Fix ICA #36

wmvanvliet · 2018-04-27T13:54:02Z

ICA ignored the EEG channels, which means EOG artifacts where not cleaned from the EEG sensors, which means autoreject started rejecting a lot of data based on blinks in the EEG.

This should fix it. I'm re-running our conpy pipeline now (takes a while...) to see if we get decent results.

larsoner · 2018-04-27T13:59:17Z

According to the comments we still don't remove EOG-related artifacts. Do you need to change later steps to add this?

larsoner · 2018-04-27T13:59:46Z

scripts/processing/06-make_epochs.py

@@ -150,6 +150,7 @@ def run_epochs(subject_id, tsss=False):
    print('  Getting rejection thresholds')
    reject = get_rejection_threshold(epochs.copy().crop(None, reject_tmax),
                                     random_state=random_state)
+    del reject['eog']  # Don't reject epochs based on EOG


We probably still should reject based on EOG because we don't want to use trials with blinks. At least this was what we thought before.

larsoner · 2018-04-27T14:01:28Z

Actually 06-make_epochs.py already has ica.find_bads_eog, so I guess there is nothing else to do there.

Can you try leaving the EOG-based rejection in and see if this results in a reasonable number of trials? Or are too many rejected still?

wmvanvliet · 2018-04-27T14:22:20Z

Let's see...

larsoner · 2018-04-27T15:34:18Z

scripts/processing/06-make_epochs.py

@@ -166,6 +166,6 @@ def run_epochs(subject_id, tsss=False):

 # Here we use fewer N_JOBS to prevent potential memory problems
 parallel, run_func, _ = parallel_func(run_epochs, n_jobs=max(N_JOBS // 4, 1))
-parallel(run_func(subject_id) for subject_id in range(1, 20))
+parallel(run_func(subject_id) for subject_id in range(6, 20))


wmvanvliet · 2018-04-27T16:51:47Z

Leaving the EOG rejection limits in gives me the following rejection rates:

subject #rejected/#total (rejection rate)
Subject sub002: 175/887 (19.7 %)
Subject sub003: 183/882 (20.7 %)
Subject sub004: 404/889 (45.4 %)
Subject sub006: 178/883 (20.2 %)
Subject sub007: 011/885 (01.2 %)
Subject sub008: 309/882 (35.0 %)
Subject sub009: 622/889 (70.0 %)
Subject sub010: 426/885 (48.1 %)
Subject sub011: 384/884 (43.4 %)
Subject sub012: 351/888 (39.5 %)
Subject sub013: 392/883 (44.4 %)
Subject sub014: 241/887 (27.2 %)
Subject sub015: 177/883 (20.0 %)
Subject sub017: 642/883 (72.7 %)
Subject sub018: 005/882 (00.6 %)
Subject sub019: 234/885 (26.4 %)

larsoner · 2018-04-27T16:56:02Z

There is some reasonable scientific basis for rejecting trials where blinks occurred in such a visual paradigm. Even if the artifacts are rejected, the visual input is not the same. I am no visual expert, though, but this did play into our EOG-related decision IIRC

larsoner · 2018-04-27T19:44:06Z

IIRC most subjects don't have a usable 3-layer BEM so EEG is not used for LCMV/dSPM. So this probably won't make a huge difference for what we show in the paper, except maybe the sensor plots.

jasmainak · 2018-04-27T19:51:46Z

Marijn, can you share the html of the results once you build it?

On Fri 27 Apr 2018 at 21:44, Eric Larson ***@***.***> wrote: IIRC most subjects don't have a usable 3-layer BEM so EEG is not used for LCMV/dSPM. So this probably won't make a huge difference for what we show in the paper, except maybe the sensor plots. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#36 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/APHioixLt1KOMQWEVMPIiDPgmGqwEhqgks5ts3UHgaJpZM4TqaJ1> .

-- Sent from my iPhone

wmvanvliet · 2018-04-30T10:40:54Z

Bwah. With the ICA properly cleaning EEG and autoreject being performed on MEG+EEG only, the rejection rates are as follows:

subject #rejected/#total (rejection rate)
Subject sub002: 069/887 (07.8 %)
Subject sub003: 171/882 (19.4 %)
Subject sub004: 354/889 (39.8 %)
Subject sub006: 178/883 (20.2 %)
Subject sub007: 011/885 (01.2 %)
Subject sub008: 175/882 (19.8 %)
Subject sub009: 576/889 (64.8 %)
Subject sub010: 306/885 (34.6 %)
Subject sub011: 337/884 (38.1 %)
Subject sub012: 232/888 (26.1 %)
Subject sub013: 308/883 (34.9 %)
Subject sub014: 241/887 (27.2 %)
Subject sub015: 001/883 (00.1 %)
Subject sub017: 574/883 (65.0 %)
Subject sub018: 008/882 (00.9 %)
Subject sub019: 234/885 (26.4 %)

That's still a lot of dropped epochs. And worse yet: the grand average STC makes no sense at all. Guess it's back to the drawing board for me on this.

jasmainak · 2018-04-30T14:55:04Z

I’ll need to dig into this to see what’s actually going on ...

