Note

Go to the end to download the full example code.

The Evoked data structure: evoked/averaged data#

This tutorial covers the basics of creating and working with evoked data. It introduces the Evoked data structure in detail, including how to load, query, subset, export, and plot data from an Evoked object. For details on creating an Evoked object from (possibly simulated) data in a NumPy array, see Creating MNE-Python data structures from scratch.

As usual, we start by importing the modules we need:

# Authors: The MNE-Python contributors.
# License: BSD-3-Clause
# Copyright the MNE-Python contributors.

import mne

Creating Evoked objects from Epochs#

Evoked objects typically store EEG or MEG signals that have been averaged over multiple epochs, which is a common technique for estimating stimulus-evoked activity. The data in an Evoked object are stored in an array of shape (n_channels, n_times) (in contrast to an Epochs object, which stores data of shape (n_epochs, n_channels, n_times)). Thus, to create an Evoked object, we’ll start by epoching some raw data, and then averaging together all the epochs from one condition:

root = mne.datasets.sample.data_path() / "MEG" / "sample"
raw_file = root / "sample_audvis_raw.fif"
raw = mne.io.read_raw_fif(raw_file, verbose=False)

events = mne.find_events(raw, stim_channel="STI 014")
# we'll skip the "face" and "buttonpress" conditions to save memory
event_dict = {
    "auditory/left": 1,
    "auditory/right": 2,
    "visual/left": 3,
    "visual/right": 4,
}
epochs = mne.Epochs(raw, events, tmin=-0.3, tmax=0.7, event_id=event_dict, preload=True)
evoked = epochs["auditory/left"].average()

del raw  # reduce memory usage

You may have noticed that MNE informed us that “baseline correction” has been applied. This happened automatically during creation of the Epochs object, but may also be initiated (or disabled) manually. We will discuss this in more detail later.

The information about the baseline period of Epochs is transferred to derived Evoked objects to maintain provenance as you process your data:

print(f"Epochs baseline: {epochs.baseline}")
print(f"Evoked baseline: {evoked.baseline}")

Basic visualization of Evoked objects#

We can visualize the average evoked response for left-auditory stimuli using the plot() method, which yields a butterfly plot of each channel type:

evoked.plot()

Like the plot() methods for Raw and Epochs objects, evoked.plot() has many parameters for customizing the plot output, such as color-coding channel traces by scalp location, or plotting the global field power alongside the channel traces. See Visualizing Evoked data for more information on visualizing Evoked objects.

Subsetting Evoked data#

Evokeds are not memory-mapped

Evoked objects use a data attribute rather than a get_data() method; this reflects the fact that the data in Evoked objects are always loaded into memory and never memory-mapped from their location on disk (because they are typically much smaller than Raw or Epochs objects).

Unlike Raw and Epochs objects, Evoked objects do not support selection by square-bracket indexing. Instead, data can be subsetted by indexing the data attribute:

print(evoked.data[:2, :3])  # first 2 channels, first 3 timepoints

To select based on time in seconds, the time_as_index() method can be useful, although beware that depending on the sampling frequency, the number of samples in a span of given duration may not always be the same (see the Time, sample number, and sample index section of the tutorial on Raw data for details).

Selecting, dropping, and reordering channels#

By default, when creating Evoked data from an Epochs object, only the primary data channels will be retained: eog, ecg, stim, and misc channel types will be dropped. You can control which channel types are retained via the picks parameter of epochs.average(), by passing 'all' to retain all channels, or by passing a list of integers, channel names, or channel types. See the documentation of average() for details.

If you’ve already created the Evoked object, you can use the pick(), pick_channels(), pick_types(), and drop_channels() methods to modify which channels are included in an Evoked object. You can also use reorder_channels() for this purpose; any channel names not provided to reorder_channels() will be dropped. Note that channel selection methods modify the object in-place, so in interactive/exploratory sessions you may want to create a copy() first.

evoked_eeg = evoked.copy().pick(picks="eeg")
print(evoked_eeg.ch_names)

new_order = ["EEG 002", "MEG 2521", "EEG 003"]
evoked_subset = evoked.copy().reorder_channels(new_order)
print(evoked_subset.ch_names)

Similarities among the core data structures#

Evoked objects have many similarities with Raw and Epochs objects, including:

Loading and saving Evoked data#

Single Evoked objects can be saved to disk with the evoked.save() method. One difference between Evoked objects and the other data structures is that multiple Evoked objects can be saved into a single .fif file, using mne.write_evokeds(). The example data includes such a .fif file: the data have already been epoched and averaged, and the file contains separate Evoked objects for each experimental condition:

evk_file = root / "sample_audvis-ave.fif"
evokeds_list = mne.read_evokeds(evk_file, verbose=False)
print(evokeds_list)
print(type(evokeds_list))

