Read ICA EEGLab: problems with AMICA decomposition

(possible) Bug

read_ica_eeglab() appears to not be able to read in AMICA (note: not runica) decompositions from an EEGLab .set file.

The error upon trying to instantiate ICA:

  AssertionError                            Traceback (most recent call last)
 <ipython-input-10-a6f5a4e0501a> in <module>
       5 from mne import preprocessing
       6 
 ----> 7 ica = preprocessing.read_ica_eeglab('full_filepath')
       8 
       9 
 
 ~/opt/anaconda3/envs/mne/lib/python3.8/site-packages/mne/preprocessing/ica.py in read_ica_eeglab(fname)
    2754 
    2755     n_ch = len(ica.ch_names)
 -> 2756     assert eeg.icaweights.shape == (n_components, n_ch)
    2757     # When PCA reduction is used in EEGLAB, runica returns
    2758     # weights= weights*sphere*eigenvectors(:,1:ncomps)';
 
 AssertionError: 

Steps to reproduce

ICA was performed with EEGLab’s AMICA (not runica) with PCA reduction to rank correct for interpolated channels.

Looking at the data in EEGLab, ICAweights = n_components x n_pcs, and ICAsphere = n_pcs x n_channels. I’m guessing read_ica_eeglab is assuming the icaweights matrix is already multiplied by icasphere.

Expected results

read_ica_eeglab should correctly read in pca reduced IC decompositions from AMICA (and possibly other algorithms – this might want to be tested, such as FASTICA).

Minimal reproducible example

61 channel recording (.set/.fdt)

[mmagnuski]

@michaelcgold Yes, that should be the simplest scenario. 😃 In other words, your problem does not seem specific to AMICA (at least at the surface).

[mmagnuski]

I don’t think AMICA is supported in mne right now. For example AMICA allows for multiple ICA models and provides a timecourse for their likelihood. Some of the ICs can be even shared between these models. Because of these differences mne-python could be able to import AMICA model only in the simplest scenario. But you are right - there might be an inconsistency in how ICA is read from eeglab in general, expecting no PCA reduction.