Including bad channels in autoreject

💬 Support & Discussions
1

Hi all,

I use autoreject to find optimal peak-to-peak rejection thresholds on my MEG data. Before running autoreject (local), I exclude bad channels that are notoriously bad. For some parts, autoreject seems to respect this exclusion (there is a warning, that these bad channels will be ignored, and in the progress bars, the reduced number of good channels are shown).
However, when I plot the channels x epoch rejection matrix (good, bad, interpolated), Two of the channels that I had marked as bad, are consistently shown in blue across all epochs, ie. they are interpolated (see figure below).

image
image854×1076 27.5 KB

What is happening here? Is this just a problem with the illustration or are these channels used after all in the determination of the thresholds, which channels are interpolated, etc?

Thanks,
Eduard

2

Hello,

Could you share some code snippet to look at, and the output of the .info attribute on the objects passed to autoreject?

Mathieu

3

Hi Eduard,

I would suggest starting with one of the examples and the sample dataset. Do you observe this problem also there? If yes, we might have a bug or compatibility issue with newer MNE versions. Sharing a short code snippet as Mathieu suggested would also help in pinpointing the issue.

Mainak

4

Hi both of you,

Here the code snippet that produced the plot from my starting post:

# set parameters for autoreject
n_interpolates = np.array((1, 2, 3, 5, 7, 9))
consensus_percs = np.linspace(0.2, 0.7, 11)

# init autoreject object, picks are grads or mags respectively
rej = AutoReject(n_interpolates, consensus_percs, picks=picks,
                    thresh_method='bayesian_optimization')

# run the autoreject (local) algorithm
rej.fit(epochs)

# extract the log file
rej_log = rej.get_reject_log(epochs)

# plot
xlabels = reject_log.ch_names
image = reject_log.labels
image[image == 2] = 0.5  # move interp to 0.5
legend_label = {0: 'good', 0.5: 'interpolated', 1: 'bad'}
cmap = colors.ListedColormap(['white', 'blue', 'red'])
img = axes.imshow(image.T, cmap=cmap, vmin=0, vmax=1, interpolation='none')
plt.setp(axes, yticks=range(0, len(xlabels)), yticklabels=xlabels)
plt.setp(axes.get_yticklabels(), fontsize=2)
#add red box around rejected epochs
for idx in np.where(reject_log.bad_epochs)[0]:
    axes.add_patch(patches.Rectangle((idx - 0.5, -0.5), 1, len(xlabels),
                 linewidth=1, edgecolor='r', facecolor='none'))

The output of info I shall provide in about 2 weeks, as I am out of office for the next 2 weeks.

Eduard

5

Hi Eduard,

Can you see this: https://github.com/autoreject/autoreject/issues/309

looks like someone hit the same issue as you but didn’t get around to making a pull request. If you have bandwidth, any help would be welcome.

Mainak