Best Practices for Exclusion Criteria in fMRI Data with High Noise and Variance due to Noise (Electromagnetic?)

I have collected an fMRI dataset under a study where each subject completes four runs in a single session: one reference run with no external stimulus and three runs with distinct external stimuli. Unfortunately, due to suspected electromagnetic interference from the external stimulus device, the fMRI data for some subjects/runs are excessively noisy and exhibit high temporal variance.

I preprocessed the data using fMRIPrep, which has provided useful QC metrics, but the noise and variance issues persist in some cases.
Given this situation:

  1. What academically accepted exclusion criteria could I employ to identify and exclude problematic subjects or runs in a consistent and justifiable manner?
  2. Are there established metrics (e.g., temporal signal-to-noise ratio, high temporal variance thresholds, fMRIPrep QC outputs, etc.) or preprocessing-derived measures that are typically used to quantify and standardize exclusion decisions?
  3. How can I justify the exclusion of entire subjects (i.e., all their runs) versus individual runs in cases where some runs appear to meet quality standards while others do not?

I cannot rely on motion-based exclusion criteria as the primary factor, as motion artifacts are minimal in my dataset. Additionally, because of privacy constraints, I cannot share specific images or examples of noisy data.

I would greatly appreciate references to relevant literature, guidelines, or community-accepted practices to improve pre-processing of my data and ensure the criteria are defensible in a publication.

Thank you for your insights!

Hi-

Quality control (QC) is an important topic. You might find useful information in this recent community wide project on describing QC procedures for FMRI data:
FMRI Open QC Project: Demonstrating Quality Control (QC) Procedures in fMRI
The emphasis on the project was to help build up QC as not just a way to filter “good” and “bad” data, but as a way to better understand data properties and have more confidence in final results.

There are 10 contributions from various teams around the field, using a wide range of software, discussion QC considerations and issues (and one Editorial summarizing findings from around the whole project). Each team’s contribution provides many examples. Provins et al. (2023) there specifically use fMRIPrep.

–pt