Preventing losing/altering events when downsampling

[skjerns]

I have the following problem

Our trigger channel sometimes doesn't switch on all bits simultaneously, producing events in the STI101 channel (1000 Hz) like

0, 0, 0, 4, 5, 5, 5, 5, 5, 0, 0, 0

when I downsample these data using mne-bids-pipeline to 100 Hz, the resulting trigger channel values are downsampled to 0, 4, 0, such that the event is interpreted as a 4 instead of a 5.

Usually I would avoid this by extracting events beforehand with appropriate min_duration and shortest_event and resampling jointly (raw, events = raw.resample(100, events=events)). However, when using mne-bids-pipeline I don't have this option. The events are correctly inserted as annotations into the resampled pipeline data. However, converting back from annotations to events via mne.events_from_annotations(raw) will scramble the trigger code values, unless manually specified.

Is there any way to avoid this without relying on the _events.json sidecar (in which the correct event information is saved)?

[hoechenberger]

Hi!

The events are correctly inserted as annotations into the resampled pipeline data. However, converting back from annotations to events via mne.events_from_annotations(raw) will scramble the trigger code values, unless manually specified.

I don't really understand what you're saying here, could you please provide an example? Doesn't need to be an MWE, just some illustration of what happens vs what you would expect. This will also help us triage whether this is a Pipeline or an MNE-Python problem. Are you trying to "eject" your data from the Pipeline? Or where does your issue occur? Thank you!

[skjerns]

I don't really understand what you're saying here, could you please provide an example? Doesn't need to be an MWE, just some illustration of what happens vs what you would expect. This will also help us triage whether this is a Pipeline or an MNE-Python problem. Are you trying to "eject" your data from the Pipeline? Or where does your issue occur? Thank you!

Sure! I use various trigger codes that do not appear in order, e.g. a list of events extrated from mne.find_events would read 4, 99, 2, 4, 234. During the pipeline, these events (before resampling) are added as annotations to the raw object, with the use of the events.tsv (which got created by write_raw_bids(raw=raw, .., events=events,). Then later, if I want to convert from annotations back to events, the event codes would be simply enumerated in the order that they appeared, e.g. 1,2,3,1,4. However, as triggers are not in the same order for each participant, the resulting events code would be scrambled across participants. However, I'm also unable to access the regular events object after resampling, as some triggers get resampled in a wrong way. I could rely on the annotations as strings, but that makes things a bit more difficult and error prone, but would definitively be a workaround.

I am trying to use MNE to preprocess my data, without epoching. I.e. running ICA, downsampling, filtering etc, and then having the clean_raw.fif as input for all further processing :)

[hoechenberger]

Ah okay, so you're basically running all pre-processing steps to get the "clean" raw data, and that's when you stop using the pipeline and want to proceed working with the data in your own manual code / pipeline, did I get this right? 🙂

[skjerns]

Ah okay, so you're basically running all pre-processing steps to get the "clean" raw data, and that's when you stop using the pipeline and want to proceed working with the data in your own manual code / pipeline, did I get this right? 🙂

exactly! our final analysis is a bit more intricate than what mne-bids-pipeline offers

[hoechenberger]

Thanks! Could you share what the annotations on that clean file look like?

[skjerns]

sure! annotation.zip

I think this is intended behaviour in MNE, as far as I can tell, the original trigger values are not saved when running mne.annotations_from_events and when running the reverse mne.events_from_annotations, the resulting values are just enumerations?

[hoechenberger]

You can pass the event_id to control which IDs will be used :)