SKUNKWRK 256 channel ICA filtering to remove eye artifacts

Original EEG data -> Flagclip(+-800 microVolt) -> 50 Hz lowpass filter (already highpassed during recording)

EEGSCORE used to find bad channels

Matlab scripts writes data in EEGLAB format with bad channels removed

EEGLAB used to clean out  time segments with several bad channels due to movement or other non-eye artifacts

This "clean" data will be input to the ICA computations
 

Created four sets of interleaved electrodes
    Adequate training for each ICA weight  (~30-40 training samples per weight)
    Matlab memory constraints

Exactly the same time samples are submitted to ICA for each interleave

ICA eye movement components are identified for each interleave (same number for each interleave)

These components are removed using EEGLAB and the ICA filtered interleaved sets are used to reconstruct the ICA filtered 256 channel
data set with MASK file identifying bad channels and cleaining done in EEGLAB prior to ICA

Rule of thumb for ICA:  at least 3*NCHAN^2 training points or 196608 for 256 channels.  Skunkwrk runs are 7 minutes at 500Hz sampling for 210,000 samples.  Thus there are just enough samples for training.  However, Matlab runs into memory problems with running this large an ICA.  Decided to break data into four interleaved sets of channels.  Each interleaved set contains the eye channels (Horizontal: 1,53,227,251; Vertical: 11,18,36,37,241,231) and consist of 71,75,71,69 channels, respectively.

For these interleaved sets there are approximately 210,000/(70^2) ~ 40 training samples per ica weight. 

Data is first passed through a 50Hz lowpass filter to reduce EMG, 60Hz line noise and other potential non-brain noise sources.

Used 'extended' ICA option to check for sub-Gaussian sources.  Training 'stop' limit set to 1E-07 (better than default 1E-06)

ICA components for sknk222r1_bcr_cl_intrlv1:

ICA component activations for sknk222r1_bcr_cl_intrlv1:

Original data for sknk222r1_bcr_cl_intrlv1:

ICA filtered data (components 1,2,3,5 removed) for sknk222r1_bcr_cl_intrlv1_icaF:

Power spectrum for all 420 seconds of data at FP1 for sknk222r1_bcr_cl_intrlv1:


Power spectrum for artifact free data (Gold standard, 133 sec). Over two-thirds of data contained artifacts and was scored out.
 

Plot in time domain of ICA filtered blink for sknk222_r1 at FP1 (which is one of the eye channels on the GSN256)

Blue - orig. data; ICA filtered data for interleave sets 1 - 4 are shown in Brown, Black, Green and Magenta, respectively. Vert. microVolts; Horiz. Time (secs)

Result for three data processing paths:

sknk211





sknk221






sknk222

Plot of data from ICA eye components only (1,2,3,5):


Scalp plot of power for these data:

Effect of ICA filtering on variance of asymmetry scores.  Used EEGLAB to look at sknk222r1_intrlv1_bcr_cl and the ICA filtered _icaF version. Data were converted to average reference in EEGLAB using the 61 channels in interleave 1.  Electrode power estimates and asymmetry scores were estimated from one second time windows for the entire length of the cleaned data with no windows crossing discontinuities where bad data were removed.  Means and variances were computed from log10(powerRight)-log10(powerLeft) alpha 8-13Hz for each of these one second windows.

sknk211r1_bcr_cl_intrlv1                                      sknk221r1_bcr_cl_intrlv1                   sknk222r1_bcr_cl_intrlv1
                            Mean        Variance                            Mean        Variance                            Mean        Variance
E36/E18 FP1/2    0.182       0.082                                 0.200           0.114                             0.060           0.051
ICA filtered          0.002       0.062                               -0.091           0.078                            -0.004           0.022

E39/E225 F3/4      -                 -                                  -0.334           0.117                            -0.168           0.181
ICA filtered            -                 -                                  -0.231           0.104                            -0.131           0.164

E54/E223 F7/8  -0.136        0.171                               -0.502           0.125                            -0.044           0.201
ICA filtered        -0.231        0.191                              -0.477            0.127                              0.106          0.148

ICA EOG filtering decreases variances in almost all cases and looks like it is most important at FP1/2 as would be expected

Shown below are histograms of the sknk221 data at FP1/2 with ICA filtered on right: