Parsing events from raw data

This tutorial describes how to read experimental events from raw recordings, and how to convert between the two different representations of events within MNE-Python (Events arrays and Annotations objects).

In the introductory tutorial we saw an example of reading experimental events from a “STIM” channel; here we’ll discuss events and annotations more broadly, give more detailed information about reading from STIM channels, and give an example of reading events that are in a marker file or included in the data file as an embedded array. The tutorials Working with events and Annotating continuous data discuss how to plot, combine, load, save, and export events and Annotations (respectively), and the latter tutorial also covers interactive annotation of Raw objects.

We’ll begin by loading the Python modules we need, and loading the same example data we used in the introductory tutorial, but to save memory we’ll crop the Raw object to just 60 seconds before loading it into RAM:

import os
import numpy as np
import mne

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder, 'MEG', 'sample',
raw =


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
Reading 0 ... 36037  =      0.000 ...    60.000 secs...

The Events and Annotations data structures

Generally speaking, both the Events and Annotations data structures serve the same purpose: they provide a mapping between times during an EEG/MEG recording and a description of what happened at those times. In other words, they associate a when with a what. The main differences are:

  1. Units: the Events data structure represents the when in terms of samples, whereas the Annotations data structure represents the when in seconds.

  2. Limits on the description: the Events data structure represents the what as an integer “Event ID” code, whereas the Annotations data structure represents the what as a string.

  3. How duration is encoded: Events in an Event array do not have a duration (though it is possible to represent duration with pairs of onset/offset events within an Events array), whereas each element of an Annotations object necessarily includes a duration (though the duration can be zero if an instantaneous event is desired).

  4. Internal representation: Events are stored as an ordinary NumPy array, whereas Annotations is a list-like class defined in MNE-Python.

What is a STIM channel?

A stim channel (short for “stimulus channel”) is a channel that does not receive signals from an EEG, MEG, or other sensor. Instead, STIM channels record voltages (usually short, rectangular DC pulses of fixed magnitudes sent from the experiment-controlling computer) that are time-locked to experimental events, such as the onset of a stimulus or a button-press response by the subject (those pulses are sometimes called TTL pulses, event pulses, trigger signals, or just “triggers”). In other cases, these pulses may not be strictly time-locked to an experimental event, but instead may occur in between trials to indicate the type of stimulus (or experimental condition) that is about to occur on the upcoming trial.

The DC pulses may be all on one STIM channel (in which case different experimental events or trial types are encoded as different voltage magnitudes), or they may be spread across several channels, in which case the channel(s) on which the pulse(s) occur can be used to encode different events or conditions. Even on systems with multiple STIM channels, there is often one channel that records a weighted sum of the other STIM channels, in such a way that voltage levels on that channel can be unambiguously decoded as particular event types. On older Neuromag systems (such as that used to record the sample data) this “summation channel” was typically STI 014; on newer systems it is more commonly STI101. You can see the STIM channels in the raw data file here:

raw.copy().pick_types(meg=False, stim=True).plot(start=3, duration=6)
plot 20 events from raw


Removing projector <Projection | PCA-v1, active : False, n_channels : 102>
Removing projector <Projection | PCA-v2, active : False, n_channels : 102>
Removing projector <Projection | PCA-v3, active : False, n_channels : 102>

You can see that STI 014 (the summation channel) contains pulses of different magnitudes whereas pulses on other channels have consistent magnitudes. You can also see that every time there is a pulse on one of the other STIM channels, there is a corresponding pulse on STI 014.

Converting a STIM channel signal to an Events array

If your data has events recorded on a STIM channel, you can convert them into an events array using mne.find_events(). The sample number of the onset (or offset) of each pulse is recorded as the event time, the pulse magnitudes are converted into integers, and these pairs of sample numbers plus integer codes are stored in NumPy arrays (usually called “the events array” or just “the events”). In its simplest form, the function requires only the Raw object, and the name of the channel(s) from which to read events:

events = mne.find_events(raw, stim_channel='STI 014')
print(events[:5])  # show the first 5


86 events found
Event IDs: [ 1  2  3  4  5 32]
[[27977     0     2]
 [28345     0     3]
 [28771     0     1]
 [29219     0     4]
 [29652     0     2]]

If you don’t provide the name of a STIM channel, find_events() will first look for MNE-Python config variables for variables MNE_STIM_CHANNEL, MNE_STIM_CHANNEL_1, etc. If those are not found, channels STI 014 and STI101 are tried, followed by the first channel with type “STIM” present in raw.ch_names. If you regularly work with data from several different MEG systems with different STIM channel names, setting the MNE_STIM_CHANNEL config variable may not be very useful, but for researchers whose data is all from a single system it can be a time-saver to configure that variable once and then forget about it.

find_events() has several options, including options for aligning events to the onset or offset of the STIM channel pulses, setting the minimum pulse duration, and handling of consecutive pulses (with no return to zero between them). For example, you can effectively encode event duration by passing output='step' to mne.find_events(); see the documentation of find_events() for details. More information on working with events arrays (including how to plot, combine, load, and save event arrays) can be found in the tutorial Working with events.

