S8: Confounds & Physio Regressors

Step Code: S8_confounds_and_physio_regressors Depends on: S3 (frame metrics), S4 (motion), S6 (warps), S7 (PAM50 levels) Required by: S10 (QC Aggregation & Release)

Purpose

S8 builds the nuisance regressors — the time-courses an analyst later removes from the BOLD signal so that physiological pulsation, motion and scanner drift do not masquerade as neural activity. It does not regress anything itself and does not resample the BOLD: it computes the regressors in native functional space and writes them to a table for the downstream general linear model (GLM).

The output is one BIDS-Derivatives *_desc-confounds_timeseries.tsv per run, with a JSON sidecar describing each column.

No slice-timing correction. The pipeline never temporally resamples the BOLD. The slice-timing metadata in the BIDS sidecar is used here only to phase the RETROICOR cardiac/respiratory regressors — following the cord-fMRI field standard (Eippert 2017; Kaptan 2023).

Regressor families

S8 emits five families, a mix of single columns and per-slice columns. CSF and RETROICOR are computed per slice because cord GLMs are often fit slice by slice.

1. Motion and outlier flags (always present)

• Translations trans_x, trans_y: slice-mean translation per volume from S4's motion parameters, plus first derivatives.
• Framewise displacement (FD): |Δtrans_x| + |Δtrans_y| — the 2D cord adaptation of Power's FD.
• DVARS / RefRMS: already computed at S3; repacked here with outlier flags.
• Outlier flags: a volume is flagged if FD > 0.5 mm (Power 2014 lenient scrubbing cutoff for the cord) or if DVARS/RefRMS exceeds a Tukey Q3 + 1.5·IQR fence (the fMRIPrep convention; robust to the cord's heavy-tailed DVARS).

2. CSF aCompCor, per slice (when a CSF mask exists)

The top 5 principal components of the cerebrospinal-fluid (CSF) signal in each slice (csf_slice{z}_pc{k}). aCompCor (Behzadi 2007) captures shared physiological fluctuation. Computed with FSL fslmeants --eig after per-voxel constant+linear detrending, on a subject-specific CSF mask (S2 canal minus cord, warped to native func via S6; falls back to the PAM50 CSF mask). No erosion (the cord-CSF rim is too thin); slices with fewer than 5 voxels are skipped.

3. RETROICOR, per slice (when readable physiology is present)

Cardiac and respiratory phase regressors via FSL PNM, auto-disabled when physio recordings are absent. With cardiac order 4, respiratory order 4 and interaction order 2, that is 8 + 8 + 16 = 32 regressors per slice (Kaptan 2023 / Dabbagh 2024 cord recipe), aggregated to one column per regressor by averaging across cord-bearing slices. Optional heart-rate and respiration-volume-per-time (RVT) slow regressors are added.

4. Cosine drift basis (default ON)

A discrete-cosine high-pass basis — the regressor-based equivalent of high-pass filtering — with a 0.01 Hz (1/100 s) cutoff (Kaptan 2023 / Dabbagh 2024 cord standard). Disable when the analyst's GLM owns high-pass filtering.

5. SpinalCompCor (default ON)

A cord-specific anatomical-noise component method (Hemmerling 2025). It dilates the cord-plus-CSF mask by 18 mm, subtracts the original to define a surrounding "noise" region, and extracts the top 5 principal components of that region's signal. By default these are taken globally with a fixed component count; an optional, slower per-slice parallel-analysis selection (IAAFT surrogates) is available for research use.

Step Metric and QC

The step-local check is the design-matrix condition number — a number that flags when regressors are so correlated that the GLM becomes numerically unstable. Because the per-slice CSF regressors are applied slice-locally downstream, S8 scores each slice's real design (global regressors plus that slice's own columns) and gates on the worst slice.

High motion only soft-warns and never fails: whether to drop a high-motion run is the analyst's call at GLM time.

Outputs

derivatives/spineprep/{dataset}/sub-{id}/func/
├── sub-{id}_task-{task}_desc-confounds_timeseries.tsv   # all regressor columns
└── sub-{id}_task-{task}_desc-confounds_timeseries.json  # column descriptions

derivatives/spineprep/{dataset}/sub-{id}/figures/
└── sub-{id}_..._desc-S8_confound_columns.png            # regressor overview

References

1. aCompCor: Behzadi et al. NeuroImage 37(1):90–101 (2007). DOI
2. FD / scrubbing: Power et al. NeuroImage 84:320–341 (2014). DOI
3. Cord RETROICOR recipe: Kaptan et al. NeuroImage (2023); Dabbagh et al. (2024).
4. SpinalCompCor: Hemmerling et al. (2025).

Last updated: June 2026