Channel interpolation

Why do I need this step?

1. The average reference needs channels spread over the whole head. A missing channel on one side tilts the average toward the other side.
2. Comparing participants needs the same channels for everyone.
3. Scalp maps draw a gap where a channel is missing.

But interpolation has a price. An interpolated channel is a mix of the other channels, so it adds no new information. That is why each interpolated channel reduces the data rank by one. You will need this in the ICA lessons.

How do I do it in EEGLAB?

Start with the Starting point. Its script puts the dataset from before channel removal, D1 hp0.1 notch60, in dataset 1, and the main dataset D1 badch in dataset 2. You will need the number of dataset 1 for your own data, because it holds the locations of all channels.

First: practice on a good channel

1. With D1 badch current, open Edit › then Select data. Type C3 in Channel(s), check the box under on->remove these, then press OK.
2. Name the new dataset D1 noC3. In the main window, Channels per frame is now 31.
3. Open Tools › then Interpolate electrodes. A window titled Interpolate channel(s) — pop_interp() appears.
4. Press Use all channels from other dataset, type 2, the number of the D1 badch dataset, in Dataset index, then OK. The name C3 appears under What channel(s) do you want to interpolate.
5. In Interpolation method, keep Spherical, leave Time range [min max] (s) empty, then press OK.
6. The command window prints Interpolating 1 channels.... Name the dataset D1 C3 back. The channel count is 32 again.
7. Compare C3 in D1 badch and in D1 C3 back with Plot › then Channel data (scroll). The two signals are very similar.
8. These are exercise copies, so do not save them. Go back to D1 badch.

Second: interpolation on your own data

1. With the dataset you removed bad channels from as current, open Tools › then Interpolate electrodes.
2. Press Use all channels from other dataset and type the number of the dataset from before the removal. The names of the removed channels appear.
3. Keep Spherical and press OK. Name the result D1 interp and save it.
4. In D1 no channel was removed, so if you try, EEGLAB says No new channels selected. The main dataset stays D1 badch, and the re-referencing lesson starts from it.

How do I write it as code?

% Rebuild the removed channels from the full channel list (fullLocs)
if ~isempty(badChans)
    EEG = pop_interp(EEG, fullLocs, 'spherical');
end
EEG.setname = 'D1 interp';
fprintf('Channels after interpolation: %d\n', EEG.nbchan);
EEG = pop_saveset(EEG, 'filename', 'D1_interp.set', 'filepath', pwd);

% Exercise: remove a good channel, rebuild it, compare with the original
testChan = 'C3';
EEG_drop = pop_select(EEG, 'rmchannel', {testChan});
EEG_back = pop_interp(EEG_drop, fullLocs, 'spherical');
iOrig = find(strcmp({EEG.chanlocs.labels}, testChan));
iBack = find(strcmp({EEG_back.chanlocs.labels}, testChan));
r = corrcoef(double(EEG.data(iOrig, :)), double(EEG_back.data(iBack, :)));
fprintf('Correlation between original and interpolated %s: %.2f\n', testChan, r(1, 2));

The variable fullLocs comes from the Starting point script: a copy of all channel locations before any removal. Always keep it at the start of processing.

Which value should I choose?

How do I check it worked?

1. The channel count is complete: Channels per frame in the main window equals the original number. In D1 it is 32.
2. Same order: in Plot › then Channel locations › then By name the interpolated channels appear in their places.
3. The exercise: the script prints a correlation close to 0.98 between the original and the interpolated C3.
4. The signal makes sense: in Plot › then Channel data (scroll) the interpolated channel looks like its neighbors, with no flat line and no jumps.

What mistakes do people make?

• Re-referencing before interpolating. An average reference over missing channels leans toward the side with more channels. Interpolate first, then re-reference.
• Losing the original locations. Once a channel is removed, EEGLAB only knows where it was from an earlier dataset or from fullLocs. Keep the locations before any removal.
• Forgetting that the rank dropped. If you interpolate two channels, the data rank is two below the channel count, and three below after an average reference: 32 channels give rank 29. Write this number down and compare it with what pop_runica prints when you run ICA.
• Trusting edge channels too much. A channel at the edge of the cap, such as T7, has neighbors on one side only, so its interpolation is less accurate than for a central channel such as C3.

How do I write it in my paper?

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

Removed channels were interpolated using spherical spline interpolation (pop_interp, EEGLAB 2026.1).

Going further

• EEGLAB Tutorial: https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html (Issues with data rank deficiencies)
• Makoto's preprocessing pipeline: https://eeglab.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline (§7 Interpolate all the removed channels; §7.2)
• 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 7: Inspecting the EEG and Interpolating Bad Channels)