In sensor/apply_ica, we should not apply the ICA rejection threshold to the data before rejecting components

[hoechenberger]

We have a setting ica_reject to remove bad epochs before fitting ICA. The epochs we feed into ICA are separate from the epochs we want to clean: they're high-pass filtered and they have drop_bad(reject=ica_reject) applied. This ensures ICA can be fit on epochs without flux / voltage spikes etc

Now in the apply_ica step, when we load the "actual" epochs we wish to clean, instead of simply applying ICA, we first call drop_bad(reject=ica_reject) too. In my opinion this is a bug, as normally we should be assuming that applying ICA may at least remove some of those high-amplitude artifacts. But if we reject those epochs before cleaning them via ICA, they're gone forever.

mne-bids-pipeline/mne_bids_pipeline/steps/preprocessing/_07a_apply_ica.py

Lines 102 to 115 in 1bc1e82

	epochs = mne.read_epochs(in_files.pop("epochs"), preload=True)
	ica_reject = _get_reject(
	subject=subject,
	session=session,
	reject=cfg.ica_reject,
	ch_types=cfg.ch_types,
	param="ica_reject",
	)
	epochs.drop_bad(ica_reject)
	# Now actually reject the components.
	msg = f'Rejecting ICs: {", ".join([str(ic) for ic in ica.exclude])}'
	logger.info(**gen_log_kwargs(message=msg))
	epochs_cleaned = ica.apply(epochs.copy()) # Copy b/c works in-place!

I think we should remove the drop_bad() call here. Any remaining large-amplitude artifacts after ICA cleaning will be caught in the following step, preprocessing/ptp_reject.

In an EEG dataset I'm currently working with, applying the PTP rejection threshold before applying ICA leads to a loss of many epochs that could have been saved through ICA

cc @agramfort

[larsoner]

Agreed @hoechenberger feel free to fix it