Creating MNE-Python data structures from scratch

This tutorial shows how to create MNE-Python’s core data structures using an existing NumPy array of (real or synthetic) data.

We begin by importing the necessary Python modules:

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

import numpy as np

import mne

Creating Info objects

Info objects

For full documentation on the Info object, see The Info data structure.

The core data structures for continuous (Raw), discontinuous (Epochs), and averaged (Evoked) data all have an info attribute comprising an mne.Info object. When reading recorded data using one of the functions in the mne.io submodule, Info objects are created and populated automatically. But if we want to create a Raw, Epochs, or Evoked object from scratch, we need to create an appropriate Info object as well. The easiest way to do this is with the mne.create_info function to initialize the required info fields. Additional fields can be assigned later as one would with a regular dictionary.

To initialize a minimal Info object requires a list of channel names, and the sampling frequency. As a convenience for simulated data, channel names can be provided as a single integer, and the names will be automatically created as sequential integers (starting with 0):

# Create some dummy metadata
n_channels = 32
sampling_freq = 200  # in Hertz
info = mne.create_info(n_channels, sfreq=sampling_freq)
print(info)

<Info | 7 non-empty values
 bads: []
 ch_names: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, ...
 chs: 32 misc
 custom_ref_applied: False
 highpass: 0.0 Hz
 lowpass: 100.0 Hz
 meas_date: unspecified
 nchan: 32
 projs: []
 sfreq: 200.0 Hz
>

You can see in the output above that, by default, the channels are assigned as type “misc” (where it says chs: 32 MISC). You can assign the channel type when initializing the Info object if you want:

ch_names = [f"MEG{n:03}" for n in range(1, 10)] + ["EOG001"]
ch_types = ["mag", "grad", "grad"] * 3 + ["eog"]
info = mne.create_info(ch_names, ch_types=ch_types, sfreq=sampling_freq)
print(info)

<Info | 7 non-empty values
 bads: []
 ch_names: MEG001, MEG002, MEG003, MEG004, MEG005, MEG006, MEG007, MEG008, ...
 chs: 3 Magnetometers, 6 Gradiometers, 1 EOG
 custom_ref_applied: False
 highpass: 0.0 Hz
 lowpass: 100.0 Hz
 meas_date: unspecified
 nchan: 10
 projs: []
 sfreq: 200.0 Hz
>

If the channel names follow one of the standard montage naming schemes, their spatial locations can be automatically added using the set_montage method:

ch_names = ["Fp1", "Fp2", "Fz", "Cz", "Pz", "O1", "O2"]
ch_types = ["eeg"] * 7
info = mne.create_info(ch_names, ch_types=ch_types, sfreq=sampling_freq)
info.set_montage("standard_1020")

	General
	MNE object type	Info
	Measurement date	Unknown
	Participant	Unknown
	Experimenter	Unknown
	Acquisition
	Sampling frequency	200.00 Hz
	Channels
	EEG	7
	Head & sensor digitization	10 points
	Filters
	Highpass	0.00 Hz
	Lowpass	100.00 Hz

Info consistency

When assigning new values to the fields of an Info object, it is important that the fields stay consistent. if there are N channels:

• The length of the channel information field chs must be N.

• The length of the ch_names field must be N.

• The ch_names field should be consistent with the name field of the channel information contained in chs.

Note the new field dig that includes our seven channel locations as well as theoretical values for the three cardinal scalp landmarks.

Additional fields can be added in the same way that Python dictionaries are modified, using square-bracket key assignment:

info["description"] = "My custom dataset"
info["bads"] = ["O1"]  # Names of bad channels
print(info)

<Info | 10 non-empty values
 bads: 1 items (O1)
 ch_names: Fp1, Fp2, Fz, Cz, Pz, O1, O2
 chs: 7 EEG
 custom_ref_applied: False
 description: My custom dataset
 dig: 10 items (3 Cardinal, 7 EEG)
 highpass: 0.0 Hz
 lowpass: 100.0 Hz
 meas_date: unspecified
 nchan: 7
 projs: []
 sfreq: 200.0 Hz
>

Creating Raw objects

Units

The expected units for the different channel types are:

