This repository contains the codes used to obtain the results in the article Embedding neurophysiological signals.

To install the necessary dependencies, you can simply use conda:

conda env create -f environment.yml

Then you must activate it to be able to reproduce the experiments:

conda activate embedding_eeg_2022

Before starting anything, you must create a file named local_config.yaml that suits your hardware configuration. To do this, you can make use of the template named local_config_TEMPLATE.yaml.

Then, you can run the experiments. The scripts named fbcsp_within-subject.py, eegnet_linear_probing.py, and eegnet_cross-subject.py will respectively run the MOABB evaluations for the FBCSP, EEGNet+LP, and EEGNet pipelines (see article).

Before running eegnet_linear_probing.py and eegnet_cross-subject.py, you should first execute the script eegnet_cross-subject_lightning.py. These two pipelines use the same pre-trained neural networks and the eegnet_cross-subject_lightning.py script will actually train these networks and save them for later use.

By default, the configuration file is set so that the FBCSP and EEGNet+LP pipelines are evaluated by 5-folds cross-validation on all the calibration trials available. The results with random permutation cross-validation that use restricted number of calibration trials can be computed by editing the configuration files:

1. In config.yaml, comment the lines:

n_perms: null
data_size: null

2. In config.yaml, uncomment the lines:

#    data_size:
#      policy: per_class
#      value: [1, 2, 4, 8, 16, 32, 64, 96]
#    n_perms: [50, 34, 23, 15, 10, 7, 5, 5]

3. In local_config.yaml, change the suffix field so that the previous results are not overwritten.

The plots and figures present in the article can be reproduced using the jupyter notebook visualization.ipynb.

To cite our work in a scientific publication, please use:

@article{
    TODO
}