Creating EpochsArray with explicit timestamps instead of sampling frequency

đź’¬ Support & Discussions
1

Hello,

I have EEG epochs stored in MatLab files with the following arrays:

  • allTargets: array of all target epochs with shape (n_epochs, n_samples, n_channels),
  • allNTargets: array of all non-target epochs,
  • tSCALE: array with sample timestamps (from range [-0.2;1] seconds) with shape (1, n_samples).

As the EpochsArray constructor requires an Info object, I created one with mne.create_info. Because I don’t have sampling frequency explicitly, I calculated it as sfreq = n_samples / 1.2. However, when I compare the tSCALE array with the epochsArray.times array (with np.allclose), they are not even close.
Is it possible to specify timestamps explicitly? Can I edit the EpochsArray.times attribute directly?

I am sorry if there is an answer somewhere already. I tried to search for it, but maybe I searched for the wrong keywords.
I really appreciate any help you can provide.

2

just a quick sanity check: did you do EpochsArray(..., tmin=-0.2)? Can you tell us more about how “not even close” the arrays are? (do you get a constant value if you subtract one from the other? what value?)

3

Yes, I did

I typed the sampling frequency formula into a calculator and got a result of 511.66… so I tried to round it to 512. When I create the EpochsArray with this sampling frequency, the difference is an almost constant vector with a single step in the middle at index 306.

I’m not sure if this makes any sense. Please let me know if you need to make something more clear.

4

that makes sense. Sampling frequency need not be an integer. I suggest when you create the info object you pass sfreq=n_samples / duration (i.e., whatever expression yielded the number 511.66…), and see if that gives the correct timestamps.

5

That is what I did earlier but then the difference between the arrays is not constant at all.

6

can you share a link to the relevant files and post a code sample showing what you’ve done so far (i.e., code that loads the files, then creates the info and the EpochsArray)?

7

I created a minimalistic example at colab:

Well, it seems like I was fooled by PyCharm’s debugger, which tells me that the difference between arrays is not constant, but it is. Sorry for the confusion.

8

OK, I realized what the problem is. MNE-Python requires when creating epochs that there be a sample at exactly time=0. So even though you specify tmin=-0.2, you end up with tmin equalling -0.19921875 because the time span tmin to zero does not accommodate an integer number of samples given your sampling frequency of 512 Hz. Put another way: in your data sfreq * tmin = 102.4 samples, so MNE-Python shifts everything by 0.4 samples to make sure we get a sample aligned at exactly zero.

The solution is to set tmin=0 when creating the epochs, and then shift times afterward:

import scipy.io
import numpy as np
from mne import create_info, EpochsArray

path = 'example.dat.mat'
subject = scipy.io.loadmat(path)


orig_times = subject['tSCALE'][0]
n_samp = subject['tSCALE'].shape[-1]
tmin, tmax = orig_times[[0, -1]]
duration = tmax - tmin
sfreq = (n_samp - 1) / duration
assert sfreq == 1 / np.diff(orig_times)[0]


ch_names = [x[0] for x in subject['electrodes'].flat]
ch_types = ['eeg'] * len(ch_names)

info = create_info(ch_names=ch_names, ch_types=ch_types, sfreq=sfreq)
info.set_montage('standard_1020')

# transpose to (n_epochs, n_channels, n_samples) 
data = subject['allTARGETS'].transpose(0, 2, 1)
n_epochs = data.shape[0]
n_samples = data.shape[2]

# create epochs
event_no = 1
events_dict = {'target': event_no}
events = np.column_stack((
    np.arange(0, n_epochs * n_samples, n_samples),
    np.zeros(n_epochs, dtype=int),
    np.full((n_epochs, ), event_no)
))
epoch = EpochsArray(
    data,
    info,
    tmin=0,               # <- changed
    baseline=(0, -tmin),  # <- note this change too
    events=events,
    event_id=events_dict
)

epoch = epoch.shift_time(tmin, relative=False)

np.testing.assert_allclose(orig_times, epoch.times)
9

Okay, good to know there has to be a sample at time=0.
Thank you for your help.