On Mon 30 Apr 2018 at 12:40, Marijn van Vliet ***@***.***> wrote: Bwah. With the ICA properly cleaning EEG and autoreject being performed on MEG+EEG only, the rejection rates are as follows: subject #rejected/#total (rejection rate) Subject sub002: 069/887 (07.8 %) Subject sub003: 171/882 (19.4 %) Subject sub004: 354/889 (39.8 %) Subject sub006: 178/883 (20.2 %) Subject sub007: 011/885 (01.2 %) Subject sub008: 175/882 (19.8 %) Subject sub009: 576/889 (64.8 %) Subject sub010: 306/885 (34.6 %) Subject sub011: 337/884 (38.1 %) Subject sub012: 232/888 (26.1 %) Subject sub013: 308/883 (34.9 %) Subject sub014: 241/887 (27.2 %) Subject sub015: 001/883 (00.1 %) Subject sub017: 574/883 (65.0 %) Subject sub018: 008/882 (00.9 %) Subject sub019: 234/885 (26.4 %) That's still a *lot* of dropped epochs. And worse yet: the grand average STC makes no sense at all. Guess it's back to the drawing board for me on this. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#36 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/APHiol17OPJJvsx8GyH3iHkv7AQ4HEW7ks5ttuo3gaJpZM4TqaJ1> .

-- Sent from my iPhone

wmvanvliet · 2018-06-19T16:46:20Z

I've played around a bit with the data. It seems the mags and EEG are just that noisy. In our Frontiers submission we've ended up setting fixed, lenient, rejection thresholds in order to retain enough data for connectivity analysis.

agramfort · 2018-06-20T06:59:09Z

you don't correct for ECG/EOG with ICA? just EOG and big artifacts with reject? how did you set the reject thresholds?

wmvanvliet · 2018-06-20T22:02:58Z

sorry for the confusion. We do apply ICA. We replaced autoreject with a fixed threshold.

agramfort · 2018-06-21T03:12:45Z

ok so I understand that for your use case autoreject was removing too much data.

wmvanvliet · 2018-06-21T05:46:40Z

Yes. I thought this was because ICA was not cleaning the EEG data (which it wasn't), but fixing the ICA did not resolve the issue.

jasmainak · 2018-06-21T05:52:15Z

Just to be clear, this was autoreject (global), right? I think the issue might be that it does not include data augmentation, but I'll have to try it

larsoner · 2018-06-21T13:12:48Z

Yes. I thought this was because ICA was not cleaning the EEG data (which it wasn't), but fixing the ICA

So there is something wrong with what ICA does currently that we could fix? If so, what is it / how did you fix it?

wmvanvliet · 2018-06-21T17:11:57Z

The ICA is currently not fitted to the EEG channels. This PR fixes this.

agramfort · 2018-06-21T17:38:14Z

I am afraid this will change a bit the results of our paper. I don't have time to rerun all to check if it's the case... :(

larsoner · 2018-06-21T19:23:07Z

Assuming MEG and EEG are fitted separately (?) then it should only affect the EEG sensor space estimates, as we could not do source imaging with EEG. But indeed someone will need to re-run from this step forward and upload the result to check if we want to incorporate.

The review is almost done (?) so if we do change it, we should make it clear in the README that a particular commit was used for the paper figures.

agramfort · 2018-06-22T11:45:40Z

ICA is fitted jointly so I guess it will have some impact. Hopefully tiny...

larsoner reviewed Apr 27, 2018

View reviewed changes

wmvanvliet force-pushed the fix_ica branch 2 times, most recently from 68845eb to f64adaf Compare June 21, 2018 17:01

Also fit ICA to EEG channels

3ea356e

wmvanvliet force-pushed the fix_ica branch from f64adaf to 3ea356e Compare June 21, 2018 17:09

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix ICA #36

Fix ICA #36

wmvanvliet commented Apr 27, 2018 •

edited

larsoner commented Apr 27, 2018

larsoner Apr 27, 2018

larsoner commented Apr 27, 2018

wmvanvliet commented Apr 27, 2018

larsoner Apr 27, 2018

wmvanvliet commented Apr 27, 2018 •

edited

larsoner commented Apr 27, 2018

larsoner commented Apr 27, 2018

jasmainak commented Apr 27, 2018 via email

wmvanvliet commented Apr 30, 2018

jasmainak commented Apr 30, 2018 via email

wmvanvliet commented Jun 19, 2018

agramfort commented Jun 20, 2018 via email

wmvanvliet commented Jun 20, 2018

agramfort commented Jun 21, 2018 via email

wmvanvliet commented Jun 21, 2018

jasmainak commented Jun 21, 2018

larsoner commented Jun 21, 2018

wmvanvliet commented Jun 21, 2018

agramfort commented Jun 21, 2018

larsoner commented Jun 21, 2018

agramfort commented Jun 22, 2018 via email

Fix ICA #36

Are you sure you want to change the base?

Fix ICA #36

Conversation

wmvanvliet commented Apr 27, 2018 • edited

larsoner commented Apr 27, 2018

larsoner Apr 27, 2018

Choose a reason for hiding this comment

larsoner commented Apr 27, 2018

wmvanvliet commented Apr 27, 2018

larsoner Apr 27, 2018

Choose a reason for hiding this comment

wmvanvliet commented Apr 27, 2018 • edited

larsoner commented Apr 27, 2018

larsoner commented Apr 27, 2018

jasmainak commented Apr 27, 2018 via email

wmvanvliet commented Apr 30, 2018

jasmainak commented Apr 30, 2018 via email

wmvanvliet commented Jun 19, 2018

agramfort commented Jun 20, 2018 via email

wmvanvliet commented Jun 20, 2018

agramfort commented Jun 21, 2018 via email

wmvanvliet commented Jun 21, 2018

jasmainak commented Jun 21, 2018

larsoner commented Jun 21, 2018

wmvanvliet commented Jun 21, 2018

agramfort commented Jun 21, 2018

larsoner commented Jun 21, 2018

agramfort commented Jun 22, 2018 via email

wmvanvliet commented Apr 27, 2018 •

edited

wmvanvliet commented Apr 27, 2018 •

edited