Exporting Epochs to Pandas DataFrames

This tutorial shows how to export the data in Epochs objects to a Pandas DataFrame, and applies a typical Pandas split-apply-combine workflow to examine the latencies of the response maxima across epochs and conditions.

We’ll use the Sample dataset, but load a version of the raw file that has already been filtered and downsampled, and has an average reference applied to its EEG channels. As usual we’ll start by importing the modules we need and loading the data:

import os
import seaborn as sns
import mne

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder, 'MEG', 'sample',
                                    'sample_audvis_filt-0-40_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file, verbose=False)

Next we’ll load a list of events from file, map them to condition names with an event dictionary, set some signal rejection thresholds (cf. Rejecting Epochs based on channel amplitude), and segment the continuous data into epochs:

sample_data_events_file = os.path.join(sample_data_folder, 'MEG', 'sample',
                                       'sample_audvis_filt-0-40_raw-eve.fif')
events = mne.read_events(sample_data_events_file)

event_dict = {'auditory/left': 1, 'auditory/right': 2, 'visual/left': 3,
              'visual/right': 4}

reject_criteria = dict(mag=3000e-15,     # 3000 fT
                       grad=3000e-13,    # 3000 fT/cm
                       eeg=100e-6,       # 100 µV
                       eog=200e-6)       # 200 µV

tmin, tmax = (-0.2, 0.5)  # epoch from 200 ms before event to 500 ms after it
baseline = (None, 0)      # baseline period from start of epoch to time=0

epochs = mne.Epochs(raw, events, event_dict, tmin, tmax, proj=True,
                    baseline=baseline, reject=reject_criteria, preload=True)
del raw

Out:

288 matching events found
Applying baseline correction (mode: mean)
Not setting metadata
Created an SSP operator (subspace dimension = 4)
4 projection items activated
Loading data for 288 events and 106 original time points ...
    Rejecting  epoch based on EEG : ['EEG 003']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 004', 'EEG 006', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001']
    Rejecting  epoch based on EEG : ['EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on EEG : ['EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on MAG : ['MEG 1411', 'MEG 1421', 'MEG 1441']
    Rejecting  epoch based on MAG : ['MEG 1711']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on MAG : ['MEG 1411', 'MEG 1421']
    Rejecting  epoch based on MAG : ['MEG 1711']
    Rejecting  epoch based on EOG : ['EOG 061']
    Rejecting  epoch based on EEG : ['EEG 008']
    Rejecting  epoch based on EEG : ['EEG 008']
    Rejecting  epoch based on EOG : ['EOG 061']
    Rejecting  epoch based on EEG : ['EEG 007', 'EEG 008']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on MAG : ['MEG 1411', 'MEG 1421']
    Rejecting  epoch based on MAG : ['MEG 1411']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on EEG : ['EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on MAG : ['MEG 1331', 'MEG 1421']
    Rejecting  epoch based on EEG : ['EEG 007']
    Rejecting  epoch based on EOG : ['EOG 061']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
    Rejecting  epoch based on EEG : ['EEG 001', 'EEG 002', 'EEG 003', 'EEG 007']
35 bad epochs dropped

Converting an Epochs object to a DataFrame

Once we have our Epochs object, converting it to a DataFrame is simple: just call epochs.to_data_frame(). Each channel’s data will be a column of the new DataFrame, alongside three additional columns of event name, epoch number, and sample time. Here we’ll just show the first few rows and columns:

time condition epoch MEG 0113 MEG 0112 MEG 0111 MEG 0122 MEG 0123 MEG 0121 MEG 0132
0 -200 visual/left 1 21.320252 11.567619 -73.248584 -10.609895 -18.700540 -68.352746 -17.908667
1 -193 visual/left 1 7.049675 -9.702938 -34.168267 17.480975 -40.182237 -121.225684 2.072452
2 -186 visual/left 1 5.283711 -13.006895 -26.912317 52.130290 -65.157198 -117.842466 6.986073
3 -180 visual/left 1 23.040471 1.643716 -37.835772 48.517952 -65.214631 -82.043144 9.672744
4 -173 visual/left 1 25.688943 7.974112 -85.561683 6.450115 -43.778315 -111.795987 17.277138


Scaling time and channel values

By default, time values are converted from seconds to milliseconds and then rounded to the nearest integer; if you don’t want this, you can pass time_format=None to keep time as a float value in seconds, or convert it to a Timedelta value via time_format='timedelta'.

