clean_rawdata & ASR

Why do I need this step?

Artifact Subspace Reconstruction (ASR) finds these bursts automatically. It is part of the clean_rawdata plugin, which ships with EEGLAB. As the official tutorial explains, it works in two stages:

1. It finds clean portions of the recording, the calibration data, and learns how large normal activity is.
2. It marks every short window whose activity is many times larger than the calibration data. The default is 20 standard deviations.

For recorded data the official tutorial advises removing the marked portions rather than correcting them. Correction was designed for real-time use, and its effect on offline analysis is not clearly understood.

How do I do it in EEGLAB?

Start with D1 hp0.1 notch60 loaded, as in the Starting point. It is already high-pass filtered.

1. Open Tools › then Reject data using Clean Rawdata and ASR.
2. Leave Remove channel drift (data not already high-pass filtered) unchecked, because the data are already filtered at 0.1 Hz.
3. Keep Process/remove channels checked, check Ignore these channels (ECG, EMG, …), and type EOG1 EOG2 next to it.
4. Uncheck the three channel criteria: Remove channel if it is flat for more than (seconds), Max acceptable high-frequency noise std dev, and Min acceptable correlation with nearby chans [0-1]. Bad channels belong to their own lesson.
5. Keep Perform Artifact Subspace Reconstruction bad burst correction/rejection checked, with 20 in Max acceptable 0.5 second window std dev.
6. Keep Remove bad data periods (when uncheck, correct using ASR) checked, and leave Use Riemanian distance metric (not Euclidean) - beta unchecked.
7. Keep Additional removal of bad data periods checked, with -Inf 7 in Acceptable [min max] channel RMS range (+/- std dev) and 25 in Maximum out-of-bound channels (%).
8. Keep Pop up scrolling data window with rejected data highlighted checked and press OK.
9. A scrolling window shows the original and cleaned data together. Scroll through it and look at what was removed.
10. Name the new dataset D1 asr. In the main window, compare Frames per epoch and Events with D1 hp0.1 notch60. The extra events are the boundary events.
11. Do not save D1 asr. The main path continues from D1 hp0.1 notch60.

How do I write it as code?

% ASR on a copy: no channel criteria (bad channels lesson), EOG channels ignored
EEG_asr = pop_clean_rawdata(EEG, 'FlatlineCriterion', 'off', 'ChannelCriterion', 'off', ...
    'LineNoiseCriterion', 'off', 'Highpass', 'off', 'BurstCriterion', 20, ...
    'WindowCriterion', 0.25, 'BurstRejection', 'on', 'Distance', 'Euclidian', ...
    'WindowCriterionTolerances', [-Inf 7], 'channels_ignore', {'EOG1', 'EOG2'});
EEG_asr.setname = 'D1 asr';

% How much data was kept, and how many boundary events were added
fprintf('Data kept: %.1f of %.1f s (%.1f%%)\n', EEG_asr.pnts / EEG_asr.srate, ...
    EEG.pnts / EEG.srate, 100 * EEG_asr.pnts / EEG.pnts);
fprintf('Boundary events: %d\n', sum(strcmp({EEG_asr.event.type}, 'boundary')));

% Effect on epochs: pop_epoch drops every epoch that contains a boundary event
epBefore = pop_epoch(EEG, {'square'}, [-0.2 0.8]);
epAfter  = pop_epoch(EEG_asr, {'square'}, [-0.2 0.8]);
fprintf('Epochs around square: %d before ASR, %d after\n', epBefore.trials, epAfter.trials);

% Original and cleaned data in one scrolling window
vis_artifacts(EEG_asr, EEG);

The script prints how many seconds were kept, how many boundary events were added, and how many epochs around square survive. pop_epoch is from station 5; here it only counts the epochs lost.

Which value should I choose?

When to use ASR, and when to avoid it:

How do I check it worked?

1. The scrolling window showed the original and cleaned data, and you saw what was removed.
2. The script check: it prints the seconds kept, the number of boundary events, and the epochs before and after.
3. The events increased in the main window by the number of boundary events.
4. The path is unchanged: D1 asr is not saved, and the next lesson starts from D1 hp0.1 notch60.

What mistakes do people make?

• Using correction instead of removal on recorded data. Keep Remove bad data periods (when uncheck, correct using ASR) checked.
• A threshold that is too low. At 3 or 5, ASR may remove blinks and large but normal brain activity. Start at 20, look at what was removed, then decide.
• Forgetting the boundary events. Each removed portion becomes a boundary event, and every epoch that contains one is dropped. Count the epochs before and after.
• Running ASR on data with a slow drift. ASR needs high-pass filtered data, as the plugin page notes. Filter first, or check Remove channel drift (data not already high-pass filtered).
• Not looking at the result. The official plugin page recommends that new users compare the original and cleaned data by eye. The scrolling window is there for that.

How do I write it in my paper?

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

High-amplitude bursts were identified with Artifact Subspace Reconstruction (burst criterion 20 SD; clean_rawdata plugin 2.11, EEGLAB 2026.1), and the affected periods were removed together with windows in which more than 25% of channels exceeded 7 SD ([N] s, [P]% of the data); EOG channels were excluded.

Going further

• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/cleanrawdata.html (6. Reject artifacts > c. Automated rejection)
• EEGLAB plugin: https://eeglab.org/plugins/clean_rawdata/ (clean_rawdata)
• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html (Issues associated with aggressive automated artifact removal before ICA)
• EEGLAB preprocessing #7: Removing bad data segments: https://www.youtube.com/watch?v=ZvbRiX84mgc (Full video)