Automated pipelines: when to use them

Why do I need this step?

These pipelines save a lot of time, especially with large datasets. But each one makes dozens of decisions for you: which channel is bad, which component is an eye, which segment to delete. If you do not know these decisions, you cannot defend them in your paper or notice when they go wrong.

A ready-made pipeline saves time, but it does not excuse you from understanding every step in it.

That is why this lesson comes after you built the pipeline yourself. In it you compare the six common pipelines, then run one of them (PREP) on D1 and check its output the same way you would check any pipeline.

Six pipelines in one table

All of them are separate tools, not bundled with EEGLAB. Versions are as they were when this lesson was written.

Whatever pipeline you choose, check three things in its output before you continue:

1. The channels: how many channels were removed or interpolated, and which ones? Compare them with what you see in Plot › then Channel data (scroll).
2. The spectrum: compare the spectrum before and after. The line peak should become clearly smaller, and the rest of the spectrum should not collapse.
3. The report: read the pipeline’s report for each participant. A participant with too many removed channels or components needs a look, or exclusion.

How do I do it in EEGLAB?

First: install PREP

1. Open File › then Manage EEGLAB extensions, search for PrepPipeline and install it (the latest release when this lesson was written is 0.57.0). Restart eeglab, and the item Run PREP pipeline appears in the Tools menu.

Second: look at the data before

2. Start from D1 raw, as in “Starting point”. These are the data as downloaded, without filtering.
3. Open Plot › then Channel spectra and maps and keep the defaults. Note the sharp peak at 60 Hz, as in the line noise lesson.

Third: run PREP

4. Open Tools › then Run PREP pipeline. The PREP pipeline control panel window opens with six buttons, one for each stage.
5. Click Reference. In Reference channels, Evaluation channels and Re-referenced channels, type [1 3:5 7:32], that is, all channels except EOG1 (channel 2) and EOG2 (channel 6). Click Ok.
6. Click Line noise and check that Line frequencies removed shows 60. PREP computes it from the sampling rate: multiples of 60 below Nyquist, which is 64 Hz in D1. Click Ok.
7. Click Ok in the control panel. PREP prints each stage in the command window and takes about a minute on D1. When it finishes, it writes a summary file (.html) and a detailed report (.pdf) in your working folder, and asks you to name the new dataset. Name it D1 prep.

Fourth: check the output

8. Open the .html summary file in a browser. Look for the list of interpolated channels and the list of channels still noisy after the reference.
9. Open Plot › then Channel spectra and maps again on D1 prep. The 60 Hz peak is smaller but still there, clearly so in a few channels. PREP’s line noise step reduces the peak; it does not remove it completely. If the remaining peak matters for your analysis, add the band-stop filter from the line noise lesson after PREP.
10. Compare with your result in the bad channels lesson: there clean_rawdata did not remove any channel from D1. If PREP interpolated a channel, open Plot › then Channel data (scroll) and look at it yourself before you accept the decision.

How do I write it as code?

Continues from “Starting point”. PREP’s own function prepPipeline is the clearest way to call it from a script, as in the example that ships with the plugin. The script does not write the report; it prints the results directly.

% PREP is not bundled with EEGLAB: install it from the extension manager
plugin_askinstall('PrepPipeline', 'prepPipeline', true);

% Keep a copy of the raw data to compare spectra
EEG_raw = EEG;

% Scalp channels only; EOG is cleaned of line noise but not referenced
eegChans = find(~ismember({EEG.chanlocs.labels}, {'EOG1', 'EOG2'}));

params = struct();
params.name = 'D1';
params.lineFrequencies = 60;
params.detrendChannels = 1:EEG.nbchan;
params.lineNoiseChannels = 1:EEG.nbchan;
params.referenceChannels = eegChans;
params.evaluationChannels = eegChans;
params.rereferencedChannels = eegChans;

% Line noise reduction, robust reference, bad channel interpolation
[EEG, params] = prepPipeline(EEG, params);
EEG.setname = 'D1 prep';

% What did PREP decide?
nd = EEG.etc.noiseDetection;
if ~strcmp(nd.errors.status, 'good')
    error('PREP stopped with status: %s', nd.errors.status);
end
fprintf('PREP version: %s\n', nd.version);
fprintf('Interpolated channels: %s\n', ...
    strjoin({EEG.chanlocs(nd.interpolatedChannelNumbers).labels}, ', '));
fprintf('Still noisy after reference: %s\n', ...
    strjoin({EEG.chanlocs(nd.stillNoisyChannelNumbers).labels}, ', '));

% Spectrum before and after
figure; pop_spectopo(EEG_raw, 1, [], 'EEG', 'freqrange', [1 64], 'freq', [10 60]);
figure; pop_spectopo(EEG, 1, [], 'EEG', 'freqrange', [1 64], 'freq', [10 60]);

% PREP leaves the data unfiltered: high-pass before ICA or epochs
EEG = pop_eegfiltnew(EEG, 'locutoff', 0.1);

EEG.filename = ''; EEG.filepath = ''; EEG.datfile = ''; EEG.saved = 'no';
[ALLEEG, EEG, CURRENTSET] = pop_newset(ALLEEG, EEG, 0, 'setname', 'D1 prep', 'gui', 'off');
eeglab redraw;

Which value should I choose?

How do I check it worked?

• The command window prints PREP version: PrepPipeline0.57.0.
• The list of interpolated channels is printed, even if empty. Compare it with what you found in the bad channels lesson.
• The 60 Hz peak in D1 prep is clearly smaller than in the raw spectrum (about 11 dB above its neighbours instead of about 23 dB), and the alpha peak at 10 Hz keeps its shape.
• The channel count of D1 prep is still 32.

What mistakes do people make?

• Trusting the pipeline without opening its report. Every pipeline makes mistakes on some participants. Read the report for each participant, and look at the data of any participant with many removed channels.
• Leaving the line frequency at its default for 50 Hz data. PREP computes multiples of 60 by default. In data recorded with 50 Hz mains, it works on the wrong frequency and leaves the real peak.
• Expecting PREP to remove the line peak completely. In D1 the 60 Hz peak drops by about half in dB and stays visible. Look at the spectrum after PREP, and add a band-stop filter if you need the peak gone.
• Assuming PREP filtered the data. The output of PREP is not high-passed. If you go straight to ICA or epochs, the slow drift is still there.
• Changing pipelines between participants. Every participant in one study goes through the same pipeline with the same settings. If you change a setting, rerun everyone.
• Applying a pipeline outside its population. HAPPE and MADE were designed for children, and BeMoBIL for movement. Their settings may not suit adult data recorded at rest, and the reverse.

How do I write it in my paper?

A template: edit the data, settings and versions to match what you ran.

Line noise at 60 Hz was attenuated, and robust average referencing with bad channel detection and interpolation was performed, with the PREP pipeline (prepPipeline, PrepPipeline plugin [0.57.0], EEGLAB 2026.1), using the scalp channels as reference and evaluation channels; [n] channels were interpolated.

Going further

• EEGLAB Tutorial: https://eeglab.org/tutorials/11_Scripting/automated_pipeline.html (11. Scripting > e. Automated pipeline)
• PREP pipeline (EEG-Clean-Tools): https://github.com/VisLab/EEG-Clean-Tools (README, Releases (0.57.0))
• RELAX: https://github.com/NeilwBailey/RELAX (README)
• What is the optimal automated EEG pipeline?: https://www.youtube.com/watch?v=mW7u2YEK3ZY (Full video)