ICA-based AR (AR2)

ICA1 decomposition is used to screen IC timecourses for unusual activity measured as deviation from typical values. Moreover, bad sensors can again be detected using unusual topography patterns of ICs.

Bad channels detection

To provide additional detection of bad channels that were not classified as bad during the first stage of AR, the topographies of ICs are analyzed. We use the fact that normal brain sources are always picked up by many adjacent sensors. Usually, the source is most visible at the closest sensor(s), with the power gradually decreasing as the distance from the sensor increases. On the other hand, unique noise on a particular channel will appear as a topography with a very steep gradient that involves a single electrode only and is not visible at neighboring electrodes. This allows for detecting and removing such noisy channels. The recommended value for the ar.ica_topoZThr variable is 7.

FT7 electrode marked as bad due to its absolute weight z-score exceeding 7. Histogram shows the distribution of weights.
ar.ica_topoZThr = 7;             % rejction of sensors that deviate significantly / topography based

The component topography plots with histograms are saved in the 4_ICA1/[subjectID]/ICmixing_hist subdirectory.

ICA-based trial removal

The idea of trial-based artifact rejection is similar to that performed during the AR1 procedure. However, instead of channel time courses, here time courses of ICs are screened for artifacts. Usually only the first components are to be checked, which is set by the ar.ica_ic2screen parameter, with a typical value around 5 to 8. In the first step, individual ICs are screened, and trials where deviations are found above ar.ica_singleicZthe z-threshold are marked for rejection. Second, smaller deviations exceeding ar.ica_multiicZThr z-threshold but visible on several (ar.ica_multiicZCnt) components simultaneously at the same time can also be rejected. Exact values will differ depending on data quality and measurement environment, but the below values can be used as a rough guess:

% bad trials ic-based removal
ar.ica_ic2screen    = 6;    % how many IC to screen for artifacts. 
ar.ica_singleicZthe = 20;   % trial rejection based on z-scores of individual IC
ar.ica_multiicZThr  = 7;    % trial rejection based on z-scores of multiple IC simultaneously 
ar.ica_multiicZCnt  = 5;    % trial rejection based on z-scores of multiple IC simultaneously - no of IC that exceed ar.ica_multiicZThr

IC timecourses and topographies are saved in the 4_ICA1/[subjectID] subdirectory.

Sample component timecourse with trial 33 showing significant deviation from typical values. Y axis is scaled is z-points.

Customizing AR parameters

The resulting files and logs are stored in the 5_ICA1_AR subfolder. ICA1_AR_report.txt contains information on removed channels and trials, as in the example below:

DS: 4
04_CTR_SES1_SEGM

REMOVED 2/368 (1%)
1 trials removed due to single IC var (thr=20)
   300
2 trials removed due to multiple IC simultaneous var (thr=10; cnt=6)
   280
   300
1 removed sensors (thr=7)
 (T8) 

Saving data after AR2

Datasets after ICA-based AR are saved in the 5_ICA1_AR directory as *_ICA1_arj.mat files. They contain uncorrected (no EOG correction) signals with channels / trials removed in the AR2 procedure.

Last updated