Ch 8: EEG Artifacts, Their Detection, Influence, and Removal

This chapter goes over EEG artifacts, what they look like in data and the problems they cause, and how to remove them (and ramifications of trying to do so). Noise generally stems from the physical (eye movements, blinks, amplifier saturation), but also the cognitive (irrelevant processing). While this chapter discusses how to attenuate a lot of it, it is stressed that EEG ultimately is not a noise free environment.

8.1 Removing Data Based on Independent Components Analysis

Independent Components Analysis is a technique where one decomposes the EEG time series data into a set of independent components for the purposes of identifying independent sources of variants. Real world example would be using multiple microphones around a room of talking people, and using the weighted averages of the sounds from each microphone to deduce who is saying what. The result of independent component analysis would be a set of weights for each microphone such that the weighted sum of all the microphones isolates the voice of one person. With EEG, independent components analysis results in a set of weights of all electrodes such that each component is a weighted sum of activity at all electrodes, that can be used to isolate sources of brain electrical signals.

This can be used to clean EEG data by identifying components that hold artifacts. This can be judged via topography, and time-domain /frequency spectrum. However, it is never truly possible to completely separate noise and data. Its best to remain conservative and only remove what is clearly noise (such as with blink artifacts, which are the easiest to identify due to their topographical and time based characteristics).

[Figure 8.1]

The maximum number of components is the number of electrodes used in the experiment. Practically speaking, if you have many electrodes, it may be reasonable to pick a smaller number of components.

8.2 Removing Trials because of Blinks

It is not always practical to remove every trial that happens to contain blinks. Thankfully, blinks do not destroy EEG signal, but linearly sum on-top of it, which means it can be removed relatively easily. This can be done with regression techniques, but also with independent components analysis, which is effective for blinks and ocular motor artifacts. 

Telling subjects to blink and specific times can cause increased cognitive load that may not be intended. However, if user blinks during a stimulus, it may impact their response to it. In addition, long blinks may indicate tiredness, in which the data may need to be rejected for that reason.

8.3 Removing Trials because of Oculomotor Activity

Oculomotor activity also adds noise, particularly to frontal electrodes. These artifacts can be minimized by keeping visual stimuli to the center of the subjects vision. Eye tracking can also be used to aid in rejecting trials with eye movement. EOGs can also be used, but are not as effective in detecting the smaller movements known as microsaccades. These microsaccades can have impact on high frequency EEG data and is a topic of discussion. One of the best ways to minimize it is to make sure the stimulus is small so one does not need to move their eyes to capture the whole thing. 

Similar to before, while independent components analysis can b used to largely mitigate this noise, a subject who has been looking around a lot during the experiment may not be engaged and their data may need to be tossed.

8.4 Removing Trials based on EMG in EEG Channels

EMG noise shows as bursts of large amplitude signal within 20-40 Hz, and particular occur on electrodes around the face, neck, and ears. Large bursts caused by sneezing, coughing, or any large movement should result in the trial being removed.

[FIgure 8.3]

However, some subjects show some EMG noise even when they are relaxing. If testing hypothesis far from muscle groups and within the beta-band, it may be possible to still salvage the trial. 

If EMG noise has equal amplitude before and after trial period (during normalization period), this noise will be removed during baseline normalization. If EMG noise is present in all conditions, it will cancel out during condition comparisons. Independent component analysis can be used to remove the EMG noise if it is well separated.

8.5 Removing Trials Based on Task Performance

Cognitive Noise or Cognitive Artifacts refers to actual brain activity but that is not relevant to the testing hypothesis. One source of this may be error trials (in which the subject did something wrong). These types of trials should be removed or at least separated since the brain did not respond in the expected way. Trials preceding and following the erroneous trial should also be considered for removal as it may represent a brain state about to / over-correcting from the mistake.

In addition, consider removing trials that had a long reaction time (over 3 standard devveations slower), or extremely fast reaction time (<200 ms), or trials that had multiple presses or none at all. All of these indicate subjects not paying attention.

8.6 Removing Trials Based on Response Hand EMG

Building on the previous section, if possible use EMGs in addition to buttons to record user responses. This can be used to record partial errors (user about to press one button, but changing their mind and pushing the other). This is because EEG brain responses of partial errors are closer to that of EEG data of users making errors than that of users choosing the correct answer. This is easier to set up physically on the thumb (due to bigger muscles) and if the button takes some effort to push.

[Figure 8.4B]

Cohen describes they method they use to do this. They take the Z-transform of the derivative of the EMG signal (of each hand) before being rectified (absolute value). This is done to minimize variation between subjects / their arms. A partial error is identified when the incorrect hands Z-derivative signal exceeds 2 standard deviations during the time period between stimulus and actual button press. In addition, the magnitude of the peak should be at least 2 times larger than any peak in the time period from 300 ms before the stimulus to the stimulus itself (to aid in filtering out EMG noise). 

However, the above should be tuned for your particular experiment.

[Figure 8.4A]

8.7 Train Subjects to Minimize Artifacts

Telling subjects what not to do (moving, blinking, etc) helps minimize the odds of them producing EEG noise. Showing subjects the data coming in when they do motor actions is a great way to demonstrate your point as well.

8.8 Minimize Artifacts during Data Collection

Once again, you cannot rescue bad data. Observe EEG signal during the experiments, and intervene if a source of noise is detected.