ARTIFACT REJECTION

1
ARTIFACT REJECTION

• In cognitive electrophysiological experiments
• Based on Luck 2005

2
What are the artifacts?

• Blinks
• Eye movements
• Skin Potentials
• Muscle activity

3
Why artifacts are problematic

• Large relative to signal, so just a few may
  significantly reduce S/N (depending on how many
  trials you have)
• Some artifacts may be systematic, either
  occurring more in one condition than another, or
  being loosely time-locked to data
• E.g., eye-blinks to certain types of visual stim
• NOTE any kind of artifact systematicity is very
  problematic, and the following techniques will
  not really help, so try to prevent this on
  front-end!!

4
Artifact rejection as signal detection

• Select a threshold value, and if signal exceeds,
  make a response (exclude)
• Hit correctly excluded artifactual response
• Miss failed to exclude artifactual response
• False alarm excluded a clean response
• Correct rejection failed to exclude a clean
  response
• Same trade-offs in setting threshold as in
  classical signal detection situationmore hits at
  expense of more false alarms

5
By-hand rejection perspectives

• Luck
• The only difference between by hand
  rejection and automatic rejection is that it
  uses the experimenters visual system instead of
  a computer algorithm to determine the extent to
  which an artifact appears to be present and uses
  an informal, internal threshold to determine
  which trials to reject. The advantage of the
  approach is that the human visual system can be
  trained to do an excellent job of differentiating
  between real artifacts and normal EEG noise.
  However, a well-designed computer algorithm may
  be just as sensitive, if not more so.
• Me
• Automatic rejection is worth time-savings, IF you
  choose the algorithm yourselfnot just the
  default provided by your softwareand you test it
  and tune it before letting it go. Even though
  this will take time, if you have a dataset of any
  complexity, it will still save time in the end

6
Eye blinks

• Each eye has electrical gradient with positive at
  front and negative at back blinking results in
  deflections of this voltage
• Mainly monophasic deflection of 50-100 mV with
  duration 200-400 ms

7
Eye blinks polarity

• Opposite in polarity for sites above vs below
  eyethis is why it is nice to have a VEOG
• Provides clear signature of blink

8
Eye-blinks polarity

• Wise Note from Luck
• Because of the polarity reversal exhibited by
  blinks, you should always be suspicious of an
  experimental effect that is opposite in polarity
  at electrode sites above versus below the eyes
• Also, remember that any signals that are picked
  up more by the reference electrode than others
  will appear in inverted form in all electrodes
  using the reference

9
Eye-blink measures

• Hard voltage threshold at channels close to eyes
• Problem variation in baseline voltage can bring
  it over threshold, resulting in false alarms
• Peak-to-peak voltage measure
• Difference between minimum and maximum voltages
  within an epoch
• Less distorted by slow changes in baseline

10
Eye-blink measures step function

• Even peak-to-peak might catch slow shifts if they
  are big enough
• Can minimize this by using a step function
• Define step of x width, e.g. 100 ms
• For each point, take the difference between the
  average amplitude of preceding 100 ms and
  following 100 ms
• At end of epoch, take max diff and compare to
  threshold
• Step function also filters out high-frequency
  activity (blink isnt high frequency)
• Differential step function subtracts location
  above eye from below, takes advantage of polarity

11
Eye-blinks Choosing threshold

• Better to tailor to individual, although depends
  on type of artifact
• But this can be problematic in between-subjects
  designs, if the groups differ in type or number
  of artifacts
• Luck recommends for each subject, starting with a
  default threshold based on previous experience,
  then apply to set of trials and see if it is
  working
• Easy to test with polarity signature from VEOG

12
Eye-blinks how to prevent

• Ask subjects to wear glasses instead of contacts
• Can keep individual-use eyedrops handy
• Use short trial blocks or miniblocks (with
  automatic breaks of 20-30 s every so often).
• Dont be shy to tell people that they need to
  stop blinkinga lot of people just forget, all it
  takes is a few words from you to remind them to
  watch it

13
Eye movements

• Effect of eye movements have been systematically
  measured
• A bipolar recording of voltage between electrodes
  at locations immediately adjacent to the two eyes
  will yield deflection of 16 mV per degree of eye
  movement
• Voltage falls off predictably with electrode
  distance from eyes

14
Eye Movements

• Can use a step function to detect on HEOG
• Since less recognizable than blinks, may want to
  ask subject to move eyes back and forth a bit at
  beginning to get a baseline for the artifact
• If youre afraid that some conditions might
  engender systematic eye movement, you can compare
  averages at HEOG for the relevant conditions to
  show that movement isnt present

15
Slow voltage shifts

• Caused by changes in skin impedance due to
  sweating
• Try to keep room cool
• Can be limited by reducing impedance successfully
  at beginning of experiment
• The greater the initial impedance, the greater
  the changes in impedance
• Slight movement of electrodes also results in
  slow shifts
• As long as relatively rare, not a big deal to
  include

16
Amplifier saturation

• Slow voltage shifts can sometimes cause amplifier
  to saturate (reach highest value representable
  until AC reset)
• Also known as (confusingly) blocking
• If frequent, can reduce by changing gain on
  amplifier
• Good algorithm for finding is X-within-Y-of-peak,
  which finds maximum value within a trial and then
  counts number of points at or near maximum
• Applied to each channel separately, as they can
  saturate independently

17
Alpha waves

• Occur most when subjects are tired
• Although some people have them even when they are
  alert
• Particularly problematic when using constant
  stimulus rate b/c alpha rhythm can become
  entrained to stimulation rate such that they
  arent reduced in averaging
• Include 50 ms of jitter
• Too close to frequency range of ERP data to be
  easy to reject without a lot of false positives

18
Muscle activity

• Very high freq much is eliminated by low-pass
  filter in amplifier
• Luck says not usually necessary to reject trials
  with EMG, but maybe this depends on your
  amplifier
• Rejection
• Perform Fourier Transform on each trial and
  reject based on amount of high-frequency power
• Calculate difference in voltage between each
  consecutive pair points in trial and threshold on
  this difference
• Some stimuli elicit reflexive muscle
  twitchesusually short lived, so need to look at
  time domain and not frequency for these

19
Artifact Correction

• Various options
• Record response associated with artifact and
  subtract from waveforms, after multiplied by a
  propagation factor
• Do a source analysis (dipole modeling) to isolate
  ocular activity
• Do ICA
• But a little dangerous

20
Useful references from Luck