Supplementary Material for “EEG Signal Processing in MI-BCI Applications with Improved Covariance Matrix Estimators”

Citation Author(s):
Universidad de Sevilla
Universidad de Sevilla
M. Auxiliadora
Universidad de Sevilla
Universidad de Sevilla
Submitted by:
Sergio Cruces
Last updated:
Tue, 05/17/2022 - 22:17
Data Format:
Research Article Link:
0 ratings - Please login to submit your rating.



This material is associated with the PhD Thesis of Javier Olias (which is supervised by Sergio Cruces) and the article: 

EEG Signal Processing in MI-BCI Applications with Improved Covariance Matrix Estimators” by J.Olias, R. Martin-Clemente, M.A. Sarmiento-Vega and S. Cruces,
which was accepted in 2019 by IEEE Transactions on Neural Systems and Rehabilitation Engineering.



In you will find the following files:

1) Data.mat
    Synthetic dataset of simulated EEG filtered recordings.
    It can be replaced by the datasets of the BCI competions for real testing.
    The data is stored in a MatLab struct type with the following fields:

    -x: Simulated EEG trials of dimension (n. samples)x(n. sensors)x(n. trials).
    -y: Classes of the trials in a vector of dimension (n. trials) x 1.
    -TrueCovClass: Tensor that stores the covariance of the 2 classes.
                   Its dimension is (n. sensors)x(n. sensors) x 2
2) Python demo file that illustrates the improvements obtained with the proposed power-normalization of the trials contained in Data.mat. The proposal can be interpreted as a generalization of Tyler's method for multiclass samples. The demo shows the improvement in the scale-invariant Riemannian distance of the estimated covariance matrices of the classes with respect to their true values once the proposed normalization is applied. > The average distance before normalization is 2.84 > The average distance after normalization is 1.88 It also reports the accuracy of the classification results obtained with the CSP+LDA classifier and the CSP+Tangent Space Logistic Regression classifiers, with and without normalization. The normalized versions nCSP+LDA and nCSP+TSLR clearly outperform the unnormalized ones CSP+LDA and CSP+TSLR. Accuracy CSP nCSP LDA 0.8175 0.8475 TSLR 0.8025 0.8700 3) Stores the functions implementing the proposed normalization and the scale-invariant Riemannian distance. 4)
Auxiliary functions for the demo. 5) Suplementary.pdf Supplementary material of the main article with extra figures and tables. It presents the obtained results of the algorithms for the multi-class paradigm.
For additional inquiries you can contact us by e-mail:,