# The events and Annotations data structures¶

Events and Annotations are quite similar. This tutorial highlights their differences and similarities, and tries to shed some light on which one is preferred to use in different situations when using MNE.

Here are the definitions from the Glossary.

events
Events correspond to specific time points in raw data; e.g., triggers, experimental condition events, etc. MNE represents events with integers that are stored in numpy arrays of shape (n_events, 3). Such arrays are classically obtained from a trigger channel, also referred to as stim channel.
annotations
An annotation is defined by an onset, a duration, and a string description. It can contain information about the experiment, but also details on signals marked by a human: bad data segments, sleep scores, sleep events (spindles, K-complex) etc.

Both events and Annotations can be seen as triplets where the first element answers to when something happens and the last element refers to what it is. The main difference is that events represent the onset in samples taking into account the first sample value (raw.first_samp), and the description is an integer value. In contrast, Annotations represents the onset in seconds (relative to the reference orig_time), and the description is an arbitrary string. There is no correspondence between the second element of events and Annotations. For events, the second element corresponds to the previous value on the stimulus channel from which events are extracted. In practice, the second element is therefore in most cases zero. The second element of Annotations is a float indicating its duration in seconds.

See Reading an event file for a complete example of how to read, select, and visualize events; and Rejecting bad data (channels and segments) to learn how Annotations are used to mark bad segments of data.

## An example of events and annotations¶

The following example shows the recorded events in sample_audvis_raw.fif and marks bad segments due to eye blinks.

import os.path as op
import numpy as np

import mne

data_path = mne.datasets.sample.data_path()
fname = op.join(data_path, 'MEG', 'sample', 'sample_audvis_raw.fif')


Out:

Opening raw data file /home/circleci/mne_data/MNE-sample-data/MEG/sample/sample_audvis_raw.fif...
Read a total of 3 projection items:
PCA-v1 (1 x 102)  idle
PCA-v2 (1 x 102)  idle
PCA-v3 (1 x 102)  idle
Range : 25800 ... 192599 =     42.956 ...   320.670 secs


First we’ll create and plot events associated with the experimental paradigm:

# extract the events array from the stim channel
events = mne.find_events(raw)

# Specify event_id dictionary based on the meaning of experimental triggers
event_id = {'Auditory/Left': 1, 'Auditory/Right': 2,
'Visual/Left': 3, 'Visual/Right': 4,
'smiley': 5, 'button': 32}
color = {1: 'green', 2: 'yellow', 3: 'red', 4: 'c', 5: 'black', 32: 'blue'}

mne.viz.plot_events(events, raw.info['sfreq'], raw.first_samp, color=color,
event_id=event_id)


Out:

320 events found
Event IDs: [ 1  2  3  4  5 32]


Next, we’re going to detect eye blinks and turn them into Annotations:

# find blinks
eog_events = mne.preprocessing.find_eog_events(raw)

# Turn blink events into Annotations of 0.5 seconds duration,
# each centered on the blink event:
onset = eog_events[:, 0] / raw.info['sfreq'] - 0.25
annot = mne.Annotations(onset, duration, description,
orig_time=raw.info['meas_date'])

# plot the annotated raw


Out:

EOG channel index for this subject is: [375]
Filtering the data to remove DC offset to help distinguish blinks from saccades
Setting up band-pass filter from 2 - 45 Hz
Filter length of 8192 samples (13.639 sec) selected
Setting up band-pass filter from 1 - 10 Hz
Filter length of 8192 samples (13.639 sec) selected
Now detecting blinks and generating corresponding events
Number of EOG events detected : 46


## Working with Annotations¶

An important element of Annotations is orig_time which is the time reference for the onset. It is key to understand that when calling raw.set_annotations, given annotations are copied and transformed so that raw.annotations.orig_time matches the recording time of the raw object. Refer to the documentation of Annotations to see the expected behavior depending on meas_date and orig_time. Where meas_date is the recording time stored in Info. You can find more information about Info in The Info data structure.

We’ll now manipulate some simulated annotations. The first annotations has orig_time set to None while the second is set to a chosen POSIX timestamp for illustration purposes. Note that both annotations have different onset values.

# Create an annotation object with orig_time undefined (default)
annot_none = mne.Annotations(onset=[0, 2, 9], duration=[0.5, 4, 0],
description=['foo', 'bar', 'foo'],
orig_time=None)
print(annot_none)

# Create an annotation object with orig_time
orig_time = '2002-12-03 19:01:31.676071'
annot_orig = mne.Annotations(onset=[22, 24, 31], duration=[0.5, 4, 0],
description=['foo', 'bar', 'foo'],
orig_time=orig_time)
print(annot_orig)


Out:

<Annotations  |  3 segments : bar (1), foo (2), orig_time : None>
<Annotations  |  3 segments : bar (1), foo (2), orig_time : 2002-12-03 19:01:31.676071>


Now we create two raw objects and set each with different annotations. Then we plot both raw objects to compare the annotations.

# Create two cropped copies of raw with the two previous annotations
raw_a = raw.copy().crop(tmax=12).set_annotations(annot_none)
raw_b = raw.copy().crop(tmax=12).set_annotations(annot_orig)

# Plot the raw objects
raw_a.plot()
raw_b.plot()


Note that although the onset values of both annotations were different, due to complementary orig_time they are now identical. This is because the first one (annot_none), once set in raw, adopted its orig_time. The second one (annot_orig) already had an orig_time, so its orig_time was changed to match the onset time of the raw. Changing an already defined orig_time of annotations caused its onset to be recalibrated with respect to the new orig_time. As a result both annotations have now identical onset and identical orig_time:

# Show the annotations in the raw objects
print(raw_a.annotations)
print(raw_b.annotations)

# Show that the onsets are the same
np.set_printoptions(precision=6)
print(raw_a.annotations.onset)
print(raw_b.annotations.onset)


Out:

<Annotations  |  3 segments : bar (1), foo (2), orig_time : 2002-12-03 19:01:10.720100>
<Annotations  |  3 segments : bar (1), foo (2), orig_time : 2002-12-03 19:01:10.720100>
[42.955971 44.955971 51.955971]
[42.955971 44.955971 51.955971]


Notice again that for the case where orig_time is None, it is assumed that the orig_time is the time of the first sample of data.

raw_delta = (1 / raw.info['sfreq'])
print('raw.first_sample is {}'.format(raw.first_samp * raw_delta))
print('annot_none.onset[0] is {}'.format(annot_none.onset[0]))
print('raw_a.annotations.onset[0] is {}'.format(raw_a.annotations.onset[0]))


Out:

raw.first_sample is 42.95597082905339
annot_none.onset[0] is 0.0
raw_a.annotations.onset[0] is 42.955970883369446


It is possible to concatenate two annotations with the + operator (just like lists) if both share the same orig_time

annot = mne.Annotations(onset=[10], duration=[0.5],
description=['foobar'],
orig_time=orig_time)
annot = annot_orig + annot  # concatenation
print(annot)


Out:

<Annotations  |  4 segments : bar (1), foo (3), foobar (1), orig_time : 2002-12-03 19:01:31.676071>


Note that you can also save annotations to disk in FIF format:

>>> annot.save('my-annot.fif')


Or as CSV with onsets in (absolute) ISO timestamps:

>>> annot.save('my-annot.csv')


Or in plain text with onsets relative to orig_time:

>>> annot.save('my-annot.txt')


Total running time of the script: ( 0 minutes 9.169 seconds)

Estimated memory usage: 10 MB

Gallery generated by Sphinx-Gallery