rereferencing - neuralinterfacinglab/LabManual GitHub Wiki

What is re-referencing?

In electrophysiology, we always measure the difference in electrical potential between a contact of interest and a reference contact. This is done to eliminate noise related to the ground circuit. We generally choose one specific reference contact that serves as the 'baseline' for all other contacts. In our sEEG recordings, this is usually a contact located in white matter that does not show any epileptic activity. As comparison, in scalp EEG, this is usually a centrally located contact or an electrode placed on the earlobe.

Re-referencing means changing this reference electrode artificially, after the data has been recorded. The use of one single contact for a reference can bias the signal in other contacts. For example, a contact that is near the reference might have a similar voltage signal than a contact further away. The resulting amplitude (contact - reference) can thus be lower nearer to the reference, solely due to the spatial element. It is therefore common, at least in scalp EEG, to apply a re-referencing scheme that reduces this potential bias.

What are the re-referencing schemes?

The most common re-referencing scheme, stemming from scalp EEG, is the common average reference (CAR). Here, the bias is reduced because you take the average of all signals as the reference, thereby making each channel contribute equally. However, there may still be a bias since in sEEG especially the contacts are not equally spatially distributed in the brain. Also, it may spread signals of interest and/or noise across all contacts, potentially affecting conclusions that can be made regarding the origin of the signal.

The next most common is the bipolar reference (BPR), using the next nearest contact as the reference. In this method, each contact will have a different reference. This can reduce the bias since the spatial difference between the measured signals in roughly the same between all channels. It it thought to only represent local signals, since other signals are not represented in the equation. Since we are working with electrode shafts in sEEG, we can also do an electrode shaft reference (ESR). This is a CAR, but restricted to the specific shaft. This can reduce the spatial distribution issue from the CAR.

Then there are some variations of the first two methods: the laplacian reference (LPR), the average of the two adjacent channels (one on each side), and the common-median reference (CMR), the median instead of the average in the CAR. There are also data-driven re-referencing methods. Since there are still up-and-coming, we highlight only one of these methods here. This method applies an independent component analysis (ICA) as a spatial filter. Details can be found in Michelmann et al. (2018). Besides these, there are many other methods that you may encounter in the literature, each having more potential advantages and disadvantages.

What is the effect of re-referencing in practice?

We were curious to know how all of these methods actually affect our data. So, we applied all of the above methods to our grasp (hand movement) and speech (production) data. We evaluated the impact on how the signal in each channel is correlated to the task after the re-referencing, how correlated the channels are to one another (a measure of spatial spread) and how it actually changes decoding performance, our ultimate goal in BCIs.

We found that all of these re-referencing methods actually have little effects on our data, besides on the inter-channel correlations. Here's a breakdown:

There is a gradual decrease in inter-channel correlations with more local reference methods, as has been found before (Mercier et al., 2017, Li et al., 2018). This indicates that there is quite some signal spread in the way we record our signals (RAW) and in the CAR/CMR method which can be reduced with ESR/LPR/BPR. The ICA does not decrease the signal spread. If this metric is important, we'll see later.
There is barely a difference between the methods regarding the task-related activity. Only the ICA actually caused a big decrease in the maximum height (best channel) of the correlation between the EEG and the task in the, for the speech task only. If we look at small effects (~0.05), there is a slight increase for BPR in grasp and decrease with BPR in speech. This difference may be related to low (beta) vs high (high-gamma) frequency signals used in each task respectively. A similar effect of frequency range used has been found before in Liu et al. (2021).
The multivariate decoding (binary classes: movement vs rest) results are very similar to the task-related results. On average, every method improves the decoding accuracy a very tiny little bit (~0.01) compared to RAW, with BPR the best for grasp and ICA/BPR/CMR the worst for speech. Thus, the inter-channel correlations don't seem to be related to the decoding accuracy, but the task correlation does.
While there's only tiny differences in the height of the maximum task correlation, which channel this is can be different due to the re-referencing. This is important to be aware of in case you want to make neuroscientific interpretations regarding the spatial location of your relevant signals. Shifts in location seemed to be more dramatic for the grasp task, in which we looked at low frequency signals.

So... what should I do with my sEEG data?

First, don't spend as much time on this as I did. There is not one re-referencing scheme that will be best for sEEG in general, the choice depends on the task/research question/goal and, more specifically, the frequency range of interest. In general, I would recommend to go with the ESR method, since this seems to be a good middle-man.

Here's some specific recommendations based on your potential situation:

If you look at group-level results, it is probably not worth to dive into this further.
If you look at a single subject and try to maximise performance, you may want to evaluate some different methods. You can use a simple task correlation as a predictor.
If your goal is to test some machine learning models and evaluate decoding accuracies, it is probably not worth to dive into this further.
If your goal is to interpret the spatial representation of the task, where something happens in the brain, you may want to think about this a bit further. If you're looking at low frequency signals, you'd want to select a local reference like the BPR. In this way you can be a little more sure about the location, though the origin of low frequency signals can be difficult to determine nonetheless. If you're looking at high frequency signals, try to avoid the BPR as you may remove relevant signal, but otherwise still stick with a local reference (LPR/ESR) to reduce the bias.