Reading embedded events as Annotations

Some EEG/MEG systems generate files where events are stored in a separate data array rather than as pulses on one or more STIM channels. For example, the EEGLAB format stores events as a collection of arrays in the .set file. When reading those files, MNE-Python will automatically convert the stored events into an Annotations object and store it as the annotations attribute of the Raw object:

testing_data_folder = mne.datasets.testing.data_path()
eeglab_raw_file = os.path.join(testing_data_folder, 'EEGLAB', 'test_raw.set')
eeglab_raw =


Reading /home/circleci/mne_data/MNE-testing-data/EEGLAB/test_raw.fdt
<Annotations | 154 segments: rt (74), square (80)>

The core data within an Annotations object is accessible through three of its attributes: onset, duration, and description. Here we can see that there were 154 events stored in the EEGLAB file, they all had a duration of zero seconds, there were two different types of events, and the first event occurred about 1 second after the recording began:


{'rt', 'square'}

More information on working with Annotations objects, including how to add annotations to Raw objects interactively, and how to plot, concatenate, load, save, and export Annotations objects can be found in the tutorial Annotating continuous data.

Converting between Events arrays and Annotations objects

Once your experimental events are read into MNE-Python (as either an Events array or an Annotations object), you can easily convert between the two formats as needed. You might do this because, e.g., an Events array is needed for epoching continuous data, or because you want to take advantage of the “annotation-aware” capability of some functions, which automatically omit spans of data if they overlap with certain annotations.

To convert an Annotations object to an Events array, use the function mne.events_from_annotations() on the Raw file containing the annotations. This function will assign an integer Event ID to each unique element of raw.annotations.description, and will return the mapping of descriptions to integer Event IDs along with the derived Event array. By default, one event will be created at the onset of each annotation; this can be modified via the chunk_duration parameter of events_from_annotations() to create equally spaced events within each annotation span (see Making multiple events per annotation, below, or see Making equally-spaced Events arrays for direct creation of an Events array of equally-spaced events).


Used Annotations descriptions: ['rt', 'square']
{'rt': 1, 'square': 2}
[[128   0   2]
 [217   0   2]
 [267   0   1]
 [602   0   2]
 [659   0   1]]

If you want to control which integers are mapped to each unique description value, you can pass a dict specifying the mapping as the event_id parameter of events_from_annotations(); this dict will be returned unmodified as the event_dict.

custom_mapping = {'rt': 77, 'square': 42}
 event_dict) = mne.events_from_annotations(eeglab_raw, event_id=custom_mapping)


Used Annotations descriptions: ['rt', 'square']
{'rt': 77, 'square': 42}
[[128   0  42]
 [217   0  42]
 [267   0  77]
 [602   0  42]
 [659   0  77]]

To make the opposite conversion (from Events array to Annotations object), you can create a mapping from integer Event ID to string descriptions, and use the Annotations constructor to create the Annotations object, and use the set_annotations() method to add the annotations to the Raw object. Because the sample data was recorded on a Neuromag system (where sample numbering starts when the acquisition system is initiated, not when the recording is initiated), we also need to pass in the orig_time parameter so that the onsets are properly aligned relative to the start of recording:

mapping = {1: 'auditory/left', 2: 'auditory/right', 3: 'visual/left',
           4: 'visual/right', 5: 'smiley', 32: 'buttonpress'}
onsets = events[:, 0] /['sfreq']
durations = np.zeros_like(onsets)  # assumes instantaneous events
descriptions = [mapping[event_id] for event_id in events[:, 2]]
annot_from_events = mne.Annotations(onset=onsets, duration=durations,

Now, the annotations will appear automatically when plotting the raw data, and will be color-coded by their label value:

raw.plot(start=5, duration=5)
plot 20 events from raw

Making multiple events per annotation

As mentioned above, you can generate equally-spaced events from an Annotations object using the chunk_duration parameter of events_from_annotations(). For example, suppose we have an annotation in our Raw object indicating when the subject was in REM sleep, and we want to perform a resting-state analysis on those spans of data. We can create an Events array with a series of equally-spaced events within each “REM” span, and then use those events to generate (potentially overlapping) epochs that we can analyze further.

# create the REM annotations
rem_annot = mne.Annotations(onset=[5, 41],
                            duration=[16, 11],
                            description=['REM'] * 2)
 rem_event_dict) = mne.events_from_annotations(raw, chunk_duration=1.5)


Used Annotations descriptions: ['REM']

Now we can check that our events indeed fall in the ranges 5-21 seconds and 41-52 seconds, and are ~1.5 seconds apart (modulo some jitter due to the sampling frequency). Here are the event times rounded to the nearest millisecond:

print(np.round((rem_events[:, 0] - raw.first_samp) /['sfreq'], 3))


[ 5.     6.5    8.     9.5   11.    12.501 14.001 15.501 16.999 18.499
 41.    42.5   44.    45.5   47.    48.5   50.   ]

Other examples of resting-state analysis can be found in the online documentation for mne.make_fixed_length_events(), such as Compute envelope correlations in source space.

Total running time of the script: ( 0 minutes 14.217 seconds)

Estimated memory usage: 111 MB

Gallery generated by Sphinx-Gallery