotype associations using streamline counts, fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD).
• Categorized phenotypes as trait-like or state-like and analyzed the relationship between replicability, effect size, and sample size.

Main Results:

• 36% of brain-phenotype associations were replicable across datasets using at least one diffusion MRI metric with discovery sample sizes up to 425.
• Trait-like phenotypes demonstrated higher replicability (50%) compared to state-like phenotypes (19%).
• Streamline-based connectomes showed the highest replicability (29-42%), and replicability directly correlated with effect size, with larger effects (>5% variance) requiring smaller discovery sample sizes (<300).

Conclusions:

• Trait-like phenotypes exhibit good replicability with moderate sample sizes in BWAS using structural connectomes.
• BWAS models requiring very large sample sizes (>425) likely have limited practical relevance due to small effect sizes.
• Large sample sizes remain essential for robust explainability, fairness assessment, and generalizability of neuroimaging findings.