Benchmarking resting state fMRI connectivity pipelines for classification: robust accuracy despite processing
Tatiana Medvedeva1, Irina Knyazeva2, Ruslan Masharipov2
1N. P. Bechtereva Institute of the Human Brain, Russian Academy of Sciences, St. Petersburg, Russia. tanai2010@mail.ru.
Brain Informatics
|May 5, 2026
Summary
Machine learning models reliably distinguish brain states using functional connectivity (FC) from fMRI data, even across different labs. This demonstrates the robustness of FC for biomarker discovery in neuroscience.
Area of Science:
- Neuroscience
- Machine Learning
- Neuroimaging
Background:
- Machine learning (ML) shows promise in neuroscience, but reproducibility in neuroimaging is a challenge.
- Inconsistent preprocessing and functional connectivity (FC) calculations limit ML generalizability in brain activity analysis.
- Variability in brain states and data acquisition further complicates ML applications in neuroimaging.
Purpose of the Study:
- To systematically assess the impact of preprocessing and FC pipeline variations on ML classification of fMRI data.
- To benchmark 256 distinct FC analysis pipelines for classifying between eyes-open and eyes-closed brain states.
- To evaluate model generalizability across different laboratories and datasets using cross-site validation and few-shot domain adaptation.
Main Methods:
- Compared 256 functional connectivity (FC) analysis pipelines for fMRI data classification.
- Utilized two independent datasets from healthy participants collected in different laboratories.
- Employed direct cross-site validation and few-shot domain adaptation for model testing.
- Assessed classification accuracy and reproducibility across diverse preprocessing, parcellation, and connectivity metrics.
Main Results:
- Consistently high classification accuracy (around 90%) was achieved in discriminating between eyes-open and eyes-closed states.
- FC-based models demonstrated robust performance across different acquisition sites, despite methodological variations.
- Optimal accuracy and stability were achieved using Pearson correlation and tangent space parametrization for FC, Brainnetome atlas, and CompCor-based confound regression.
Conclusions:
- Resting-state fMRI (rs-fMRI) functional connectivity (FC) characteristics are resilient to methodological variations.
- FC-based models can robustly differentiate well-defined brain states across different sites.
- These findings support the utility of rs-fMRI FC for discovering reliable biomarkers, especially in stable brain states.
