How to handle unstable mixing matrix estimation warning when using ICA.fit()

šŸ’¬ Support & Discussions
1
  • MNE-Python version: 0.24
  • MAC OS 11.6.1

Iā€™m encountering a warning when attempting to fit my ica to my raw data. I have a 10-minute EEG recording consisting of 46 channels with sampling rate 250 Hz.

Iā€™ve tried to use the syntax given under the API reference for ICA.

import mne
from mne.preprocessing import ICA
raw = mne.io.read_raw_edf('myfile.edf',preload=True)
raw_filt = raw.filter(l_freq=1, h_freq=None)
ica = ICA(method='fastica',n_components=.999999, random_state=6,max_iter="auto")
ica.fit(raw_filt)

In the fitting step, I receive the following warning:

> RuntimeWarning: Using n_components=0.999999 (resulting in n_components_=6) may lead to an unstable mixing matrix estimation because the ratio between the largest (46) and smallest (2.2e-05) variances is too large (> 1e6); consider setting n_components=0.999999 or an integer <= 5

Iā€™m not sure how to address this, particularly since Iā€™ve followed the suggestion and set n_components=.999999 . I observed that setting n_components=5 did not produce a warning, but Iā€™m not sure why.

Also, do I apply the fit to the original raw data or the filtered version?

`

2

Hi @PaulF,

I have just encountered the same issue and I am not sure how to fix it. Did you end up finding a solution or an explanation to the problem?

Giuseppe

3

@Giuseppe-1993 what kind of data are you using? MEG, EEG, ā€¦? How many channels? Thanks!

4

Hi @richard,

thank you for your attention. I am using EEG data, 64 channels, recorded using BioSemi. I have a total of 30 participants and this ICA issue seems to happen only on particular datasets. I am now inspecting the data to see if perhaps the dataset itself is corrupted but it does not look like it.

5

@Giuseppe-1993 would you mind sharing one of the affected datasets and a minimal working example to reproduce the issue youā€™re seeing?

You can share a link to the data via email: richard.hoechenberger@gmail.com

6

Hi Richard,

Unfortunately, I am not able to share either the dataset or the code due to IP agreements. However, after careful inspection, I can conclude that the behaviour was very likely due to the low signal-to-noise ratio. After rejecting manually segments of the data that seemed problematic, the ICA issue seems to stop. I will keep you updated should this happen again.

Giuseppe