• Volts: eeg, eog, seeg, dbs, emg, ecg, bio, ecog

• Teslas: mag

• Teslas/meter: grad

• Molar: hbo, hbr

• Amperes: dipole

• Arbitrary units: misc

To create a Raw object from scratch, you can use the mne.io.RawArray class constructor, which takes an Info object and a NumPy array of shape (n_channels, n_samples). Here, we’ll create some sinusoidal data and plot it:

times = np.linspace(0, 1, sampling_freq, endpoint=False)
sine = np.sin(20 * np.pi * times)
cosine = np.cos(10 * np.pi * times)
data = np.array([sine, cosine])

info = mne.create_info(
    ch_names=["10 Hz sine", "5 Hz cosine"], ch_types=["misc"] * 2, sfreq=sampling_freq
)

simulated_raw = mne.io.RawArray(data, info)
simulated_raw.plot(show_scrollbars=False, show_scalebars=False)

Creating RawArray with float64 data, n_channels=2, n_times=200
    Range : 0 ... 199 =      0.000 ...     0.995 secs
Ready.
Using qt as 2D backend.

Creating Epochs objects

To create an Epochs object from scratch, you can use the mne.EpochsArray class constructor, which takes an Info object and a NumPy array of shape (n_epochs, n_channels, n_samples). Here we’ll create 5 epochs of our 2-channel data, and plot it. Notice that we have to pass picks='misc' to the plot method, because by default it only plots data channels.

data = np.array(
    [
        [0.2 * sine, 1.0 * cosine],
        [0.4 * sine, 0.8 * cosine],
        [0.6 * sine, 0.6 * cosine],
        [0.8 * sine, 0.4 * cosine],
        [1.0 * sine, 0.2 * cosine],
    ]
)

simulated_epochs = mne.EpochsArray(data, info)
simulated_epochs.plot(picks="misc", show_scrollbars=False, events=True)

Not setting metadata
5 matching events found
No baseline correction applied
0 projection items activated
You seem to have overlapping epochs. Some event lines may be duplicated in the plot.

Since we did not supply an events array, the EpochsArray constructor automatically created one for us, with all epochs having the same event number:

print(simulated_epochs.events[:, -1])

[1 1 1 1 1]

If we want to simulate having different experimental conditions, we can pass an event array (and an event ID dictionary) to the constructor. Since our epochs are 1 second long and have 200 samples/second, we’ll put our events spaced 200 samples apart, and pass tmin=-0.5, so that the events land in the middle of each epoch (the events are always placed at time=0 in each epoch).

events = np.column_stack(
    (
        np.arange(0, 1000, sampling_freq),
        np.zeros(5, dtype=int),
        np.array([1, 2, 1, 2, 1]),
    )
)
event_dict = dict(condition_A=1, condition_B=2)
simulated_epochs = mne.EpochsArray(
    data, info, tmin=-0.5, events=events, event_id=event_dict
)
simulated_epochs.plot(
    picks="misc", show_scrollbars=False, events=events, event_id=event_dict
)

Not setting metadata
5 matching events found
No baseline correction applied
0 projection items activated

You could also create simulated epochs by using the normal Epochs (not EpochsArray) constructor on the simulated RawArray object, by creating an events array (e.g., using mne.make_fixed_length_events) and extracting epochs around those events.

Creating Evoked Objects

If you already have data that was averaged across trials, you can use it to create an Evoked object using the EvokedArray class constructor. It requires an Info object and a data array of shape (n_channels, n_times), and has an optional tmin parameter like EpochsArray does. It also has a parameter nave indicating how many trials were averaged together, and a comment parameter useful for keeping track of experimental conditions, etc. Here we’ll do the averaging on our NumPy array and use the resulting averaged data to make our Evoked.

# Create the Evoked object
evoked_array = mne.EvokedArray(
    data.mean(axis=0), info, tmin=-0.5, nave=data.shape[0], comment="simulated"
)
print(evoked_array)
evoked_array.plot()

<Evoked | 'simulated' (average, N=5), -0.5 – 0.495 s, baseline off, 2 ch, ~9 KiB>

In certain situations you may wish to use a custom time-frequency decomposition for estimation of power spectra. Or you may wish to process pre-computed power spectra in MNE. Following the same logic, it is possible to instantiate averaged power spectrum using the SpectrumArray or EpochsSpectrumArray classes.

# compute power spectrum

psd, freqs = mne.time_frequency.psd_array_welch(
    data, info["sfreq"], n_fft=128, n_per_seg=32
)

psd_ave = psd.mean(0)

info = mne.create_info(["Ch 1", "Ch2"], sfreq=sampling_freq, ch_types="eeg")
spectrum = mne.time_frequency.SpectrumArray(
    data=psd_ave,
    freqs=freqs,
    info=info,
)

spectrum.plot(spatial_colors=False, amplitude=False)

Effective window size : 0.640 (s)
Plotting power spectral density (dB=True).