Insufficient data for ICA

AdamHull · July 11, 2025, 5:19pm

Hi there,

I'm trying to use ICA (PICARD) to remove ocular artifacts from my MEG data:

but when I do so I get this warning in MATLAB:

as well as this warning in the Brainstorm window:

Do I really need 48 minutes of data for ICA to work properly? Or is this warning caused by the apparent rank inconsistency? The warning says this inconsistency is most likely because I am using Linux 64-bit Matlab, but I am running Matlab (2023a) on MacOS. I saw this thread on this rank deficiency problem but it does not seem to have been resolved.

Curiously, the ICA seems to work okay despite the warning, in that so far I have been able to consistently identify IC's that correlate with the VEOG and HEOG channels, and removing them seems to do a good job removing the ocular artifacts.

Any guidance would be much appreciated!

Raymundo.Cassani · July 15, 2025, 4:46pm

This comes from a very conservative rule of thumb, where number of time samples is a k multiplied by nChannels^2 . This comes from the fact that the unmixing matrix that is estimated with ICA has nChannels^2. So in Brainstorm the warning is raised if there are less than 10 time samples for each element in the unmixing matrix.

10*(270^2) = 729000 samples --> 2916 s with fs =250

As mentioned, this is very conservative requirement, and it is possible to find artifactual components with less samples.

Are there flat channels in your recording? Or channels that were interpolated? Those will reduce the rank.

Since you are working with MEG sensors and there are VEOG and HEOG channels, you may want to tests your results against artifact removal with SSP:

https://neuroimage.usc.edu/brainstorm/Tutorials/ArtifactsSsp

AdamHull · July 16, 2025, 2:55pm

Oh I see, thank you!

No flat channels, and I don't believe there was any interpolation unless that was done in the downsampling (from 2400 to 250Hz)? When I export the recording to Matlab and calculate the rank I get 270.

Raymundo.Cassani · July 16, 2025, 3:11pm

The issue is in the way the rank is computed inside picard:

github.com/mind-inria/picard

matlab_octave/picard.m

5f9754589


      
          function tmprank2 = getrank(tmpdata)
              % Function originally written in EEGLAB by Arnaud Delorme
              tmprank = rank(tmpdata);
              %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
              %Here: alternate computation of the rank by Sven Hoffman
              %tmprank = rank(tmpdata(:,1:min(3000, size(tmpdata,2)))); old code
              covarianceMatrix = cov(tmpdata', 1);
              [E, D] = eig (covarianceMatrix);
              rankTolerance = 1e-7;
              tmprank2=sum (diag (D) > rankTolerance);
              if tmprank ~= tmprank2
                  fprintf('Warning: fixing rank computation inconsistency (%d vs %d) most likely because running under Linux 64-bit Matlab\n', tmprank, tmprank2);
                  tmprank2 = min(tmprank, tmprank2);
              end
          end

Specifically on the test against rankTolerance = 1e-7, as eigenvalues for MEG signals are smaller

So, there is no reason to worry on this warning

AdamHull · July 16, 2025, 5:00pm

Ah okay I see, thank you for your help!