Notice that mne.read_evokeds() returned a list of Evoked objects, and each one has an evoked.comment attribute describing the experimental condition that was averaged to generate the estimate:

for evok in evokeds_list:
    print(evok.comment)

If you want to load only some of the conditions present in a .fif file, read_evokeds() has a condition parameter, which takes either a string (matched against the comment attribute of the evoked objects on disk), or an integer selecting the Evoked object based on the order it is stored in the file. Passing lists of integers or strings is also possible. If only one object is selected, the Evoked object will be returned directly (rather than inside a list of length one):

right_vis = mne.read_evokeds(evk_file, condition="Right visual")
print(right_vis)
print(type(right_vis))

Previously, when we created an Evoked object by averaging epochs, baseline correction was applied by default when we extracted epochs from the Raw object (the default baseline period is (None, 0), which ensures zero mean for times before the stimulus event). In contrast, if we plot the first Evoked object in the list that was loaded from disk, we’ll see that the data have not been baseline-corrected:

evokeds_list[0].plot(picks="eeg")

This can be remedied by either passing a baseline parameter to mne.read_evokeds(), or by applying baseline correction after loading, as shown here:

# Original baseline (none set)
print(f"Baseline after loading: {evokeds_list[0].baseline}")

# Apply a custom baseline correction
evokeds_list[0].apply_baseline((None, 0))
print(f"Baseline after calling apply_baseline(): {evokeds_list[0].baseline}")

# Visualize the evoked response
evokeds_list[0].plot(picks="eeg")

Notice that apply_baseline() operated in-place. Similarly, Evoked objects may have been saved to disk with or without projectors applied; you can pass proj=True to the read_evokeds() function, or use the apply_proj() method after loading.

Combining Evoked objects#

One way to pool data across multiple conditions when estimating evoked responses is to do so prior to averaging (recall that MNE-Python can select based on partial matching of epoch labels separated by /; see Subselecting epochs for more information):

left_right_aud = epochs["auditory"].average()
left_right_aud

This approach will weight each epoch equally and create a single Evoked object. Notice that the printed representation includes (average, N=145), indicating that the Evoked object was created by averaging across 145 epochs. In this case, the event types were fairly close in number:

left_aud = epochs["auditory/left"].average()
right_aud = epochs["auditory/right"].average()
print([evok.nave for evok in (left_aud, right_aud)])

However, this may not always be the case. If for statistical reasons it is important to average the same number of epochs from different conditions, you can use equalize_event_counts() prior to averaging.

Another approach to pooling across conditions is to create separate Evoked objects for each condition, and combine them afterwards. This can be accomplished with the function mne.combine_evoked(), which computes a weighted sum of the Evoked objects given to it. The weights can be manually specified as a list or array of float values, or can be specified using the keyword 'equal' (weight each Evoked object by \(\frac{1}{N}\), where \(N\) is the number of Evoked objects given) or the keyword 'nave' (weight each Evoked object proportional to the number of epochs averaged together to create it):

left_right_aud = mne.combine_evoked([left_aud, right_aud], weights="nave")
assert left_right_aud.nave == left_aud.nave + right_aud.nave

Note that the nave attribute of the resulting Evoked object will reflect the effective number of averages, and depends on both the nave attributes of the contributing Evoked objects and the weights with which they are combined. Keeping track of effective nave is important for inverse imaging, because nave is used to scale the noise covariance estimate, which in turn affects the magnitude of estimated source activity (see The minimum-norm current estimates for more information, especially the Whitening and scaling section). Note that mne.grand_average() does not adjust nave to reflect the effective number of averaged epochs; it simply sets nave to the number of evokeds that were averaged together. For this reason, it is best to use mne.combine_evoked() rather than mne.grand_average() if you intend to perform inverse imaging on the resulting Evoked object.

Other uses of Evoked objects#

Although the most common use of Evoked objects is to store averages of epoched data, there are a few other uses worth noting here. First, the method epochs.standard_error() will create an Evoked object (just like epochs.average() does), but the data in the Evoked object will be the standard error across epochs instead of the average. To indicate this difference, Evoked objects have a kind attribute that takes values 'average' or 'standard error' as appropriate.

Another use of Evoked objects is to represent a single trial or epoch of data, usually when looping through epochs. This can be easily accomplished with the epochs.iter_evoked() method, and can be useful for applications where you want to do something that is only possible for Evoked objects. For example, here we use the get_peak() method (which is not available for Epochs objects) to get the peak response in each trial:

for ix, trial in enumerate(epochs[:3].iter_evoked()):
    channel, latency, value = trial.get_peak(ch_type="eeg", return_amplitude=True)
    latency = int(round(latency * 1e3))  # convert to milliseconds
    value = int(round(value * 1e6))  # convert to µV
    print(f"Trial {ix}: peak of {value} µV at {latency} ms in channel {channel}")

Estimated memory usage: 0 MB

Gallery generated by Sphinx-Gallery