It’s not what it looks like: Baseline Shifts

Cook et al. 2013
The implantable device from Cook et al. 2013

I’ve just come back from the fantastic IWSP7: Epilepsy Mechanisms, Prediction and Control conference in Melbourne. Having apparently outgrown the initial meetings’ focus on seizure prediction, this year covered all aspects from computational models, intracranial devices, to imaging in epilepsy. For those who don’t know – Melbourne is a great place for such a conference, since Mark Cook and colleagues have managed a couple of years ago to pull off a clinical trial of implantable intracranial recording devices designed for long-term ambulatory recordings, in addition to the potential for responsive neuromodulation. The set-up can be seen on the right (an image the conference conveners seemed to love), and was a first in the world of seizure prediction. [1]

But the meeting was brimming with exciting talks – many of which addressed quite fundamental aspects of what it is we are measuring with our current ways of recording the electrical activity of the brain. This is the first of a few blog posts that will deal with the different features we do not currently pick up with our electrical recordings of the brain.

Shifts in overall voltage are usually ignored

EEG recordings can be difficult, because what we are measuring (tiny little currents caused by brain activity that reach the scalp) so easily disappear in noise. Because of this, we do a whole lot of preprocessing, such as filtering and re-referencing of the original data. This makes the lines look nicer, and EEGs in general easier to read, but it also gets rid of some important information, such as large, relatively slow shifts in the baseline.

That this may be important can be seen from theoretical analyses, as well as other experimental work. In Victor Jirsa’s fantastic paper on modelling seizures last year [2], the modelling showed a very dramatic drop in the baseline just preceding a seizure, which allegedly was not present in the empirical data that the group tried to model. Yet a second look at the original data prompted by the modelling showed that indeed, the seizures in the original recordings, showed a large drop in baseline voltage, just as predicted by the modelling.

Jirsa et al. 2014
This two graphs show a computational seizure model (A) and an empirical recording of seizure-like events (B). Note the large drops in baseline when the seizure starts in the model, as well as in the empirical data. This is quite a convincing example of computational modelling predicting an unexpected data phenomenon. (Jirsa et al. 2014)


We also know that other measurement methods, such as magneto-encephalography (MEG), which has a better signal-to-noise ratio and does not require filtering out the baseline shifts, can show quite dramatic baseline-shift effects in experimental setups. Experiments that are in many ways akin to typical event-related-potentials in the EEG literature can huge effects on overall baseline, in addition to smaller event-related fluctuations, something we do not see in the EEG based literature, probably because of the different preprocessing steps (see image below from Maria Chait’s group [3]).

Andreou et al. 2015
This is an MEG experiment on complex repetitive auditory patterns. The researchers here were interested in the latency of the offset responses (which are fast and marked by arrows). Here, I wanted to highlight the huge shifts in the baseline that occur as well – something we generally exclude from analysis with EEG, but is easily shown in this MEG study (Andreou et al. 2015)


It is not clear exactly what occurs during these large baseline shifts, but they certainly present themselves as a phenomenon to be considered, both in spontaneous brain activity and seizures, and in evoked brain responses so commonly studied with EEG recordings. There are ways in which EEG may be able to capture these, particularly with newer, high signal-to-noise ratio recording techniques. This includes some changes in the preprocessing (e.g. to not filter out slow fluctuations, and to reference to an independent electrode), but technically could be possible. I’m excited to see what comes of this!

[1] Cook MJ et al. (2013) Lancet Neurol 12: 563-71. doi:10.1016/S1474-4422(13)70075-9
[2] Jirsa VK et al. (2014) Brain 137: 2210-30. doi:10.1093/brain/awu133
[3] Andreou LV et al. (2015) NeuroImage 110: 194-204. doi:10.1016/j.neuroimage.2015.01.052