Process a big data EEG resource (TUH EEG Corpus)

In this example, we showcase usage of the Temple University Hospital EEG Corpus ( including simple preprocessing steps as well as cutting of compute windows.

# Author: Lukas Gemein <>
# License: BSD (3-clause)

import tempfile

import numpy as np
import matplotlib.pyplot as plt
import mne

from braindecode.datasets import TUH
from braindecode.preprocessing import (
    preprocess, Preprocessor, create_fixed_length_windows, scale as multiply)'seaborn')
mne.set_log_level('ERROR')  # avoid messages everytime a window is extracted

If you want to try this code with the actual data, please delete the next section. We are required to mock some dataset functionality, since the data is not available at creation time of this example.

from braindecode.datasets.tuh import _TUHMock as TUH  # noqa F811

We start by creating a TUH dataset. First, the class generates a description of the recordings in TUH_PATH (which is later accessible as tuh.description) without actually touching the files. This will parse information from file paths such as patient id, recording data, etc and should be really fast. Afterwards, the files are sorted chronologically by year, month, day, patient id, recording session and segment. In the following, a subset of the description corresponding to recording_ids is used. Afterwards, the files will be iterated a second time, slower than before. The files are now actually touched. Additional information about subjects like age and gender are parsed directly from the EDF file header. If existent, the physician report is added to the description. Furthermore, the recordings are read with with preload=False. Finally, we will get a BaseConcatDataset of BaseDatasets each holding a single which is fully compatible with other braindecode functionalities.

TUH_PATH = 'please insert actual path to data here'
N_JOBS = 2  # specify the number of jobs for loading and windowing
tuh = TUH(
    n_jobs=1 if TUH.__name__ == '_TUHMock' else N_JOBS,  # Mock dataset can't
    # be loaded in parallel

We can easily create descriptive statistics using the description DataFrame, for example an age histogram split by gender of patients.

fig, ax = plt.subplots(1, 1, figsize=(15, 5))
genders = tuh.description.gender.unique()
x = [tuh.description.age[tuh.description.gender == g] for g in genders]
    bins=np.arange(100, dtype=int),
ax.set_xlabel('Age [years]')
plot tuh eeg corpus


Text(152.97222222222223, 0.5, 'Count')

Next, we will perform some preprocessing steps. First, we will do some selection of available recordings based on the duration. We will select those recordings, that have at least five minutes duration. Data is not loaded here.

def select_by_duration(ds, tmin=0, tmax=None):
    if tmax is None:
        tmax = np.inf
    # determine length of the recordings and select based on tmin and tmax
    split_ids = []
    for d_i, d in enumerate(ds.datasets):
        duration = d.raw.n_times /['sfreq']
        if tmin <= duration <= tmax:
    splits = ds.split(split_ids)
    split = splits['0']
    return split

tmin = 5 * 60
tmax = None
tuh = select_by_duration(tuh, tmin, tmax)

Next, we will discard all recordings that have an incomplete channel configuration (wrt the channels that we are interested in, i.e. the 21 channels of the international 10-20-placement). The dataset is subdivided into recordings with ‘le’ and ‘ar’ reference which we will have to consider. Data is not loaded here.

short_ch_names = sorted([
    'A1', 'A2',
    'FP1', 'FP2', 'F3', 'F4', 'C3', 'C4', 'P3', 'P4', 'O1', 'O2',
    'F7', 'F8', 'T3', 'T4', 'T5', 'T6', 'FZ', 'CZ', 'PZ'])
ar_ch_names = sorted([
    'EEG A1-REF', 'EEG A2-REF',
    'EEG FP1-REF', 'EEG FP2-REF', 'EEG F3-REF', 'EEG F4-REF', 'EEG C3-REF',
    'EEG C4-REF', 'EEG P3-REF', 'EEG P4-REF', 'EEG O1-REF', 'EEG O2-REF',
    'EEG F7-REF', 'EEG F8-REF', 'EEG T3-REF', 'EEG T4-REF', 'EEG T5-REF',
le_ch_names = sorted([
    'EEG A1-LE', 'EEG A2-LE',
    'EEG FP1-LE', 'EEG FP2-LE', 'EEG F3-LE', 'EEG F4-LE', 'EEG C3-LE',
    'EEG C4-LE', 'EEG P3-LE', 'EEG P4-LE', 'EEG O1-LE', 'EEG O2-LE',
    'EEG F7-LE', 'EEG F8-LE', 'EEG T3-LE', 'EEG T4-LE', 'EEG T5-LE',
    'EEG T6-LE', 'EEG FZ-LE', 'EEG CZ-LE', 'EEG PZ-LE'])
assert len(short_ch_names) == len(ar_ch_names) == len(le_ch_names)
ar_ch_mapping = {ch_name: short_ch_name for ch_name, short_ch_name in zip(
    ar_ch_names, short_ch_names)}
le_ch_mapping = {ch_name: short_ch_name for ch_name, short_ch_name in zip(
    le_ch_names, short_ch_names)}
ch_mapping = {'ar': ar_ch_mapping, 'le': le_ch_mapping}