Note also that, by default, channel measurement values are scaled so that EEG data are converted to µV, magnetometer data are converted to fT, and gradiometer data are converted to fT/cm. These scalings can be customized through the scalings parameter, or suppressed by passing scalings=dict(eeg=1, mag=1, grad=1).

df = epochs.to_data_frame(time_format=None,
                          scalings=dict(eeg=1, mag=1, grad=1))
df.iloc[:5, :10]
time condition epoch MEG 0113 MEG 0112 MEG 0111 MEG 0122 MEG 0123 MEG 0121 MEG 0132
0 -0.199795 visual/left 1 2.132025e-12 1.156762e-12 -7.324858e-14 -1.060990e-12 -1.870054e-12 -6.835275e-14 -1.790867e-12
1 -0.193135 visual/left 1 7.049675e-13 -9.702938e-13 -3.416827e-14 1.748098e-12 -4.018224e-12 -1.212257e-13 2.072452e-13
2 -0.186476 visual/left 1 5.283711e-13 -1.300690e-12 -2.691232e-14 5.213029e-12 -6.515720e-12 -1.178425e-13 6.986073e-13
3 -0.179816 visual/left 1 2.304047e-12 1.643716e-13 -3.783577e-14 4.851795e-12 -6.521463e-12 -8.204314e-14 9.672744e-13
4 -0.173156 visual/left 1 2.568894e-12 7.974112e-13 -8.556168e-14 6.450115e-13 -4.377831e-12 -1.117960e-13 1.727714e-12


Notice that the time values are no longer integers, and the channel values have changed by several orders of magnitude compared to the earlier DataFrame.

Setting the index

It is also possible to move one or more of the indicator columns (event name, epoch number, and sample time) into the index, by passing a string or list of strings as the index parameter. We’ll also demonstrate here the effect of time_format='timedelta', yielding Timedelta values in the “time” column.

df = epochs.to_data_frame(index=['condition', 'epoch'],
                          time_format='timedelta')
df.iloc[:5, :10]
time MEG 0113 MEG 0112 MEG 0111 MEG 0122 MEG 0123 MEG 0121 MEG 0132 MEG 0133 MEG 0131
condition epoch
visual/left 1 -1 days +23:59:59.800204 21.320252 11.567619 -73.248584 -10.609895 -18.700540 -68.352746 -17.908667 45.379736 -53.123525
1 -1 days +23:59:59.806864 7.049675 -9.702938 -34.168267 17.480975 -40.182237 -121.225684 2.072452 27.734826 -95.339465
1 -1 days +23:59:59.813524 5.283711 -13.006895 -26.912317 52.130290 -65.157198 -117.842466 6.986073 5.996158 -177.627037
1 -1 days +23:59:59.820184 23.040471 1.643716 -37.835772 48.517952 -65.214631 -82.043144 9.672744 1.748030 -211.024973
1 -1 days +23:59:59.826844 25.688943 7.974112 -85.561683 6.450115 -43.778315 -111.795987 17.277138 2.123811 -173.917677


Wide- versus long-format DataFrames

Another parameter, long_format, determines whether each channel’s data is in a separate column of the DataFrame (long_format=False), or whether the measured values are pivoted into a single 'value' column with an extra indicator column for the channel name (long_format=True). Passing long_format=True will also create an extra column ch_type indicating the channel type.

long_df = epochs.to_data_frame(time_format=None, index='condition',
                               long_format=True)
long_df.head()

Out:

Converting "epoch" to "category"...
Converting "channel" to "category"...
Converting "ch_type" to "category"...
epoch time channel ch_type value
condition
visual/left 1 -0.199795 MEG 0113 grad 21.320252
visual/left 1 -0.199795 MEG 0112 grad 11.567619
visual/left 1 -0.199795 MEG 0111 mag -73.248584
visual/left 1 -0.199795 MEG 0122 grad -10.609895
visual/left 1 -0.199795 MEG 0123 grad -18.700540


Generating the DataFrame in long format can be helpful when using other Python modules for subsequent analysis or plotting. For example, here we’ll take data from the “auditory/left” condition, pick a couple MEG channels, and use seaborn.lineplot() to automatically plot the mean and confidence band for each channel, with confidence computed across the epochs in the chosen condition:

channels = ['MEG 1332', 'MEG 1342']
data = long_df.loc['auditory/left'].query('channel in @channels')
# convert channel column (CategoryDtype → string; for a nicer-looking legend)
data['channel'] = data['channel'].astype(str)
sns.lineplot(x='time', y='value', hue='channel', data=data)
plot 40 epochs to data frame

We can also now use all the power of Pandas for grouping and transforming our data. Here, we find the latency of peak activation of 2 gradiometers (one near auditory cortex and one near visual cortex), and plot the distribution of the timing of the peak in each channel as a violinplot():

df = epochs.to_data_frame(time_format=None)
peak_latency = (df.filter(regex=r'condition|epoch|MEG 1332|MEG 2123')
                .groupby(['condition', 'epoch'])
                .aggregate(lambda x: df['time'].iloc[x.idxmax()])
                .reset_index()
                .melt(id_vars=['condition', 'epoch'],
                      var_name='channel',
                      value_name='latency of peak')
                )

ax = sns.violinplot(x='channel', y='latency of peak', hue='condition',
                    data=peak_latency, palette='deep', saturation=1)
plot 40 epochs to data frame

Total running time of the script: ( 0 minutes 17.466 seconds)

Estimated memory usage: 921 MB

Gallery generated by Sphinx-Gallery