Ch 9: Overview of Time-Domain EEG Analysis

This chapter discusses various aspect of analysis of Time-Series data, particularly in regards to the ERP

9.1 Event-Related Potentials

Since noise is theoretically random and signal isn't. If you take a bunch of trials, phase lock them at a t = 0 event and average them, the noise gets attenuated and the signal remaining is the ERP. This much knowledge is sufficient when using ERP for data inspection. For making inferences on cognitive processes, be wary, and research component overlap, component quantification, appropriate interpretation, and statistical procedures first.

9.2 Filtering ERPs

Time domain signal averaging can act as a low pass filter. One way this happens is with non-phased locked activity, which is lost during averaging because it inst locked to a phase, non-phased locked activities in the brain tend to be above 15 Hz. In addition, short high frequency events often jitter temporally between subjects, and thus like non-phase locked activity get lost during averaging.

Further filtering frequencies from ERPs is not always necessary, and is a debated topic. The benefits it can bring include reducing residual high frequency fluctuations, which reduces the possibility that a peak is only a noise spike. However there are also risks that a poorly executed filter results in ringing artifacts, as seen in the following figure with over-zealous 0-10 Hz and 5-15hz band-pass filters. Meanwhile the 0-40 Hz filter much closer resembles the raw data.

The particular danger with these ringing artifacts is that they can appear to be oscillations, which can lead to incorrect conclusions in your research.

Ripple artifacts can largely be avoided by constructing filters with gentle transition zones. Chapter 14 discusses proper filter construction to avoid this and other artifacts.

Another more fundamental issue with filtering ERPs is that low pass filters reduce temporal resolution of the data, which is arguably one of the main points of the ERP, and this effect becomes more pronounced the lower the top frequency cutoff is in the band-pass filer.

When filtered, generally the high end is ~20-30 Hz and the low end is ~5-10 Hz. Note that mathematically speaking, applying the same filter to all trials then averaging them is the same as averaging all the trials then applying the filter to it.

9.3 Butterfly Plots and Global Field Power / Topographical Variance Plots

A butterfly plot shows the ERP from all electrodes overlaid on the same figure (See A). It is generally useful for spotting obviously noisy electrodes.

GFP (Global Field Power) is the standard deviation of activity of all electrodes, and is computed by taking the standard deviation of all electrodes at all points in time (See B). Note that this is NOT the same as average amplitude, rather GFP shows how different the activity is in different brain areas, ergo if every brain region was active GFP would be low. GFP is useful as a visual inspection tool to ensure the timing of responses align with task events.

9.4 The Flicker Effect

The Flicker Effect is a general phenomena of brain activity syncing with an external rhythm.  For example, looking at a flickering 20 Hz light will cause 20 Hz rhythmic activity in the visual cortex, observable as a narrow band power or phase alignment increase., or directly in the ERP.

This effect also has other names on more specific contexts:
Visual -> SSVEP (Steady-State Visual Evoked Potential)
Auditory -> SSAEP (Steady-State Auditory Evoked Potential)

This effect can be used to "tag" processing of single stimulus (IE, showing multiple stimuli to the subject that flicker at different frequencies, then using SSVEP to deduce the one they were looking at). However, the flicker effect has a low temporal resolution as it takes a few hundred milliseconds to stabilize.

It should also be noted that the flicker effect is known to affect beyond just the initial sensory cortical area, with some research indicating that flickering at specific frequencies may increase cognitive processes that run at said frequencies.

The effect works up to around 100 Hz, though lower ones may invoke it stronger.

There is debate whether the flicker effect is the direct result of the ERPs from the sensing of the flicker, or if it results from neural oscillations.

9.5 Topographical Maps

A topographical map allows for seeing activities detected by electrodes over the area of the subjects head. Its a useful tool for determining if any electrodes are obviously bad, but it is also useful for glancing at brain activities over various areas over time.

Topographical maps interpolate data between the electrode points. The more electrodes, the more accurate the map will be (the less space will be calculated by interpolation). 

9.6 Microstates

In EEG and ERP map series, there are brief sub-second time periods where brain states tend to remain quasi-stable, before changing into drastically different landscapes. We call these quasi-stable states microstates. Their durations vary but tend to be in the alpha range (70-130 ms), and their topology tend to fit to 4 or 5 patterns. These states have been linked to various cognitive tasks, such as memory recall and language.

To identify microstates, consider a period of time in which topologies are relatively constant (relatively little change from one instant to the next). The temporal difference (also known as global map dissimilarity), in other words, is zero. This can last several hundreds of milliseconds, then suddenly from one point to the next there is major change in the topology, which  see as a sharp increase in the temporal difference. This sharp change is the changing from one microstate to another.

The stable maps can be used in an hierarchical clustering analysis to identify topographical distributions that characterize the topographical maps during times of stability, which is known as cluster mapping. From here, the time points can be labeled following the cluster map that is the most similar, which opens the door to other task and statistical analysis.

9.7 ERP Images

ERP images are another way to represent ERP data from a single electrode. Rather than averaging all trials together, they are stacked vertically and have their amplitude changes color coded in each line / trial. Once again this can be used as a tool to vet for bad data. 

However, in by controlling in what order trials are stacked, we can seek various things. For example, sticking by reaction time we can see if amplitude patterns center around reaction (the figure in A doesn't support this in this case). By aligning by voltage at a specific time, we can also look for patterns that way (as in Figure B)

ERP images can be smoothed using a 2D Gaussian.

This methodology of stacking data isn't only done with time-domain EEG data, but can also be done with frequency band specific data as well.