Benchmarking functional connectome-based predictive models for resting-state fMRI.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomarker Discovery

Background:

• Functional connectomes derived from resting-state fMRI (rest-fMRI) hold potential as biomarkers for psychological and clinical traits.
• Significant variability exists in analytical pipelines used to generate these connectomes, hindering reproducibility.

Purpose of the Study:

• To identify optimal modeling choices for predictive analyses of functional connectome edge weights.
• To benchmark different pipeline components across diverse neuroimaging cohorts.

Main Methods:

• Systematic benchmarking of over 240 pipelines across 6 cohorts (2000+ individuals) covering neurodegenerative, neuropsychiatric, drug use, and psychological traits.
• Evaluation of region definition strategies (pre-defined vs. data-driven), functional connectome representation methods (e.g., covariance, correlation), and supervised learning models (e.g., logistic regression).

Main Results:

• Data-driven region definitions consistently outperformed pre-defined regions.
• Tangent-based parametrization of covariances proved superior for capturing between-region interactions compared to simple correlations or partial correlations.
• Simple linear models, particularly logistic regression, demonstrated the best and most consistent prediction performances.

Conclusions:

• The study outlines robust modeling choices for reproducible functional connectome biomarker development.
• Findings pave the way for more reliable neuroimaging-based biomarkers in clinical settings.
• Optimized pipelines enhance the discovery of biomarkers for neurological and psychiatric conditions.