FIX/ENH: loading EDF annotations

[mbillingr]

I made sure that the stim channel does not get resampled, and that the int16 values (each containing two characters) are not clamped.

Parsing the stim channel to get events of the form [onset, duration, 'annotation'] remains to be done.

I'm not sure where to store these events, though. Currently I consider not to store the events at all. Instead, a member function would simply parse and return the event list. Any thoughts?

[agramfort]

@t3on @christianbrodbeck I guess this is for you :)

[coveralls]

Coverage increased (+0.12%) when pulling a83f0c9 on kazemakase:fix-edf-annotations into 9a88a37 on mne-tools:master.

[mbillingr]

I've added a parser as a private member of RawEDF. Not sure how to expose the functionality, though.
What is the correct format of the event list?

BTW, how do you get coveralls to automatically comment on PRs? :)

[agramfort]

we need the help of our EDF experts to review :) ping @t3on @christianbrodbeck

I have no experience with this format...

in the MNE way events are used to generate a stim channel that can be parsed with find_events function

[mbillingr]

in the MNE way events are used to generate a stim channel that can be parsed with find_events function

In that case I might have got it wrong. I thought the stim channel was specific to EDF.
Maybe it is necessary to convert the EDF annotation channel to an MNE compatible stim channel?

[agramfort]

In that case I might have got it wrong. I thought the stim channel was
specific to EDF.
Maybe it is necessary to convert the EDF annotation channel to an MNE
compatible stim channel?

isn't that the case already?

find_events works on an EDF raw?

[mbillingr]

It does not work on data where events are stored as EDF+ annotations: http://www.edfplus.info/specs/edfplus.html#edfplusannotations

[larsoner]

@kazemakase Coveralls will automatically comment once the build is passing (it's currently not). But it's actually a bit better to rely on your own local nose/coverage output, though, since Travis only runs a subset of the full test suite.

[christianbrodbeck]

@t3on ? I've never worked with EDF either...

[mbillingr]

@eric89gxl I was just wondering because i can't get coveralls to comment on PRs in the SCoT repo even when the build is passing.

[larsoner]

Ahh, for a different repo... you have to enable it on coveralls.io somewhere.

Assuming you've seen that option and enabled it, you also have to have .travis.yml configured properly. For some reason some of the repos I work on it works properly, and others it doesn't. I haven't tracked down what subtle differences between the repos cause some to get commented on, and others not. It appears to work for this repo, so you could try tracking down the differences between your travis config and ours.

[mbillingr]

Yep, that option is enabled and coveralls receives coverage information from travis. Good to know it doesn't work everywhere. Maybe it doesn't like capital letters in the repo name? :) I'll have another look at .travis.yml.

[teonbrooks]

sorry, just got notified. my email was silent on this ;) I am not clear on what the issue is. there is a private function, _read_annot, that takes care of the annotation file with the format described in the MNE manual. if the triggering is stored in the data, it is also capable of parsing it with the specification of the stim channel.

[mbillingr]

@t3on I assume you refer to the stim channel and event files as described here: http://martinos.org/mne/stable/manual/browse.html#events-and-annotations

Everything is fine if the EDF files in question were created according to that specification. However, I am trying to load data recorded by someone who never even heard of MNE. There is no stim channel or annotation file. Instead, there is a channel that contains ascii encoded tags (see my link above), which the current implementation is unable to read correctly.

[teonbrooks]

this shouldn't be deleted. this is used for when the sampling across different channels aren't unified. it does a resampling of the lower sampled channels to match the max sampling rate.

[mbillingr]

I did not delete it entirely (see line 242). Instead, I added separate handling of the annotation channel (which I thought was stim_channel).

[teonbrooks]

ah, sorry, my mistake, didn't see the elif.

[mbillingr]

btw, what happens if you resample the stim channel (assuming it is a correct MNE stim channel)? Wouldn't the implicit filter hurt it?

[teonbrooks]

Instead, there is a channel that contains ascii encoded tags (see my link above), which the current implementation is unable to read correctly.

do you have an example of a file that has the text annotation as channel in the data?

[mbillingr]

Sure, I was working with that file: http://www.physionet.org/physiobank/database/eegmmidb/S001/S001R06.edf

[mbillingr]

If i got this right, MNE expects a certain stim_channel where events are encoded. There are two ways to obtain a stim_channel for an edf file:
a) it is stored in the edf
b) an annotation file is parsed that replaces the stim channel

I have added a parameter tal_channel which can be used to specify an EDF+ annotation channel. I plan to create the stim_channel from the tal_channel if no annotation file is specified.

Parsing the tal channel works already, but how are events encoded in the stim channel?

[agramfort]

Parsing the tal channel works already, but how are events encoded in the stim channel?

Have a look at the find_events function for example

[coveralls]

Coverage remained the same when pulling 4a96214 on kazemakase:fix-edf-annotations into 9a88a37 on mne-tools:master.

[coveralls]

Coverage increased (+0.14%) when pulling 9c25cf7 on kazemakase:fix-edf-annotations into 9a88a37 on mne-tools:master.

[mbillingr]

Is there a way to encode simultaneous events in the stim channel?

[agramfort]

No. You need unique events

[larsoner]

If you need to, you can conceptually do something like this using a binary representation. Each event number n becomes encoded as 2 ** (n - 1). So event type 3 is encoded as the number 4, while event number 5 is encoded as the number 16. Thus simultaneous 3 and 5 events would be encoded as 20 (their sum). It's not too tough to synthesize this representation, or use bit masking to get back the original values.

If this is too annoying, you can also include multiple stim channels in a FIFF file, they just need different channel names. IIRC usually STI101 is just a binary combination (as above) of several other channels (which only take on values of zero or one) named something like STI001, STI002 or so.

[dengemann]

whitespaces missing between operators