- Powerful data surveyability: It is probably the only tool, as far as I know, that allows to check ALL the channels and ALL the data points in one window. One can perform visual sanity check for outliers and data stationarity at a glance. It saves you from eyeballing hundreds of pages of raw EEG data.
- Simple data cleaning: If you find a problem in your channels or time-series waveforms, you can seamlessly perform an essential-level data cleaning by chopping off bad channels/datapoints. Unlike clean_rawdata(), trimOutlier() does not add anything to data. Hence recommended for purists and skeptics of any automated data cleaning/correction algorithm. For example, if your cleaning strategy is a classical 'rejected trials EEG exceeds +/- 200 uV', then trimOutlier() does the job for you. When called from GUI, it provides interactive and intuitive user environment with which one can determine rejection criteria based on feedback from cut-and-try processes.
It keeps a log in EEG.etc.trimOutlier.cleanChannelMask [nbchan x 1 logical] for channel rejection and EEG.etc.trimOutlier.cleanChannelMask [data_length x 1 logical] for data point rejection.
This tutorial assumes the user understands the basic steps of loading data into EEGLAB. For more information, visit the EEGLAB tutorial.
To begin, load a continuous (un-epoched) data set and start the plugin under “Tools” of the EEGLAB window. Several figures and charts will pop up along with a prompt to clean the data.
The top left scalp map shows the topography of channel SD during all recording time, and the top right one shows the same image but after removing the channels with top 25% highest SD. By comparing these two scalp maps, you can see the scalp distribution of high amplitudes that is usually associated with artifacts.
(05/14/2024 Updated) Since 2019, the SD bar graph shows error bars that indicates 1 SD. The SD of SD is calculated by dividing the data into 100 bins with equal length with trimming, then calculate SD for each bin. My calculating the mean and SD across the 100 SDs, you can get mean and SD of channel SD values. This approach helps distinguish the following two cases: (1) High-variance state is stationary (i.e. better to reject the channel; in this case, the error bar is short); (2) High-variance state is transiend (i.e. better to retain the channel; the error bar is long).
Before continuing by clicking “Yes,” be sure to study the graphs to estimate threshold values. There is no need to worry about finding the perfect threshold values on the first try however; again, an advantage to this plugin is the interactive nature of the process.
The first input for filtering is the channel standard deviation upper bound. For the example data set, we chose an upper bound of 200 MICROV based on the “Standard deviation of all channels” graph. After clicking “Ok,” the plugin generates a graph of channel standard deviation data before and after rejection, and also displays the number of rejected channels and asks for confirmation (names of channels removed will be displayed during final step). New upper bound values may be chosen if the number of rejected channels is too high; for our experiment, we were comfortable with rejecting up to 10% of our channels.
If you have multiple subjects in your experiment you may wish to determine a threshold common to all subjects. One way to do this is by finding the standard deviation of the EEG data for each channel of each subject and sorting the values, then examining the plot of all the subject’s sorted data.
Click “Ok” again to confirm upper bound channel removal. A similar process occurs for rejection based on lower bound. Enter a lower bound value, click “Ok,” inspect the result, and alter the value or confirm by clicking “Ok” again. While the upper bound is meant to take care of “crazy” channels, the lower bound addresses “dead” channels. Therefore, it is best to avoid using 0 as the lower bound. 2 MICROV is a good value.
The next step is to reject data points. The colour key is the same as the first image: black represents mean, red is +/- 2SD, and blue is the envelope. Here we use a threshold value of 300 MICROV and a point spread width of 1ms. Click “Ok,” and examine the “After datapoint rejection” graph. Different values may be used for threshold and point spread width if the rejection is not satisfactory. Finalize the process by confirming.
Authors: Clement Lee and Makoto Miyakoshi. SCCN, INC, UCSD