Polarity of source current density estimates

Hi all ~ The recommendation in Brainstorm is to take the absolute values of current strength estimates. This is justified by stating that the negatives are difficult to interpret.

I was hoping you could comment a little bit on this topic. What I find in the source estimate solutions, is that the source waveforms roughly follow the ERP waveforms. That is, they will cross zero with a similar profile.

Consequently - the situation can arise where condition A is >0, condition B = 0 and condition C < 0. Psychologically we interpret this as a modulation of the response across the three conditions. However if I were to take the absolute values - then cond A and C would > than cond B. This is not consistent with the experimental manipulation and interpretation of ERPs.

Can you comment a little on how to interpret the negatives in the MNEs?
Many thanks.
Agatha

This is a great question, Agatha and you are pointing at some of the ambiguity in interpreting the sign of source time courses.

The sign of source currents depends on the physiology of neural activity (inward vs outward current flows through cortical layers) but is also ambiguous because of the limited spatial resolution of MEG and EEG source imaging, although in a somewhat counterintuitive way. Indeed, imagine the activity is strong and flowing outward (i.e., positive sign) on one wall of a sulcus and pretty silent elsewhere. Source imaging will tend to exaggerate the spread of this activity, reaching to the opposite side of the same sulcus. However, the cortical source currents are all oriented by default pointing outward of the cortical surface. Hence the logic of limited spatial resolution imposes that the spurious currents contribute similarly to the original sensor data as the actual neural activity. Hence spurious currents will flow inward (i.e., with negative sign) on the sulcus wall opposite to where the actual activity is located. Of course, you can interchange positive for negative and negative for positive in a symmetric, equivalent situation.
This appears very clearly when you uncheck the Absolute Value option when viewing your 3D source models. The blue/red border follows really well the anatomical folds.

Hence generally, it is not possible to tell whether the true currents are negative or positive (also because true brain activity is a mix of negative and positive current flows). Unless you contrast two conditions in the same individual: with the same folding pattern, the bias of spatial resolution is the same in the 2 conditions and therefore computing a difference will preserve consistency of the signs. When you need to go for an analysis of group maps across individuals, the projection onto another folding pattern (e.g., a common brain template) imposes that current flows are projected on another anatomy, hence their signs should not be preserved.

You will find a lot of the source imaging literature taking absolute values of current for this reason. Hence brain activity is reduced to activation, which is for measuring the magnitude of activity only. Of course, there is a lot of interest in preserving signs also for preserving the frequency and the phase of neural oscillations. If your interest is in time-frequency decompositions, then of course you don’t want to take the absolute values of currents because rectifying the time courses would essentially and artificially double the actual frequencies of neural oscillations. Time-frequency maps take care of the sign ambiguity by measuring the power of oscillations, which is a metric insensitive to sign.

To summarize, if you are interested in looking for effects where sign is expected to change across conditions, I recommend you compute contrast maps at the individual subject level and then perform statistical inference on these difference maps at the group level. It is also crucial to work with relative values when computing time-frequency decompositions or identifying oscillatory components: rectifying currents by taking their absolute values would spoil the spectral decompositions. Also, when averaging within a region, the statistics considered is usually power - hence with no effect of sign - rather than phase. For all other situations, when just the level of activation is of interest, you may rectify the signals first, then standardize them across the brain using a z-score, and run individual or group statistics.

I hope it clarifies a bit. Let me know if you still have questions: I know it’s a tricky thing to comprehend.

Cheers,

1 Like

Hi Sylvain ~ Thank you for the thorough answer. This does clarify the recommendation for absolute values. Indeed I do see the specific artefact that you mention in my data. In my case the comparisons are within subject, across conditions – so it seems that I am protected by having a within-subject design. I can work with condition differences and don’t need to rely on the absolute value transform.

As an aside, it seems that in my data this particular artefact varies across the brain. In particular it is more pronounced in those areas that are broad in functional delineation (association areas in particular). Whereas the primary sources show sources of one sign distributed across the cortical folds, and this activity is modulated (and changes sign) with condition.

What this means to me is that the size of the scout/ROI will to some extent determine one’s vulnerability to artefactual negatives. A very small ROI may erroneously point to negative/inward currents, whereas a larger (cross sulcus) ROI will actually produce a conservative estimate of the underlying activity (since negative/positive values will cancel out). Absolute values in the latter case may actually inflate source activation estimates.

In either case - I think I understand what needs be done in my case. Thank you for the detailed explanation - it helped me quite a bit.

Agatha