Single-Trial Decoding of EEG data from Gel, Water and Dry Systems

This Repository holds the software part of my bachelor's thesis. It's main target is to continue the work of Müller-Putz et al.¹, answering following questions:

• How well can EEG signals be decoded in a cross-participant manner? i.e. When training a classificator with single-trial EEG data of 14 different participants, how well does it perform on a 15th, previously unseen, participant?
• How well can EEG signals be decoded in a cross-system manner? i.e. When training a classificator with single-trial EEG data of 3 different recording systems with data of 14 participants each, how well does it perform on each 15th, previously unseen, participant?

Sources¹²³⁴⁵⁶ are also referenced in the respective matlab files.

Software Documentation

Definitions

Variable name	Description
signal	EEG signal (R-by-X matrix)
ampval	single amplitude value in signal ("sample" is ambiguous when talking about training)
trial	part of signal where an event happens (R-by-L matrix)
trials	used to describe numerous trials (R-by-L-by-N matrix)
trainal	selected ampvals of trial, reshaped to vector (1-by-T*R) or (1-by-L) (in literature often called "sample", but is ambiguous, s.a.)
trainals	used to describe numerous trainals (N-by-T*R) or (N-by-L)
classes	(not only the class descriptions, but) used to describe a cell array {t_1, t_2, ...} consisting of the trials/trainals for class 1, class 2, etc.
latency	start (ampval index in the signal) of a trial
frame	latency and length (in ampvals) of a trial in the signal

one-char variable name abbreviations typically used:

Variable name	Description
C	number of classes
R	number of channels
T	number of ampvals in WOI
L	number of ampvals in trial/trainal
N	number of trials/trainals
D	number of features/ampvals (when talking about training)

files names

eeglab_datasets (unversioned)

filename	description
_ica.set	after running ICA algo on data, before removing components
_rmc.set	after removing independent components, before finishing preprocessing
.mat.set	after finishing preprocessing

A word on parallelisation and GPU usage

This package uses parfor to run a for loop on multiple workers (see best_timepoint.m) and also Matlab's built-in function gpuArray (see scov.m). Both are great ways to quickly use multi core CPUs and GPUs without changing much code, but also lack some logic behin the curtains. Using both with a single GPU may result in an 'Out of memory on device' error and I found no other way to fix it, than just using fewer workers. What I tried:

• Check available GPU memory before transferring arrays, using D = gpuDevice; D.AvailableMemory. It seems that this is not real-time data and therefore too slow.
• Catch the error using try...catch. The error somehow just gets through. Maybe try only works in a non-parallel context?

---

References

Footnotes

1. [Müller-Putz et al., 2020] Schwarz, A., Escolano, C., Montesano, L., Müller-Putz, G. (2020). "Analyzing and Decoding Natural Reach-and-Grasp Actions Using Gel, Water and Dry EEG Systems" Frontiers in Neuroscience 14. https://doi.org/10.3389/fnins.2020.00849 ↩ ↩²

2. [Blankertz et al., 2011] Blankertz, B., Lemm, S., Treder, M., Haufe, S., and Müller, K.-R. (2011). "Single-trial analysis and classification of ERP components—a tutorial" NeuroImage 56, 814–825. https://doi.org/10.1016/j.neuroimage.2010.06.048 ↩

3. [Schäfer et al., 2005] Schäfer, Juliane and Strimmer, Korbinian. (2005). "A Shrinkage Approach to Large-Scale Covariance Matrix Estimation and Implications for Functional Genomics" Statistical Applications in Genetics and Molecular Biology, vol. 4, no. 1, 2005. https://doi.org/10.2202/1544-6115.1175 ↩

4. [Duda et al. 2001] Duda, R.O., Hart, P.E., Stork, D.G. (2001). "Pattern Classification", 2nd Edition. Wiley & Sons. ISBN: 978-0-471-05669-0 ↩

5. [Combrisson et al., 2015] Combrisson, E., Jerbi, K. (2015). "Exceeding chance level by chance: The caveat of theoretical chance levels in brain signal classification and statistical assessment of decoding accuracy". Journal of Neuroscience Methods, Volume 250, 2015, Pages 126-136. https://dx.doi.org/10.1016/j.jneumeth.2015.01.010 ↩

6. [Lehmann et al., 1980] Lehmann, D., Skrandies, W. (1980). "Reference-free identification of components of checkerboard-evoked multichannel potential fields." Electroenceph. Clin. Neurophysiol., 1980. 48: 609-621. https://doi.org/10.1016/0013-4694(80)90419-8 ↩