Various automated MATLAB pipelines for preprocessing EEG

This repository holds several pipelines which were used to benchmark how much differences in EEG preprocessing affect downstream results as reported in the following paper:

How sensitive are EEG results to preprocessing methods: A Benchmarking study
Kay A. Robbins, Jonathan Touryan, Tim Mullen, Christian Kothe, Nima Bigdely-Shamlo
bioRxiv 2020.01.20.913327; doi: https://doi.org/10.1101/2020.01.20.913327

The pipelines use MATLAB and EEGLAB. Some of the components use the MATLAB statistics toolbox. Computation of the spectral fingerprints and spectral samples assumes the MATLAB wavelet toolbox. The code was tested using MATLAB 2019a and EEGLAB v2019.1.

The following EEGLAB plugins were installed in the default configuration:

1. clean_rawdata2.1
2. dipfit
3. firfilt2.3
4. ICLabel1.1

We also installed the following EEGLAB plugins:

1. blinker1.1.2
2. MARA1.2
3. PrepPipeline0.55.3

By installation, we mean that these plugins are unzipped into the EEGLAB/plugin directory. You should add them to your MATLAB path by running eeglab, not by trying to add individual directories to your path. We recommend that you always use the latest versions of the plugins.

You will also need go download eye-catch from https://github.com/bigdelys/eye-catch. Add this directory to your path as well. For convenience we have provided a script to add the project to your MATLAB path:

runEEGPipelineProjectPaths

The script assumes that the EEG-pipelines repository and the eye-catch repository are installed at the same level under a common subdirectory.

EEG: The EEG recording should be in EEG.set file format. The recording should be at least 15 minutes in length in order for the pipelines to work effectively, as several of the components (such as PREP and Blinker) use signal statistics to set thresholds. ASR needs data to calibrate.

Channel locations: The input EEG.set file MUST have channel locations included. Furthermore, the EEG.chanlocs.type fields must be set so that the EEG channels have the type 'EEG'. PREP, Blinker, MARA and other tools rely on being able to reliably distinguish the EEG channels from auxilliary channels in order to set their defaults. In our implementation, we remove the non-EEG channels in the first step. If you do not wish to do this, you must give channels to be excluded in several steps. To simplify the implementation, this code removes non-EEG channels as the first step in the pipeline.

Reduction of number of channels: The pipeline implementation reported in our publication reduces the number of channels to 64 prior to performing most of the pipeline. To do this, you must provide a channel mapping. We have provided one channel mapping (for Biosemi 256-channel headsets into closest channels in Biosemi 64-channel headsets using standard caps). The pipelines can be run with more than 64 channels, but we have not evaluated the results for more channels.

The original pipelines were implemented using a code infrastructure that supports large-scale processing. The codes in this repository were stripped out of that infrastructure and redesigned to run on a single EEG recording file in .set file format. The code was tested on MATLAB 2019a and EEGLAB version eeglab2019_1. The paper used cudaica for some of the computations of ICA. This requires special hardware and setup. That version of the code has not been included in the repository. We assume runica from EEGLAB for the ICA decomposition.

The pipelines are fully automated, once the parameters at the top of each script are set. The supported pipelines are:

• LARG uses PREP, ICA, Blinker, and eye-catch to identify and remove artifacts. (Note: this implementation allows subtraction of a residual blink signal, but does not currently implement regression out of blinks during preprocessing.)
• MARA uses PREP, ICA and MARA to identify and remove artifacts.
• ASR uses standard application of ASR from raw data (clean_artifacts).
• ASRalt (referred to in the paper and in the diagram below as ASR*) uses PREP for line noise, robust referencing, and bad channel identification and Hamming windows for filtering (as the LARG and MARA pipelines do). It then uses only the artifact subspace removal portion of ASR (clean_asr) to remove artifacts. (This alternative is not necessarily recommended, but was developed to make a closer comparison with LARG and MARA.)