1-Hz filtering for ICA & copying weights

What does this show?

Why do I need this step?

Why 1 Hz helps ICA. Slow drift is large, and it changes over the whole recording. ICA then spends part of its effort on it, and the components become less clear. The official tutorial says ICA decompositions are notably higher quality, with less ambiguous components, when the data are high-pass filtered above 1 Hz, sometimes even 2 Hz.

Why 1 Hz harms the ERP. Slow ERP components such as P3 contain low frequencies. A 1 Hz filter removes part of them, so P3 becomes smaller. You will measure this yourself in this lesson’s script.

Why copying is valid. ICA does not learn a signal. It learns a fixed way to combine the channels, which the official tutorial calls a spatial filter. The two copies have the same channels and the same mixing in the head; only their frequencies differ. So the same weights separate the same sources in both copies.

How do I do it in EEGLAB?

Start with the two copies from the Starting point. Its script puts the main dataset in dataset 1 and the ICA copy in dataset 2.

First: check the two copies

1. Open the Datasets menu. You see D1 main as dataset 1 and D1 ica1hz as dataset 2.
2. Choose dataset 2. In the main window, ICA weights shows Yes.
3. Choose dataset 1. ICA weights shows No, and Channels per frame is the same as in dataset 2: 32 in D1.

Second: copy the weights

1. With dataset 1 current, open Edit › then Dataset info.
2. Next to ICA weights array or text/binary file (if any):, press from other dataset, type 2, then OK. Three fields fill in automatically: ALLEEG(2).icaweights, ALLEEG(2).icasphere, and ALLEEG(2).icachansind.
3. Press OK to close the window. ICA weights now shows Yes for dataset 1.

Third: compare the maps

1. With dataset 1 current, open Plot › then Component maps › then In 2-D.
2. Choose dataset 2 and open the same window. The maps are identical, in the same order. The blink component is in the same place in both.
3. Save dataset 1, the 0.1 Hz main copy and not the 1 Hz ICA copy, with File › then Save current dataset as as D1_main_ica.set. The ICLabel lesson starts from it.

When to use 2 Hz instead of 1

If many components stay ambiguous, with mixed or scattered maps, start again from dataset 1: filter a new copy at 2 Hz, run ICA on it, and copy its weights the same way. Dataset 1 stays at 0.1 Hz every time.

How do I write it as code?

% EEG_main: main dataset (0.1 Hz); EEG_ica: ICA copy (1 Hz) with weights
% Copy the decomposition back to the main dataset
EEG_main.icaweights  = EEG_ica.icaweights;
EEG_main.icasphere   = EEG_ica.icasphere;
EEG_main.icachansind = EEG_ica.icachansind;
EEG_main.icawinv     = [];   % recomputed by eeg_checkset
EEG_main.icaact      = [];
EEG = eeg_checkset(EEG_main);

% Check: same weights and scalp maps in both copies
fprintf('Same weights: %d\n', isequal(EEG.icaweights, EEG_ica.icaweights));
fprintf('Largest scalp map difference: %g\n', max(abs(EEG.icawinv(:) - EEG_ica.icawinv(:))));
fprintf('ICA components: %d\n', size(EEG.icawinv, 2));

% Why two copies: ERP at Pz around 'square' events in each copy
epMain = pop_rmbase(pop_epoch(EEG, {'square'}, [-0.2 0.8]), [-200 0]);
epIca  = pop_rmbase(pop_epoch(EEG_ica, {'square'}, [-0.2 0.8]), [-200 0]);
pz  = find(strcmp({EEG.chanlocs.labels}, 'Pz'));
win = epMain.times >= 300 & epMain.times <= 600;
erpMain = mean(epMain.data(pz, :, :), 3);
erpIca  = mean(epIca.data(pz, :, :), 3);
fprintf('Mean at Pz, 300-600 ms: 0.1 Hz copy %.2f uV, 1 Hz copy %.2f uV\n', ...
    mean(erpMain(win)), mean(erpIca(win)));
figure;
plot(epMain.times, erpMain, epIca.times, erpIca);
legend('Main dataset (0.1 Hz)', 'ICA copy (1 Hz)');
xlabel('Time (ms)'); ylabel('Amplitude (uV)');

EEG = pop_saveset(EEG, 'filename', 'D1_main_ica.set', 'filepath', pwd);

The first block is the same weight-copy lines as in the Running ICA lesson. The check prints Same weights: 1, a scalp map difference close to zero, and 31 components in D1.

The ERP block uses pop_epoch, which you will learn in station 5. Here it only shows the difference: in D1 the mean at Pz between 300 and 600 ms is about 4.3 µV in the main dataset, and about 3.1 µV in the 1 Hz copy. That is about 30% of the P3 lost if you continued on the ICA copy.

Which value should I choose?

How do I check it worked?

1. Main window: with dataset 1 current, ICA weights shows Yes.
2. Same maps: the component maps of dataset 1 and dataset 2 are identical and in the same order.
3. The script check: it prints Same weights: 1 and ICA components: 31 in D1.
4. The ERP stays whole: the mean at Pz between 300 and 600 ms is larger in the main dataset than in the 1 Hz copy, about 4.3 against 3.1 µV in D1.

What mistakes do people make?

• Continuing the ERP on the 1 Hz copy. P3 shrinks, as the script shows. The ICA copy is only for computing the weights; the analysis continues on the main dataset.
• Copying only two of the three fields. The official tutorial mentions icaweights and icasphere. Copy icachansind too, or EEGLAB does not know which channels the weights belong to. The from other dataset button copies all three.
• Different channels in the two copies. If you remove or interpolate a channel in one copy only, the weights no longer match. Do every channel step before you make the copy.
• Copying icaact from the ICA copy. Those are the time courses of the 1 Hz data. Leave icawinv and icaact empty and let eeg_checkset recompute them from the main dataset.
• Choosing a dataset number by guess. In from other dataset type the number of the ICA copy as shown in the Datasets menu. A wrong number copies other weights without an error.

How do I write it in my paper?

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

Independent component analysis (extended Infomax; Lee et al., 1999) was computed with runica ([31] components after PCA reduction to the data rank) on a copy of the data high-pass filtered at 1 Hz, and the resulting weights were applied to the 0.1 Hz high-pass filtered data.

Going further

• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html (How to deal with the aggressive high-pass filter applied before running ICA)
• EEGLAB Tutorial: https://eeglab.org/tutorials/05_Preprocess/Filtering.html (5. Preprocess data > a. Filtering)
• Makoto's preprocessing pipeline: https://eeglab.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline (§4.1; §13.5 Tips for copying ICA results to other datasets)
• Luck, Applied Event-Related Potential Data Analysis (1st edition): https://socialsci.libretexts.org/Bookshelves/Psychology/Biological_Psychology/Applied_Event-Related_Potential_Data_Analysis_(Luck) (Chapter 9: Artifact Correction with ICA)