When Cleaning fMRI Data is a Nuisance

By Neuroskeptic | June 12, 2013 3:50 pm

A flaw in data processing could be leading to biases in fMRI brain functional connectivity patterns, according to a new report: The Nuisance of Nuisance Regression.

Ironically, two high-profile recent papers about bias are amongst the victims.

The new paper, from Pittsburgh’s Michael Hallquist and colleagues, is essentially about a case where 2 + 2 = 3. Functional connectivity fMRI (fcMRI) is a powerful technique which can reveal the pattern of interactions between brain areas.

However, raw fcMRI data is full of various kinds of artefacts caused by things such as breathing and head movement. To deal with these, two clean-up techniques are commonly used: bandpass filtering, and nuisance regression. Both of these are designed to get rid of unwanted components.

Filtering the data removes signals that change either so quickly, or so slowly, that they’re unlikely to be real neural activity, and can be assumed to be noise. No more noise.

Nuisance regression is more targeted; it takes some known measures of the ‘nuisance’ variables, for instance, changes in head position, and then it removes any rsfMRI signals that are correlated with them. No more nuisances.

These are both good tools. The problem, Hallquist et al say, is that they can interfere with each other, so the benefits actually cancel out. 2 + 2 = 3. But not always. It’s all about the order.

If you bandpass first, and then do nuisance regression (“BpReg”), the second step can actually end up putting noise back into the data. The trouble is that if you filter the fcMRI data but not the nuisance variables (which is what many people have been doing), it’s invalid to then regress them together.

The opposite order, doing nuisance regression first then bandpass filtering (“RegBp”), is not subject to the same problem. Best of all is to combine both approaches in a single simultaneous step. But the main lesson is to avoid BgReg.

Most interestingly, Hallquist et al point out that a number of recent papers that warned of the dangers of head movement in rsfMRI, used BgReg. These include a Neuroskeptic favorite that has led to concern in autism research circles.

These papers warned that standard data cleaning doesn’t get rid of motion artefacts. Hallquist et al say that this is true, but only for BpReg.

Many rsfMRI researchers do use BpReg, but not all; I suspect it’s just luck which has determined who’s used which order. The lesson of this paper is clear: don’t use BpReg, and if you don’t, you can breathe a small sigh of relief.

Here at Neuroskeptic we’ve twice seen how misapplied data ‘filtering’ can actually introduce biases.

ResearchBlogging.orgHallquist, M., Hwang, K., & Luna, B. (2013). The Nuisance of Nuisance Regression: Spectral Misspecification in a Common Approach to Resting-State fMRI Preprocessing Reintroduces Noise and Obscures Functional Connectivity NeuroImage DOI: 10.1016/j.neuroimage.2013.05.116

  • DS

    I don’t think the point of the paper is that BP filtering and regression interact. I think the point is that when BpReg is used then there is a consistency issue that could cause problems.

    Also I don’t see enforcing consistency, by instead using RegBP, as a full “fix” for the general problem of regression of nuisance variables (in particular motion effects). This is because motion can happen on a time scale less than the sampling frequency (the TR) and therefore the high frequency components of motion in the fMRI data will be aliased into frequencies less than or equal to the sampling frequency. BP filtering removes aliased signal outside the pass-band but leaves the aliased signal that is in the pass-band as well. When one then regresses using BP filtered nuisance variables (motion parameter time series data) consistency is enforced but one is now removing signal within the assumed frequency band of the BOLD signal. I can see that causing problems.

    Am I wrong? I hope the authors weigh in here.

  • http://petrossa.me/ petrossa

    I guess now that just about everything that can go wrong with fMRI just about does go wrong it’s about time to retract all fMRI based papers, especially those with the more absurd conclusions (those relating to higher order functions for example)

    The technique stinks, it’s a very complex tool, with very a low resolution, in the hands of people often not exactly statistical analyses experts to say the least.

    Mostly it has become a tool to ‘prove’ whatever the researcher hopes to find.

    • Wouter

      Yes, let’s throw out the baby with the bath water.

      • http://petrossa.me/ petrossa

        if its stillborn, yes why not

        • http://blogs.discovermagazine.com/neuroskeptic/ Neuroskeptic

          A baby, maybe. Stillborn, no; fMRI is a young technique, it’s still yet to mature. Perhaps it has yet to be born at all, and these methodological controversies are the kickings of the developing fetus. A sign of health but also a reminder of immaturity.

          • http://petrossa.me/ petrossa

            The basic concept is wrong. You can’t improve something when the basic premise is faulty. This is a waste of money, time and effort. Find another method that doesn’t depend on the fallacy that oxygenation of parts of the brain are 1 on 1 positively related to the task at hand. They might just be busy inhibiting another part of the brain.



No brain. No gain.

About Neuroskeptic

Neuroskeptic is a British neuroscientist who takes a skeptical look at his own field, and beyond. His blog offers a look at the latest developments in neuroscience, psychiatry and psychology through a critical lens.


See More

@Neuro_Skeptic on Twitter


Discover's Newsletter

Sign up to get the latest science news delivered weekly right to your inbox!

Collapse bottom bar