def select_by_channels(ds, ch_mapping):
    split_ids = []
    for i, d in enumerate(ds.datasets):
        ref = 'ar' if d.raw.ch_names[0].endswith('-REF') else 'le'
        # these are the channels we are looking for
        seta = set(ch_mapping[ref].keys())
        # these are the channels of the recoding
        setb = set(d.raw.ch_names)
        # if recording contains all channels we are looking for, include it
        if seta.issubset(setb):
    return ds.split(split_ids)['0']

tuh = select_by_channels(tuh, ch_mapping)

Next, we will chain several preprocessing steps that are realized through mne. Data will be loaded by the first preprocessor that has a mention of it in brackets:

  1. crop the recordings to a region of interest

  2. re-reference all recordings to ‘ar’ (requires load)

  3. rename channels to short channel names

  4. pick channels of interest

  5. scale signals to micro volts (requires load)

  6. clip outlier values to +/- 800 micro volts (requires load)

  7. resample recordings to a common frequency (requires load)

def custom_rename_channels(raw, mapping):
    # rename channels which are dependent on referencing:
    # le: EEG 01-LE, ar: EEG 01-REF
    # mne fails if the mapping contains channels as keys that are not present
    # in the raw
    reference = raw.ch_names[0].split('-')[-1].lower()
    assert reference in ['le', 'ref'], 'unexpected referencing'
    reference = 'le' if reference == 'le' else 'ar'

def custom_crop(raw, tmin=0.0, tmax=None, include_tmax=True):
    # crop recordings to tmin – tmax. can be incomplete if recording
    # has lower duration than tmax
    # by default mne fails if tmax is bigger than duration
    tmax = min((raw.n_times - 1) /['sfreq'], tmax)
    raw.crop(tmin=tmin, tmax=tmax, include_tmax=include_tmax)

tmin = 1 * 60
tmax = 6 * 60
sfreq = 100

preprocessors = [
    Preprocessor(custom_crop, tmin=tmin, tmax=tmax, include_tmax=False,
    Preprocessor('set_eeg_reference', ref_channels='average', ch_type='eeg'),
    Preprocessor(custom_rename_channels, mapping=ch_mapping,
    Preprocessor('pick_channels', ch_names=short_ch_names, ordered=True),
    Preprocessor(multiply, factor=1e6, apply_on_array=True),
    Preprocessor(np.clip, a_min=-800, a_max=800, apply_on_array=True),
    Preprocessor('resample', sfreq=sfreq),

Next, we apply the preprocessors on the selected recordings in parallel. We additionally use the serialization functionality of braindecode.preprocessing.preprocess() to limit memory usage during preprocessing (as each file must be loaded into memory for some of the preprocessing steps to work). This also makes it possible to use the lazy loading capabilities of braindecode.datasets.BaseConcatDataset, as the preprocessed data is automatically reloaded with preload=False.


Here we use n_jobs=2 as the machines the documentation is build on only have two cores. This number should be modified based on the machine that is available for preprocessing.

OUT_PATH = tempfile.mkdtemp()  # please insert actual output directory here
tuh_preproc = preprocess(

We can finally generate compute windows. The resulting dataset is now ready to be used for model training.

window_size_samples = 1000
window_stride_samples = 1000
# generate compute windows here and store them to disk
tuh_windows = create_fixed_length_windows(

for x, y, ind in tuh_windows:

Total running time of the script: ( 0 minutes 2.135 seconds)

Estimated memory usage: 9 MB

Gallery generated by Sphinx-Gallery