I have EEG data where the recording was continuous during breaks between
blocks, so there are lots of movement artifacts during temporal spans that
I ultimately don't care about. If I just epoch the data, those temporal
spans go away, but I want to do blink rejection. Doing blink rejection on
an Epochs object doesn't seem to work out of the box, and doing it on each
epoch individually seems inefficient and possibly dodgy.
Question 1: is there a sensible way to run blink rejection on an Epochs
object that I haven't thought of?
Assuming I need to reject blinks using the Raw object, here are more
details: I don't have a separate EOG channel, so I'm using a forehead
electrode (Fp1) that seems to reflect the blinks quite well (based on
visual inspection of raw.plot()). But the movement artifacts are much
bigger than the blinks, so ICA is not working very well at catching the
actual blinks; in a ~70 minute recording it is detecting as few as 12
"blinks" for some subjects, and those events are localized around what are
clearly movement artifacts or other large between-block drifts.
Question 2: Other than manually zeroing out values in raw._data, is there a
way to mask or delete time spans of the Raw object that you want ICA to
ignore? I know about raw.crop() but that only works for beginning/end
times.
It might be helpful if you were a little more clear/specific about what
you're currently doing, so we can help suggest things you haven't thought
of
That said, I'm guessing you're using ica.find_bads_eog and ica.plot_scores?
I've been working on adapting my lab's EEGLAB workflows to MNE over the
past few months and my experience is that this routine works a lot better
for MEG than EEG data. For EEG (using various systems and channel counts),
this approach is pretty unreliable. In my experience ica.find_bads_eog at
the default threshold (3.0) often misses finding any blinks, and if you
drop the threshold down you tend to get false positives. I haven't found a
threshold that works reliably across subjects/datasets. If anyone has had
better success I'd love to hear about it though!
My approach is to fit ICA on the epochs as I want to analyze them, and
use ica.plot_properties to inspect every component and manually identify
the EOG-related components. This is usually very easy to do based on scalp
distribution, variance, and the epochs raster plots (and easy to teach
undergrads to reliably identify as well).
I've heard many people say they like to fit ICA on longer epochs than they
ultimately plan on analyzing. In principle, more data = more opportunities
to train ICA well. However, I've never done this as I've always been
satisfied with the results I get using a "normal" epoch length and manual
inspection. Fitting ICA on raw, continuous data - especially when there are
lots of very noisy periods such as breaks - is unlikely to work well
because ICA will just be trying to fit all the huge variance in your data
and so will - as you seem to be seeing - will tend to miss the blinks.
There's also a danger that real (brain) data will be lost.
My approach is to fit ICA on the epochs as I want to analyze them, and
use ica.plot_properties to inspect every component and manually identify
the EOG-related components. This is usually very easy to do based on scalp
distribution, variance, and the epochs raster plots (and easy to teach
undergrads to reliably identify as well).
Assuming the movement artifacts are much bigger than the blinks, I wonder
if this suggestion be adapted to this use case with something like:
1. Use `make_fixed_length_events` to make regularly-spaced events of some
duration (not sure if e.g., 1 sec or 10 sec would work better)
2. Create Epochs using a rejection threshold that throws away epochs based
on movement artifacts, but not the blink artifacts.
3. Run ICA on the Epochs.
one more thing. If the movement artifacts can get mapped to strong peak-to-peak
amplitudes the reject param in ICA can certainly help automatically
excluding these
segments from the ICA fit.
Thanks all, for the suggestions.? Annotations seemed like the right way to go, since I know the block start/end times already.? [Jaakko, I can report that annotations are working for me (both adding programmatically via `raw.annotations = mne.Annotations(...)` and also interactively.]? After annotating the between-block times as "bad-between-blocks", `ica.plot_components` looks good: 3 components that make sense for my experiment, and one on the forehead/face (the blinks).? Unfortunately neither `find_eog_events` nor `create_eog_epochs` seems to respect annotations (yet?).? So although I can ignore the between-blocks times when fitting the ICA, it doesn't seem to help with the next step of `ica.find_bads_eog`.? Please let me know if I'm misunderstanding something about this process; otherwise I'll report back after I've had a chance to try some of the other suggestions.
