The Spectrum and EpochsSpectrum classes: frequency-domain data

This tutorial shows how to create and visualize frequency-domain representations of your data, starting from continuous Raw, discontinuous Epochs, or averaged Evoked data.

As usual we’ll start by importing the modules we need, and loading our sample dataset:

import numpy as np

import mne

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = sample_data_folder / "MEG" / "sample" / "sample_audvis_raw.fif"
raw = mne.io.read_raw_fif(sample_data_raw_file, verbose=False).crop(tmax=60)

All three sensor-space containers (Raw, Epochs, and Evoked) have a compute_psd() method with the same options.

raw.compute_psd()

Effective window size : 3.410 (s)

Data type	Power Spectrum
Units	eeg: V²/Hz
	grad: (T/m)²/Hz
	mag: T²/Hz
Data source	Raw
Dims	channel, freq
Estimation method	welch
Number of channels	364
Number of frequency bins	1025
Frequency range	0.00 – 300.31 Hz

By default, the spectral estimation method will be the Welch[1] method for continuous data, and the multitaper method [2] for epoched or averaged data. This default can be overridden by passing method='welch' or method='multitaper' to the compute_psd() method.

There are many other options available as well; for example we can compute a spectrum from a given span of times, for a chosen frequency range, and for a subset of the available channels:

raw.compute_psd(method="multitaper", tmin=10, tmax=20, fmin=5, fmax=30, picks="eeg")

Using multitaper spectrum estimation with 7 DPSS windows

Data type	Power Spectrum
Units	eeg: V²/Hz
Data source	Raw
Dims	channel, freq
Estimation method	multitaper
Number of channels	59
Number of frequency bins	250
Frequency range	5.10 – 30.00 Hz

You can also pass some parameters to the underlying spectral estimation function, such as the FFT window length and overlap for the Welch method; see the docstrings of mne.time_frequency.Spectrum (esp. its method_kw parameter) and the spectral estimation functions psd_array_welch() and psd_array_multitaper() for details.

For epoched data, the class of the spectral estimate will be mne.time_frequency.EpochsSpectrum instead of mne.time_frequency.Spectrum, but most of the API is the same for the two classes. For example, both have a get_data() method with an option to return the bin frequencies:

with mne.use_log_level("WARNING"):  # hide some irrelevant info messages
    events = mne.find_events(raw, stim_channel="STI 014")
    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
    )
epo_spectrum = epochs.compute_psd()
psds, freqs = epo_spectrum.get_data(return_freqs=True)
print(f"\nPSDs shape: {psds.shape}, freqs shape: {freqs.shape}")
epo_spectrum

    Using multitaper spectrum estimation with 7 DPSS windows

PSDs shape: (77, 364, 301), freqs shape: (301,)

Data type	Power Spectrum
Units	eeg: V²/Hz
	grad: (T/m)²/Hz
	mag: T²/Hz
Data source	Epochs
Number of epochs	77
Dims	epoch, channel, freq
Estimation method	multitaper
Number of channels	364
Number of frequency bins	301
Frequency range	0.00 – 299.81 Hz

Additionally, both Spectrum and EpochsSpectrum have __getitem__ methods, meaning their data can be accessed by square-bracket indexing. For Spectrum objects (computed from Raw or Evoked data), the indexing works similar to a Raw object or a NumPy array:

evoked = epochs["auditory"].average()
evk_spectrum = evoked.compute_psd()
# the first 3 frequency bins for the first 4 channels:
print(evk_spectrum[:4, :3])

NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
    Using multitaper spectrum estimation with 7 DPSS windows
[[5.61863537e-23 1.13487136e-22 9.88010499e-23]
 [2.89065956e-23 4.24056397e-23 3.91160668e-23]
 [1.56762015e-25 2.65842303e-25 2.39426775e-25]
 [5.38758770e-23 1.03771739e-22 9.98519756e-23]]

Hint

If the original Epochs object had a metadata dataframe attached, the derived EpochsSpectrum will inherit that metadata and will hence also support subselecting epochs via Pandas query strings.

In contrast, the EpochsSpectrum has indexing similar to Epochs objects: you can use string values to select spectral estimates for specific epochs based on their condition names, and what you get back is a new instance of EpochsSpectrum rather than a NumPy array of the data values. Selection via hierarchical event descriptors (HEDs) is also possible:

# get both "visual/left" and "visual/right" epochs:
epo_spectrum["visual"]

Data type	Power Spectrum
Units	eeg: V²/Hz
	grad: (T/m)²/Hz
	mag: T²/Hz
Data source	Epochs
Number of epochs	37
Dims	epoch, channel, freq
Estimation method	multitaper
Number of channels	364
Number of frequency bins	301
Frequency range	0.00 – 299.81 Hz

Visualizing Spectrum objects

Both Spectrum and EpochsSpectrum objects have plotting methods plot() (frequency × power), plot_topo() (frequency × power separately for each sensor), and plot_topomap() (interpolated scalp topography of power, in specific frequency bands). A few plot options are demonstrated below; see the docstrings for full details.

evk_spectrum.plot(picks="data", exclude="bads")
evk_spectrum.plot_topo(color="k", fig_facecolor="w", axis_facecolor="w")

• 
• 

evk_spectrum.plot_topomap(ch_type="eeg", agg_fun=np.median)

Migrating legacy code

Below is a quick-reference table of equivalent code from before and after the introduction of the Spectrum and EpochsSpectrum classes.

Quick reference for common Spectral class actions

Old	New
mne.time_frequency.psd_welch(raw)	raw.compute_psd().get_data(return_freqs=True)
mne.time_frequency.psd_multitaper(raw)	raw.compute_psd(method='multitaper').get_data(return_freqs=True)
raw.plot_psd(fmin, fmax, dB, area_mode='std')	raw.compute_psd(fmin, fmax).plot(dB, ci='std')
raw.plot_psd_topo(n_fft, overlap, axes)	raw.compute_psd(n_fft, overlap).plot_topo(axes)
epochs.plot_psd_topomap(tmax, bands)	epochs.compute_psd(tmax).plot_topomap(bands)

Warning

The functions mne.time_frequency.psd_welch and mne.time_frequency.psd_multitaper have been removed; new code should use the Raw.compute_psd(), Epochs.compute_psd(), and Evoked.compute_psd() methods, and pass method='welch' or method='multitaper' as a parameter.

The class methods Raw.plot_psd(), Epochs.plot_psd(), Raw.plot_psd_topo(), and Epochs.plot_psd_topomap() have been kept in the API to support legacy code, but should be avoided when writing new code.

References

[1]

Peter D. Welch. The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms. IEEE Transactions on Audio and Electroacoustics, 15(2):70–73, 1967. doi:10.1109/TAU.1967.1161901.

[2]

David S. Slepian. Prolate spheroidal wave functions, fourier analysis, and uncertainty-V: the discrete case. Bell System Technical Journal, 57(5):1371–1430, 1978. doi:10.1002/j.1538-7305.1978.tb02104.x.

Estimated memory usage: 152 MB